• yes — protein body is stably on the outer leaflet under baseline localization via its own mechanism.
• contextual — protein body reaches the outer leaflet only under documented conditions. Transient reversible recruitment to a surface receptor does NOT count.
• no — not accessible from outside the cell.

verdict = "yes":

• classical_surface_receptor — single-pass TM with substantial extracellular domain.
• gpi_anchored — GPI anchor on the outer leaflet.
• multipass_with_exposed_loops — multi-pass TM (GPCR, transporter, channel) with extracellular loops.
• stable_complex_partner — no membrane anchor of its own, but a stable non-covalent partner of an anchored surface protein, assembled intracellularly and co-trafficked.
• other — explain the mechanism in verdict_reasoning.

verdict = "contextual":

• cell_state_induced — surfaces only under stress, oncogenic transformation, immunogenic / programmed cell death, infection, or activation-induced display.
• tissue_restricted_surface — surface display restricted to a narrow lineage (germline / reproductive, developmental, or a single specialized somatic cell type) — use this over yes even when the anchor type is unambiguous.
• lysosomal_exocytosis — lysosomal / late-endosomal TM protein reaches the PM via lysosomal exocytosis.
• dual_localization — documented PM pool alongside a dominant non-PM compartment, via active cycling or steady-state partial residence. Also covers TM proligands whose shed ectodomain is the dominant biological actor.
• stable_surface_attachment — secreted protein wash-resistantly anchored to a TM partner post-translationally (covalent or non-covalent, as long as it survives a buffer wash). Reversible binding or matrix deposition → secreted_only.
• other — explain the mechanism in verdict_reasoning.

verdict = "no":

• cytoplasmic — soluble cytoplasmic, no membrane association.
• nuclear — nuclear-resident (chromatin, nucleolar, nucleoplasmic).
• mitochondrial_internal — matrix or inner-membrane facing matrix.
• endomembrane_resident — ER / Golgi / lysosomal / peroxisomal / autophagosomal membrane only.
• nuclear_envelope — inner / outer nuclear membrane only.
• inner_leaflet_anchored — lipidated or peripheral on the cytoplasmic face of the PM.
• secreted_only — secreted with no wash-resistant surface anchoring. Covers transient non-covalent recruitment to surface receptors, matrix-deposited covalent products, and EV cargo.
• pmhc_only_intracellular — strictly intracellular; only "surface" story is MHC-presented peptides. pMHC is NOT credited for protein-body accessibility. TCR-T / TCR-mimic programs targeting an MHC-presented peptide go here.
• other — explain the mechanism in verdict_reasoning.

no is the highest-cost error: false negatives are not recoverable downstream while false positives are. Before committing to no, walk through each contextual bucket and confirm none plausibly applies for THIS specific gene (not by class analogy). Require gene-specific experimental evidence — surface biotinylation, flow cytometry on intact non-permeabilized cells, surface proteomics, or imaging on intact cells. Your verdict_reasoning must name each of the 5 contextual reasons and the specific evidence ruling each out — one short clause per bucket is enough.

Two especially-missed patterns worth checking explicitly:

• Cell-surface-directed therapeutic. An antibody / ADC / CAR-T / bispecific program engaging the protein on the cell surface is strong evidence for at least contextual. Exclude programs that bind a soluble circulating pool (anti-cytokine / anti-growth-factor / anti-complement against the secreted form). Exclude pMHC-targeting TCR-T / TCR-mimic / bispecifics — those stay no / pmhc_only_intracellular.

• Ectodomain shedding / TM precursor. "Predominantly detected as soluble" is NOT the same as "secreted-only" — if the gene encodes a single-pass TM precursor with documented sheddase / regulated-proteolysis biology, the membrane-anchored stage IS surface accessible. Stable TM precursor → yes / classical_surface_receptor; transient TM precursor of a shed-ligand-dominant gene → contextual / dual_localization. secreted_only applies only when no isoform is membrane-anchored at any stage.

When in doubt, contextual beats no.

Emit a single JSON object as your entire response. No prose around it, no fences.

{
  "verdict": "yes" | "contextual" | "no",
  "verdict_reasoning": "<= 800 chars explaining the call",
  "reason": "<one of the literals above>",
  "confidence": "low" | "medium" | "high",
  "key_uncertainty": "<= 200 chars naming the unresolved ambiguity, or null"
}

• confidence: high only when the verdict rests on explicit, unambiguous evidence. medium when well-supported but a judgment call between two plausible buckets. low when at least one contextual bucket has a plausible argument you couldn't conclusively rule out, when the call rests on absence of evidence, or when family-lineage evidence pulls against the per-gene evidence.
• key_uncertainty: when confidence != "high", name the unresolved bucket / mechanism in ≤200 chars. Set to null only when confidence = "high".
• verdict_reasoning: don't restate the verdict; argue for it. If you pick "other", name the mechanism explicitly.
• Pick the single best reason. The JSON must validate against the TriageRecordDraft schema.

• yes — protein body is stably on the outer leaflet under baseline localization via its own mechanism.
• contextual — protein body reaches the outer leaflet only under documented conditions. Transient reversible recruitment to a surface receptor does NOT count.
• no — not accessible from outside the cell.

verdict = "yes":

• classical_surface_receptor — single-pass TM with substantial extracellular domain.
• gpi_anchored — GPI anchor on the outer leaflet.
• multipass_with_exposed_loops — multi-pass TM (GPCR, transporter, channel) with extracellular loops.
• stable_complex_partner — no membrane anchor of its own, but a stable non-covalent partner of an anchored surface protein, assembled intracellularly and co-trafficked.
• other — explain the mechanism in verdict_reasoning.

verdict = "contextual":

• cell_state_induced — surfaces only under stress, oncogenic transformation, immunogenic / programmed cell death, infection, or activation-induced display.
• tissue_restricted_surface — surface display restricted to a narrow lineage (germline / reproductive, developmental, or a single specialized somatic cell type) — use this over yes even when the anchor type is unambiguous.
• lysosomal_exocytosis — lysosomal / late-endosomal TM protein reaches the PM via lysosomal exocytosis.
• dual_localization — documented PM pool alongside a dominant non-PM compartment, via active cycling or steady-state partial residence. Also covers TM proligands whose shed ectodomain is the dominant biological actor.
• stable_surface_attachment — secreted protein wash-resistantly anchored to a TM partner post-translationally (covalent or non-covalent, as long as it survives a buffer wash). Reversible binding or matrix deposition → secreted_only.
• other — explain the mechanism in verdict_reasoning.

verdict = "no":

• cytoplasmic — soluble cytoplasmic, no membrane association.
• nuclear — nuclear-resident (chromatin, nucleolar, nucleoplasmic).
• mitochondrial_internal — matrix or inner-membrane facing matrix.
• endomembrane_resident — ER / Golgi / lysosomal / peroxisomal / autophagosomal membrane only.
• nuclear_envelope — inner / outer nuclear membrane only.
• inner_leaflet_anchored — lipidated or peripheral on the cytoplasmic face of the PM.
• secreted_only — secreted with no wash-resistant surface anchoring. Covers transient non-covalent recruitment to surface receptors, matrix-deposited covalent products, and EV cargo.
• pmhc_only_intracellular — strictly intracellular; only "surface" story is MHC-presented peptides. pMHC is NOT credited for protein-body accessibility. TCR-T / TCR-mimic programs targeting an MHC-presented peptide go here.
• other — explain the mechanism in verdict_reasoning.

no is the highest-cost error: false negatives are not recoverable downstream while false positives are. Before committing to no, walk through each contextual bucket and confirm none plausibly applies for THIS specific gene (not by class analogy). Require gene-specific experimental evidence — surface biotinylation, flow cytometry on intact non-permeabilized cells, surface proteomics, or imaging on intact cells. Your verdict_reasoning must name each of the 5 contextual reasons and the specific evidence ruling each out — one short clause per bucket is enough.

Two especially-missed patterns worth checking explicitly:

• Cell-surface-directed therapeutic. An antibody / ADC / CAR-T / bispecific program engaging the protein on the cell surface is strong evidence for at least contextual. Exclude programs that bind a soluble circulating pool (anti-cytokine / anti-growth-factor / anti-complement against the secreted form). Exclude pMHC-targeting TCR-T / TCR-mimic / bispecifics — those stay no / pmhc_only_intracellular.

• Ectodomain shedding / TM precursor. "Predominantly detected as soluble" is NOT the same as "secreted-only" — if the gene encodes a single-pass TM precursor with documented sheddase / regulated-proteolysis biology, the membrane-anchored stage IS surface accessible. Stable TM precursor → yes / classical_surface_receptor; transient TM precursor of a shed-ligand-dominant gene → contextual / dual_localization. secreted_only applies only when no isoform is membrane-anchored at any stage.

When in doubt, contextual beats no.

Emit a single JSON object as your entire response. No prose around it, no fences.

{
  "verdict": "yes" | "contextual" | "no",
  "verdict_reasoning": "<= 800 chars explaining the call",
  "reason": "<one of the literals above>",
  "confidence": "low" | "medium" | "high",
  "key_uncertainty": "<= 200 chars naming the unresolved ambiguity, or null"
}

• confidence: high only when the verdict rests on explicit, unambiguous evidence. medium when well-supported but a judgment call between two plausible buckets. low when at least one contextual bucket has a plausible argument you couldn't conclusively rule out, when the call rests on absence of evidence, or when family-lineage evidence pulls against the per-gene evidence.
• key_uncertainty: when confidence != "high", name the unresolved bucket / mechanism in ≤200 chars. Set to null only when confidence = "high".
• verdict_reasoning: don't restate the verdict; argue for it. If you pick "other", name the mechanism explicitly.
• Pick the single best reason. The JSON must validate against the TriageRecordDraft schema.

• yes — protein body is stably on the outer leaflet under baseline localization via its own mechanism.
• contextual — protein body reaches the outer leaflet only under documented conditions. Transient reversible recruitment to a surface receptor does NOT count.
• no — not accessible from outside the cell.

verdict = "yes":

• classical_surface_receptor — single-pass TM with substantial extracellular domain.
• gpi_anchored — GPI anchor on the outer leaflet.
• multipass_with_exposed_loops — multi-pass TM (GPCR, transporter, channel) with extracellular loops.
• stable_complex_partner — no membrane anchor of its own, but a stable non-covalent partner of an anchored surface protein, assembled intracellularly and co-trafficked.
• other — explain the mechanism in verdict_reasoning.

verdict = "contextual":

• cell_state_induced — surfaces only under stress, oncogenic transformation, immunogenic / programmed cell death, infection, or activation-induced display.
• tissue_restricted_surface — surface display restricted to a narrow lineage (germline / reproductive, developmental, or a single specialized somatic cell type) — use this over yes even when the anchor type is unambiguous.
• lysosomal_exocytosis — lysosomal / late-endosomal TM protein reaches the PM via lysosomal exocytosis.
• dual_localization — documented PM pool alongside a dominant non-PM compartment, via active cycling or steady-state partial residence. Also covers TM proligands whose shed ectodomain is the dominant biological actor.
• stable_surface_attachment — secreted protein wash-resistantly anchored to a TM partner post-translationally (covalent or non-covalent, as long as it survives a buffer wash). Reversible binding or matrix deposition → secreted_only.
• other — explain the mechanism in verdict_reasoning.

verdict = "no":

• cytoplasmic — soluble cytoplasmic, no membrane association.
• nuclear — nuclear-resident (chromatin, nucleolar, nucleoplasmic).
• mitochondrial_internal — matrix or inner-membrane facing matrix.
• endomembrane_resident — ER / Golgi / lysosomal / peroxisomal / autophagosomal membrane only.
• nuclear_envelope — inner / outer nuclear membrane only.
• inner_leaflet_anchored — lipidated or peripheral on the cytoplasmic face of the PM.
• secreted_only — secreted with no wash-resistant surface anchoring. Covers transient non-covalent recruitment to surface receptors, matrix-deposited covalent products, and EV cargo.
• pmhc_only_intracellular — strictly intracellular; only "surface" story is MHC-presented peptides. pMHC is NOT credited for protein-body accessibility. TCR-T / TCR-mimic programs targeting an MHC-presented peptide go here.
• other — explain the mechanism in verdict_reasoning.

no is the highest-cost error: false negatives are not recoverable downstream while false positives are. Before committing to no, walk through each contextual bucket and confirm none plausibly applies for THIS specific gene (not by class analogy). Require gene-specific experimental evidence — surface biotinylation, flow cytometry on intact non-permeabilized cells, surface proteomics, or imaging on intact cells. Your verdict_reasoning must name each of the 5 contextual reasons and the specific evidence ruling each out — one short clause per bucket is enough.

Two especially-missed patterns worth checking explicitly:

• Cell-surface-directed therapeutic. An antibody / ADC / CAR-T / bispecific program engaging the protein on the cell surface is strong evidence for at least contextual. Exclude programs that bind a soluble circulating pool (anti-cytokine / anti-growth-factor / anti-complement against the secreted form). Exclude pMHC-targeting TCR-T / TCR-mimic / bispecifics — those stay no / pmhc_only_intracellular.

• Ectodomain shedding / TM precursor. "Predominantly detected as soluble" is NOT the same as "secreted-only" — if the gene encodes a single-pass TM precursor with documented sheddase / regulated-proteolysis biology, the membrane-anchored stage IS surface accessible. Stable TM precursor → yes / classical_surface_receptor; transient TM precursor of a shed-ligand-dominant gene → contextual / dual_localization. secreted_only applies only when no isoform is membrane-anchored at any stage.

When in doubt, contextual beats no.

Emit a single JSON object as your entire response. No prose around it, no fences.

{
  "verdict": "yes" | "contextual" | "no",
  "verdict_reasoning": "<= 800 chars explaining the call",
  "reason": "<one of the literals above>",
  "confidence": "low" | "medium" | "high",
  "key_uncertainty": "<= 200 chars naming the unresolved ambiguity, or null"
}

• confidence: high only when the verdict rests on explicit, unambiguous evidence. medium when well-supported but a judgment call between two plausible buckets. low when at least one contextual bucket has a plausible argument you couldn't conclusively rule out, when the call rests on absence of evidence, or when family-lineage evidence pulls against the per-gene evidence.
• key_uncertainty: when confidence != "high", name the unresolved bucket / mechanism in ≤200 chars. Set to null only when confidence = "high".
• verdict_reasoning: don't restate the verdict; argue for it. If you pick "other", name the mechanism explicitly.
• Pick the single best reason. The JSON must validate against the TriageRecordDraft schema.

• yes — protein body is stably on the outer leaflet under baseline localization via its own mechanism.
• contextual — protein body reaches the outer leaflet only under documented conditions. Transient reversible recruitment to a surface receptor does NOT count.
• no — not accessible from outside the cell.

verdict = "yes":

• classical_surface_receptor — single-pass TM with substantial extracellular domain.
• gpi_anchored — GPI anchor on the outer leaflet.
• multipass_with_exposed_loops — multi-pass TM (GPCR, transporter, channel) with extracellular loops.
• stable_complex_partner — no membrane anchor of its own, but a stable non-covalent partner of an anchored surface protein, assembled intracellularly and co-trafficked.
• other — explain the mechanism in verdict_reasoning.

verdict = "contextual":

• cell_state_induced — surfaces only under stress, oncogenic transformation, immunogenic / programmed cell death, infection, or activation-induced display.
• tissue_restricted_surface — surface display restricted to a narrow lineage (germline / reproductive, developmental, or a single specialized somatic cell type) — use this over yes even when the anchor type is unambiguous.
• lysosomal_exocytosis — lysosomal / late-endosomal TM protein reaches the PM via lysosomal exocytosis.
• dual_localization — documented PM pool alongside a dominant non-PM compartment, via active cycling or steady-state partial residence. Also covers TM proligands whose shed ectodomain is the dominant biological actor.
• stable_surface_attachment — secreted protein wash-resistantly anchored to a TM partner post-translationally (covalent or non-covalent, as long as it survives a buffer wash). Reversible binding or matrix deposition → secreted_only.
• other — explain the mechanism in verdict_reasoning.

verdict = "no":

• cytoplasmic — soluble cytoplasmic, no membrane association.
• nuclear — nuclear-resident (chromatin, nucleolar, nucleoplasmic).
• mitochondrial_internal — matrix or inner-membrane facing matrix.
• endomembrane_resident — ER / Golgi / lysosomal / peroxisomal / autophagosomal membrane only.
• nuclear_envelope — inner / outer nuclear membrane only.
• inner_leaflet_anchored — lipidated or peripheral on the cytoplasmic face of the PM.
• secreted_only — secreted with no wash-resistant surface anchoring. Covers transient non-covalent recruitment to surface receptors, matrix-deposited covalent products, and EV cargo.
• pmhc_only_intracellular — strictly intracellular; only "surface" story is MHC-presented peptides. pMHC is NOT credited for protein-body accessibility. TCR-T / TCR-mimic programs targeting an MHC-presented peptide go here.
• other — explain the mechanism in verdict_reasoning.

no is the highest-cost error: false negatives are not recoverable downstream while false positives are. Before committing to no, walk through each contextual bucket and confirm none plausibly applies for THIS specific gene (not by class analogy). Require gene-specific experimental evidence — surface biotinylation, flow cytometry on intact non-permeabilized cells, surface proteomics, or imaging on intact cells. Your verdict_reasoning must name each of the 5 contextual reasons and the specific evidence ruling each out — one short clause per bucket is enough.

Two especially-missed patterns worth checking explicitly:

• Cell-surface-directed therapeutic. An antibody / ADC / CAR-T / bispecific program engaging the protein on the cell surface is strong evidence for at least contextual. Exclude programs that bind a soluble circulating pool (anti-cytokine / anti-growth-factor / anti-complement against the secreted form). Exclude pMHC-targeting TCR-T / TCR-mimic / bispecifics — those stay no / pmhc_only_intracellular.

• Ectodomain shedding / TM precursor. "Predominantly detected as soluble" is NOT the same as "secreted-only" — if the gene encodes a single-pass TM precursor with documented sheddase / regulated-proteolysis biology, the membrane-anchored stage IS surface accessible. Stable TM precursor → yes / classical_surface_receptor; transient TM precursor of a shed-ligand-dominant gene → contextual / dual_localization. secreted_only applies only when no isoform is membrane-anchored at any stage.

When in doubt, contextual beats no.

Emit a single JSON object as your entire response. No prose around it, no fences.

{
  "verdict": "yes" | "contextual" | "no",
  "verdict_reasoning": "<= 800 chars explaining the call",
  "reason": "<one of the literals above>",
  "confidence": "low" | "medium" | "high",
  "key_uncertainty": "<= 200 chars naming the unresolved ambiguity, or null"
}

• confidence: high only when the verdict rests on explicit, unambiguous evidence. medium when well-supported but a judgment call between two plausible buckets. low when at least one contextual bucket has a plausible argument you couldn't conclusively rule out, when the call rests on absence of evidence, or when family-lineage evidence pulls against the per-gene evidence.
• key_uncertainty: when confidence != "high", name the unresolved bucket / mechanism in ≤200 chars. Set to null only when confidence = "high".
• verdict_reasoning: don't restate the verdict; argue for it. If you pick "other", name the mechanism explicitly.
• Pick the single best reason. The JSON must validate against the TriageRecordDraft schema.

• yes — protein body is stably on the outer leaflet under baseline localization via its own mechanism.
• contextual — protein body reaches the outer leaflet only under documented conditions. Transient reversible recruitment to a surface receptor does NOT count.
• no — not accessible from outside the cell.

verdict = "yes":

• classical_surface_receptor — single-pass TM with substantial extracellular domain.
• gpi_anchored — GPI anchor on the outer leaflet.
• multipass_with_exposed_loops — multi-pass TM (GPCR, transporter, channel) with extracellular loops.
• stable_complex_partner — no membrane anchor of its own, but a stable non-covalent partner of an anchored surface protein, assembled intracellularly and co-trafficked.
• other — explain the mechanism in verdict_reasoning.

verdict = "contextual":

• cell_state_induced — surfaces only under stress, oncogenic transformation, immunogenic / programmed cell death, infection, or activation-induced display.
• tissue_restricted_surface — surface display restricted to a narrow lineage (germline / reproductive, developmental, or a single specialized somatic cell type) — use this over yes even when the anchor type is unambiguous.
• lysosomal_exocytosis — lysosomal / late-endosomal TM protein reaches the PM via lysosomal exocytosis.
• dual_localization — documented PM pool alongside a dominant non-PM compartment, via active cycling or steady-state partial residence. Also covers TM proligands whose shed ectodomain is the dominant biological actor.
• stable_surface_attachment — secreted protein wash-resistantly anchored to a TM partner post-translationally (covalent or non-covalent, as long as it survives a buffer wash). Reversible binding or matrix deposition → secreted_only.
• other — explain the mechanism in verdict_reasoning.

verdict = "no":

• cytoplasmic — soluble cytoplasmic, no membrane association.
• nuclear — nuclear-resident (chromatin, nucleolar, nucleoplasmic).
• mitochondrial_internal — matrix or inner-membrane facing matrix.
• endomembrane_resident — ER / Golgi / lysosomal / peroxisomal / autophagosomal membrane only.
• nuclear_envelope — inner / outer nuclear membrane only.
• inner_leaflet_anchored — lipidated or peripheral on the cytoplasmic face of the PM.
• secreted_only — secreted with no wash-resistant surface anchoring. Covers transient non-covalent recruitment to surface receptors, matrix-deposited covalent products, and EV cargo.
• pmhc_only_intracellular — strictly intracellular; only "surface" story is MHC-presented peptides. pMHC is NOT credited for protein-body accessibility. TCR-T / TCR-mimic programs targeting an MHC-presented peptide go here.
• other — explain the mechanism in verdict_reasoning.

no is the highest-cost error: false negatives are not recoverable downstream while false positives are. Before committing to no, walk through each contextual bucket and confirm none plausibly applies for THIS specific gene (not by class analogy). Require gene-specific experimental evidence — surface biotinylation, flow cytometry on intact non-permeabilized cells, surface proteomics, or imaging on intact cells. Your verdict_reasoning must name each of the 5 contextual reasons and the specific evidence ruling each out — one short clause per bucket is enough.

Two especially-missed patterns worth checking explicitly:

• Cell-surface-directed therapeutic. An antibody / ADC / CAR-T / bispecific program engaging the protein on the cell surface is strong evidence for at least contextual. Exclude programs that bind a soluble circulating pool (anti-cytokine / anti-growth-factor / anti-complement against the secreted form). Exclude pMHC-targeting TCR-T / TCR-mimic / bispecifics — those stay no / pmhc_only_intracellular.

• Ectodomain shedding / TM precursor. "Predominantly detected as soluble" is NOT the same as "secreted-only" — if the gene encodes a single-pass TM precursor with documented sheddase / regulated-proteolysis biology, the membrane-anchored stage IS surface accessible. Stable TM precursor → yes / classical_surface_receptor; transient TM precursor of a shed-ligand-dominant gene → contextual / dual_localization. secreted_only applies only when no isoform is membrane-anchored at any stage.

When in doubt, contextual beats no.

Emit a single JSON object as your entire response. No prose around it, no fences.

{
  "verdict": "yes" | "contextual" | "no",
  "verdict_reasoning": "<= 800 chars explaining the call",
  "reason": "<one of the literals above>",
  "confidence": "low" | "medium" | "high",
  "key_uncertainty": "<= 200 chars naming the unresolved ambiguity, or null"
}

• confidence: high only when the verdict rests on explicit, unambiguous evidence. medium when well-supported but a judgment call between two plausible buckets. low when at least one contextual bucket has a plausible argument you couldn't conclusively rule out, when the call rests on absence of evidence, or when family-lineage evidence pulls against the per-gene evidence.
• key_uncertainty: when confidence != "high", name the unresolved bucket / mechanism in ≤200 chars. Set to null only when confidence = "high".
• verdict_reasoning: don't restate the verdict; argue for it. If you pick "other", name the mechanism explicitly.
• Pick the single best reason. The JSON must validate against the TriageRecordDraft schema.

• When discard, name what the paper is actually about (so the call can be audited).
• When keep_abstract, state the surface-biology claim the abstract already makes.
• When worth_fetching, name what kinds of evidence the body likely contains.

Respond ONLY with one fenced ```json block matching this AbstractTriageResponse schema:

{schema}

Stop after emitting the JSON block — no prose around it.

A clip is load-bearing if it directly evidences one of these surface-evidence categories. Err on the side of keeping when in doubt: A1's selector is the gatekeeper for final inclusion, the trim pass should just protect it from obvious noise and from A2-only material.

1. Surface-evidence methods (the hinge of A1). The assay used to detect {gene} at the cell surface, with enough methods detail that A1's selector can fill MethodObservation.method_family / method_subclass / permeabilization / antibodies / accessibility_relevance. Includes:
   • Flow cytometry on live (non-permeabilized) cells with antibody against {gene}.
   • Cell-surface biotinylation (sulfo-NHS-biotin, sulfo-NHS-SS-biotin) followed by streptavidin pulldown + WB or MS.
   • Cell-surface-capture mass spectrometry / surfaceome MS (CSC, SUSPECS, etc.).
   • Immunofluorescence — BOTH non-permeabilized IF on intact cells AND permeabilized IF when it describes a membrane-staining pattern or colocalization with a plasma-membrane marker (the permeabilization status is the selector's call; keep either).
   • IHC scored explicitly for membrane staining pattern.
   • Live-cell tracking / surface-pool measurement / internalization-rate assays.
2. Topology: single-pass / multi-pass / 7TM / GPI-anchored, ECD vs ICD length, signal peptide presence, N-terminal vs C-terminal orientation, hydropathy / DeepTMHMM-style assignments. This anchors what counts as "the extracellular face" for A1.
3. Shed or secreted forms. Soluble {gene} in supernatant / serum / plasma, sheddase activity (sheddase enzymes, matrix metalloproteinases, γ-secretase), constitutive secretion of a non-anchored isoform. Critical for risks.shed_form + risks.secreted_form downstream.
4. Epitope masking — glycan shielding, partner-protein coverage of the binding epitope (heterodimer / hetero-complex, e.g. a co-receptor covering the target's large extracellular loop), homo-oligomerization / self-association that buries the epitope at the protein's own dimer interface (e.g. tetraspanin or claudin cis-clustering), and conformational occlusion (closed vs open conformer).
5. Therapeutic engagement of the ECD. Clinical or preclinical binders (antibodies, ADCs, CAR-T) that reach {gene} at the surface — direct surface-accessibility evidence. Keep the binder and the form it engages (membrane vs secreted).
6. Methodological rigor + antibody specificity. Antibody clone IDs, RRIDs, KO-validation, paralog cross-reactivity tests, isotype controls, paired WB with fractionation step, blocking-peptide controls. These let A1 populate AntibodyRef.validation_strategy and cross_reactivity_notes — they are NOT background noise.
7. Non-surface evidence that qualifies a surface claim. RNA (RT-qPCR, RNA-seq), whole-cell western blot WITHOUT a fractionation step, total-cell IHC without membrane scoring, permeabilized IF with NO membrane pattern. KEEP these — they feed A1's non_surface_expression list (the bucket that prevents expression from being misread as accessibility: it lets the reader see "expressed, surface unconfirmed"). Mark them in your reason so the selector files them on the non-surface side. (A permeabilized IF that DOES show a membrane pattern is a surface-localization method → item 1, not here.)
8. Contradictions to surface presence. Studies finding {gene} primarily intracellular, no detectable surface signal where another paper reported one, failure-to-replicate, paralog confound (an antibody that turns out to cross-react with a paralog).
9. Structure of the ECD — PDB entries, AlphaFold predictions explicitly framed as the extracellular domain, crystal structures of the ECD with a ligand or antibody.
10. Overexpression-based surface evidence. KEEP clips that describe surface detection in transfected cell lines (HEK293, CHO, 293T, HeLa-OE, COS-7) — for orphan / under-studied genes overexpression is often the strongest available evidence. When you keep an overexpression clip, also keep the methods sentence that names the construct's signal peptide. The signal peptide source decides whether the surface localization is biologically meaningful:
    • Native / endogenous SP — phrases like "endogenous signal peptide", "native signal sequence", "untagged construct", "wild-type [GENE]", "full-length [GENE]", "[GENE] cDNA" with no leader-replacement mention → trafficking is the protein's own. Real evidence (tier as if endogenous).
    • Exogenous / foreign SP — phrases like "IgG kappa leader", "IgG κ light-chain SP", "preprotrypsin signal peptide", "PreS leader", "interleukin-2 secretion signal", "Igλ leader", "honeybee melittin SP", "N-terminal HA-tagged with [vector]-derived leader" → the SP is forcing secretory-pathway entry. A protein that's normally cytosolic can end up on the surface purely because of the foreign SP. Tier down — supportive only.
    • Mark the SP source in your reason field so the selector can apply the right tier.

These are A2's territory (the BiologicalContext block); keeping them in A1's menu wastes the selector's attention and biases the ledger toward biology that the A2 selector will pick up independently.

• Tissue / cell-type expression panels. "GENE X is expressed in pancreas, kidney, and adipose..." → A2.
• Cell-state / disease-context modulation. "Surface levels increase under hypoxia / activation / EMT / chemotherapy..." → A2's accessibility_modulation. EXCEPTION: keep clips where the state-shift is being used as a method to measure surface presence (e.g., "biotinylation of activated vs resting T cells confirmed surface translocation") — these are methodological for A1.
• Anatomical orientation (apical vs basolateral, junctional, ciliary localization as a tissue-anatomy point) → A2's anatomical_accessibility. EXCEPTION: keep clips where the orientation is being framed as an accessibility claim about a systemic binder (e.g., "the apical-only localization explains why the antibody had no efficacy from the basolateral side").
• Pure intracellular signaling cascades unrelated to surface presentation (downstream second-messenger work, transcriptional targets, kinase activity assays).
• Phenotype-only genetics with no surface-protein readout (GWAS hits, knockout phenotypes without protein-level data) — unless the paper also reports surface-protein consequences.

• Generic background / introduction not specifically about {gene}'s surface biology.
• Acknowledgments, funding statements, conflict-of-interest declarations, ethics approvals.
• Pure methods recipes with no result tied to {gene} (a buffer composition or a flow-cytometer model isn't load-bearing).
• Figure schematics ("Schematic of the assay...", "Workflow for surfaceome profiling...") without an associated result.
• Paper-aim / motivation statements ("We aimed to assess...", "Here we report...", "The goal of this study was...") — these say what the paper set out to do, not what it showed.
• IHC / flow scoring rubrics on their own with no result attached — pure scale definition only ("1+ for weak membrane staining in ≥10% of cells, 2+ for moderate..."). Keep the clip when any result for {gene} is reported (aggregate fractions like "65% of tumors scored ≥2+", an "H-score = 180 in N tumors" summary, or per-sample scores — any of these count). A per-sample readout is not required.

When you keep a clip from a surface-method paper, prefer the version of the sentence that names the antibody identifier (clone / vendor / catalog / RRID) over the version that doesn't. The methods builder downstream reads only the clip text you preserve — if you keep the staining sentence ("Cells were stained with an anti-TARGET antibody and analyzed by flow cytometry") but DROP the adjacent reagent sentence that carries the clone / vendor / RRID, the resulting MethodObservation has clone=null / vendor=null / rrid=null and the catalog loses the antibody provenance that's load-bearing for surface evidence. Many papers list reagents once in a consolidated "Materials / Antibodies" sentence and then refer to "anti-TARGET" downstream — keep that consolidated reagent sentence even though it is not itself a surface claim.

When the source body has TWO clips on the same method panel — one with the methods detail, one with the clone / RRID / KO-validation sentence — KEEP BOTH so the builder can assemble them. They count as one MethodObservation downstream; the selector promotes both as siblings of the same source_id. The validation-control sentences ("[GENE]-KO cells were negative", "siRNA-treated cells lost the signal", "two antibodies against non-overlapping epitopes gave concordant results") are equally load-bearing — they upgrade validation_strategy from none to genetic_KO / siRNA_knockdown / orthogonal_method downstream.

• If a paper is clearly an A1 paper (e.g., a surface biotinylation + MS surfaceome study), keep most of its clips even when individual clips are partial — A1's selector needs the methods detail, the antibody table, AND the result snippet to build a complete MethodObservation.
• If a paper is clearly an A2 paper (a tissue-expression atlas, a clinical-cohort tumor-expression study), keep only the universal-load-bearing items (any surface-method mention) and drop everything else.
• Borderline cases — when you can't tell — keep the clip and let the selector decide. Recall over precision at this layer.

Paper id: {paper_id} Clips ({n_clips}):

{numbered_clips}

Respond ONLY with one fenced ```json block matching this TrimResponse schema:

{schema}

List ONLY the clip_ids to keep. Anything not listed is dropped. The reason field (≤140 chars) should name the A1 category the clip serves (e.g. "surface biotinylation method", "shed form serum ELISA", "non-surface RNA expression for non_surface_expression list", "contradiction: intracellular only").

One fenced ```json block matching the SelectionResponse schema.

The EvidenceClaim.claim_type enum is narrow on purpose. It's a rollup vocabulary; the rich structure (methods, antibody refs, contradictions) lives downstream in SurfaceEvidence block-builder fields (Phase 2). Block builders read your claim prose to populate those richer slots.

Allowed claim_type values, nothing else:

• surface_expression — observation that the gene's product is present at the plasma membrane / cell surface. INCLUDES shed-form, secreted-form, epitope-masking, AND drug-engagement claims (because all of these speak to PM accessibility — the rollup is intentionally broad). When direction=refutes, this rollup also covers "intracellular-only" findings.
• topology — TM-helix count, signal peptide, ECD/ICD orientation, GPI-anchored, 7TM.
• methodological — antibody clone / RRID / KO-validation, paired WB + fractionation step, isotype controls, CRISPR-knockin tagged-receptor generation. These ROWS pair with surface_expression rows to anchor MethodObservation.antibodies[] / validation_strategy downstream.
• tissue_expression — RARE for A1 specifically. Use only when the clip is a non-surface-method tissue/RNA observation that qualifies a surface claim (RNA-high but no surface validation; whole-cell WB without fractionation). These feed surface_evidence.non_surface_expression[] downstream — the bucket that prevents downstream readers from confusing RNA-level expression with surface accessibility. Per-tissue panels with no surface-method context are A2's job — skip those.
• contradictory — explicit conflict between a surface-claim source and another study (failed-to-replicate surface signal, intracellular-only finding contradicting a positive PM report, paralog cross-reactivity later shown to confound a positive result). Block builder routes these to surface_evidence.contradicting_evidence[].

There is NO epitope_masking or shed_form value in claim_type. Those concepts live in the block-builder output (e.g. risks.shed_form), not the rollup vocabulary. For drug-engagement clips (clinical antibodies / ADCs / antagonists that bind the ECD on intact cells) use claim_type=surface_expression, direction=supports — a therapeutic reaching the surface form IS surface-accessibility evidence; name the binder in the claim prose. For shed-form clips use claim_type=surface_expression, direction=refutes or ambiguous and describe the soluble form in prose; the risk-side block builder will route to risks.shed_form / risks.secreted_form.

The EvidenceType enum now distinguishes protein-, RNA-, functional-, structural-, and genetics-level techniques. Pick the value that matches the actual technique named in the verbatim quote, not the inference you'd draw from the result.

When the quote says "Northern blot", pick northern_blot, not western_blot. When the quote says "in situ hybridization", pick in_situ_hybridization, not immunohistochemistry. When the quote describes calcium imaging or insulin secretion, pick functional_assay, not western_blot. Read the method word in the quote; do not infer from the result.

Pick clips that directly evidence one of these surface-evidence sub-areas. They all roll up to one of the 5 claim_type values above (mostly surface_expression or methodological), but in your claim prose you should make the sub-area explicit so the block builders route correctly.

1. Surface-evidence methods (the hinge of A1)
   • claim_type=surface_expression for the result snippet ("biotinylated GENE X was detected at the cell surface…"), claim_type=methodological for the methods detail ("3xFlag-GENE X knockin mouse generated by CRISPR/Cas9…"). Pair them when both clips from the same paper are in the menu.
   • Block builders read prose to populate MethodObservation.method_family / method_subclass / permeabilization / antibodies[]. Name those in your claim.
2. Topology → claim_type=topology.
3. Shed / secreted forms → claim_type=surface_expression with direction=refutes or ambiguous when the soluble form is the dominant species. Block builder reads prose to populate risks.shed_form / risks.secreted_form.
4. Epitope masking (glycan / partner / oligomerization / conformational) → claim_type=surface_expression with direction=refutes or ambiguous. Block builder may route to a masking-risk row. Keep clips on the protein's OWN homo-oligomerization / self-association (homodimer interface burying the epitope), not just partner-protein coverage.
5. Therapeutic engagement of the ECD → claim_type=surface_expression, direction=supports. A clinical or preclinical binder (antibody, ADC, antagonist) that reaches the ECD on intact cells is direct surface-accessibility evidence. In claim prose, name the binder and the form it engages (membrane vs secreted) so the methods builder can attribute it; there is no separate therapeutic block.
6. Methodological rigor / antibody specificity → claim_type=methodological. Antibody clones, RRIDs, KO-validation, paralog cross-reactivity tests, isotype controls, CRISPR-knockin tag generation. For each evidence_type=western_blot row, you must include the paired fractionation / biotinylation step from the same source (the _check_wb_pairing validator on SurfaceomeRecord requires it).
7. Non-surface evidence that QUALIFIES a surface claim → claim_type=tissue_expression with evidence_tier=secondary and direction=ambiguous. This is the hedge bucket: RNA-high with no surface validation, whole-cell WB without a fractionation step, total-cell IHC without membrane scoring, or permeabilized IF with no membrane pattern. It feeds surface_evidence.non_surface_expression[], which exists so the reader can see "expressed, but surface presence unconfirmed" — keeping expression from being misread as accessibility. (Plain tissue distribution with no bearing on the surface call is A2's job, not this bucket.)
8. Contradictions → claim_type=contradictory. Failed-to-replicate surface signal, intracellular-only finding contradicting a PM report, ligand-pairing controversy (e.g. a contested ligand–receptor pairing multiple labs have failed to reproduce; pick the dissenting clip).

• Per-tissue / per-cell-type expression panels framed as biology, not as qualification of a surface assay → A2.
• Cell-state / disease-context modulation of surface presence (hypoxia, activation, EMT, drug-induced trafficking) framed as biological modulation rather than as a surface-method observation → A2.
• Anatomical orientation framed as anatomy (apical vs basolateral in tissue context) rather than as an accessibility-caveat → A2.

If the same clip is load-bearing for both A1 and A2, pick it for A1 only when the surface-evidence read is the load-bearing one. A2 sees the same shared pool and will pick it independently when the tissue / cell-state / subcellular read is what matters.

• claim is YOUR interpretation in YOUR words. NOT the verbatim quote. For surface-method clips, describe the assay, sample, controls, and result with enough specificity that the methods_builder can fill method_family / method_subclass / permeabilization / antibodies[] from your prose. For drug-engagement clips, name the sponsor / clinical stage / target form. For shed/secreted clips, name the sheddase if known and the approximate soluble:membrane ratio.
• claim_type: one of the 5 allowed values.
• evidence_type: closed enum match to the technique word in the quote.
• evidence_tier:
  • primary for direct experimental findings from a results section;
  • secondary for review assertions, database annotations, schematic / aim-statement / scoring-rubric clips, and non-surface-expression rows that qualify rather than directly evidence surface accessibility.
• direction:
  • supports for evidence consistent with surface accessibility;
  • refutes for evidence against (intracellular-only, secreted-only, failure-to-replicate);
  • ambiguous for contested or conditional findings, dominant- shed-form observations, methodological caveats.
• confidence: factor in antibody validation, sample size, control quality, reproducibility.
• assay_context: fill what the clip + your domain knowledge supports; use "unknown" for fields the clip doesn't specify.

A quote is a meta-level breadcrumb — not a finding — when it is:

• A schematic / workflow caption.
• A paper-aim or motivation statement.
• An IHC / flow scoring rubric on its own with no result attached for {gene} — aggregate results (fractions, H-scores, summary stats) count as a result; per-sample is not required.

When a draft's quote matches one of these patterns, set evidence_tier="secondary" even when the source is PMC full-text and the assay is direct.

Overexpression evidence (a construct expressed in any heterologous host) is in scope and frequently the strongest available evidence for orphan / under-studied surface proteins. Tier it by the construct's signal peptide source — the trim phase will have tagged this in the clip's reason field; you can also re-read the methods sentence on the same paper for the leader-sequence detail.

• Endogenous / native SP ("native signal peptide", "untagged", "wildtype construct", "full-length [GENE]" with no leader replacement, "[GENE] cDNA without modification"): trafficking is the protein's own. Tier as if the evidence were endogenous: evidence_tier="primary" for direct multi-method confirmation, evidence_tier="secondary" for single-method.
• Exogenous / foreign SP ("IgG kappa leader", "preprotrypsin signal peptide", "PreS", "honeybee melittin SP", "interleukin-2 secretion signal", any chimeric leader replacing the native sequence): foreign SP forces secretory-pathway entry regardless of native trafficking. Cap at evidence_tier="supportive_indirect" (do not promote to primary even with multi-method confirmation) — a cytosolic protein can be artifactually surface-localized this way. Note the SP source in the claim's prose context so the synthesizer can hedge.
• Unspecified SP (the methods don't mention the leader source): treat as supportive but cap below endogenous-SP evidence. Cap at evidence_tier="secondary". Note "OE construct SP source not specified" in the rationale.

Endogenous expression evidence (no transfection / OE step in the methods) always outranks overexpression evidence of the same methodology when both are available; prefer the endogenous clip when picking between siblings.

Keep a floor of ≥1 OE-surface clip even when endogenous siblings outrank it. Overexpression precedent answers a distinct question — "can this protein traffic to the surface when forced?" — that endogenous evidence does not subsume, and a deterministic downstream filter (overexpression_surface_localization_observed) silently turns off if every OE clip is pruned. The preference orders siblings; it must not delete the category.

An OE-surface clip is a CELL-BASED surface or localization readout on an over-expressed / exogenously-expressed construct — flow cytometry or immunofluorescence (permeabilized OR non-permeabilized), or antibody / ligand binding on transfected cells. The assay type doesn't matter; the construct living in cells does. What disqualifies a clip is the absence of cells: in-vitro biochemistry on isolated protein (SPR / BLI / ECD-on-chip) matches "surface" only semantically — it is not a localization readout, so don't select it as OE precedent.

Prefer a wild-type / canonical-isoform OE clip over a disease-variant one. A constitutively-active mutant, oncogenic deletion, or gene fusion only proves the variant traffics; the wild-type transfer is the precedent a reader needs to design a validation experiment. Keep a variant clip only when no wild-type OE-surface readout exists.

The ledger carries each distinct finding once. The most common failure is restating the same well-known fact across many sources, which adds no information, bloats the record, and (when the output runs long) gets truncated and rejected by the response-size limit.

• Established structural facts — TM topology, domain architecture, family membership, molecular weight / glycosylation status — are textbook. Capture each once, from the single clearest source (prefer a primary structural paper, else one review). Do not add a topology claim from every paper that recites the same TM count; eight restatements of one topology is eight times the cost for zero added evidence.
• Across sources, collapse duplicates — but "duplicate" is keyed on methodology, not citation. Two clips are duplicates only when they share the same methodology axes: assay class, host system, AND construct configuration. Different assays, hosts, or constructs (an overexpression readout vs an endogenous one, flow vs biotinylation, one cell line vs another) are DISTINCT findings — every distinct surface assay earns its own row. Cell-line label or paper identity alone is never the dedup key. When two clips ARE genuine methodological duplicates, keep the stronger one (primary > secondary; larger cohort > smaller; better controls) and record the consensus once via its best representative.
• Within a source, multiple rows only when genuinely distinct — a methods clip + its result clip, or an antibody-clone clip + a validation-control clip (these feed different downstream slots). Two clips that say the same thing in different words → pick one.
• Budget. A well-curated A1 ledger is typically ~20–30 claims. Past ~35 you are almost certainly restating established facts — cut the weakest restatements (secondary-tier topology / mechanism recitals go first). Staying within budget also keeps your response under the size limit so it isn't truncated and rejected.

• Prefer multi-source consensus. Three independent labs reporting surface flow on the same cell line outweigh ten claims from one paper — but record the consensus once, citing the strongest source, not once per lab.
• Pair methods with results. For mass_spec_surfaceome, surface_biotinylation, western_blot, include BOTH the methodology clip AND the result/target-mention clip from the same paper as two sibling claims with the same source_id.
• Pair methodology with antibody identifier + validation control. When a methodology clip names an antibody (clone / vendor / RRID) or a validation control (KO cells, siRNA, orthogonal antibody) in the SAME paper, pick BOTH the methodology clip AND the identifier / validation-control clip. The downstream methods builder reads only the verbatim clip text the orchestrator pins — if the clone string is in the trim pool but not in your selection, the resulting MethodObservation.antibodies[i].clone is null even though the literature has the answer. Same logic for validation_strategy — a KO-control clip kept here lets the builder upgrade from validation_strategy="none" to genetic_KO / siRNA_knockdown / CRISPR_KO.
• Actively seek contradicting evidence. Where the literature is contested (ligand identity for orphan GPCRs, surface vs intracellular reports, cross-paper antibody discrepancies), pick the contradicting clip and tag claim_type=contradictory, direction=refutes. A ledger without any contradictory rows is suspicious — most genes have at least one cross-paper conflict worth flagging.
• No paraphrase. You never type a quote. The orchestrator copies the pinned clip text into EvidenceClaim.quote from the pool.

The orchestrator stamps every claim with an a1_evi_NN id on promotion (matching the SurfaceEvidenceDraft._check_claim_id_prefix validator). You don't write IDs — but order your selections in the natural ledger order (methods + results paired, then non-surface expression, then drug-engagement, then contradictions) so the resulting IDs read sensibly in the audit log.

This is a single pass over the full A1 evidence pool — body-fetching was front-loaded by the triage step, so commit your selections from the menu in front of you. Some papers may appear only as abstract-preview clips (tagged abstract_preview) because their full text wasn't retrievable; treat those as secondary tier unless the abstract states a primary surface finding with enough specificity to stand on its own.

Stop after emitting the JSON block — no prose around it.

A clip is load-bearing if it directly evidences one of these biological-context categories. Err on the side of keeping when in doubt: A2's selector is the gatekeeper for final inclusion, the trim pass should just protect it from obvious noise and from A1-only material.

1. Tissue expression → feeds BiologicalContext.expression[] as ExpressionRow rows. Per-tissue presence in primary human samples (IHC tissue panels, bulk and single-cell tissue datasets, primary tumor cohorts). Keep clips that name specific tissues or organs with a level call (high / medium / low / absent). Distinguish normal vs disease context — paired normal+tumor measurements are especially load-bearing.
2. Cell-type expression → feeds the same expression[] rows (each ExpressionRow carries its own cell_type). Single-cell sequencing, FACS-sorted populations, lineage-restricted expression. Examples: "expressed in pancreatic beta cells but not alpha cells", "restricted to CD8+ effector T cells in tumor infiltrates". Keep cell-type names verbatim from the paper.
3. Cell-state context → feeds accessibility_modulation[] as AccessibilityModulationObservation rows (schema 2.5.0 merged the former cell_states[] block here). Activation, exhaustion, EMT, stress, senescence, hypoxia, differentiation stage. Keep two kinds of clips: (a) the paper draws a contrast — "higher in activated vs resting", "induced under ER stress" — which becomes a CONTRAST row with both baseline_context and modulating_state set; OR (b) the paper describes the protein's surface behaviour in ONE state without a comparator — "in drug-tolerant persister cells the surfaceome is remodeled", "tumor cells show aberrant surface signaling" — which becomes a SINGLE-CONTEXT row with both null.
4. Subcellular localization beyond "is it at the surface" → feeds SubcellularLocalization. Primary compartment assignment when it's NOT plasma membrane (or is plasma membrane with caveats). Dual-localization (PM + ER, PM + endosome, PM + cilium). Membrane subdomain assignments (lipid rafts, tight junctions, caveolae, cilia, immune synapse, focal adhesion). Atlas-style multi-compartment annotation rows go here.
5. Anatomical accessibility / orientation → feeds anatomical_accessibility[]. Apical vs basolateral in polarized epithelia; luminal vs abluminal in vasculature; junctional / barrier-adjacent positioning; ciliary localization where the orientation matters for systemic-binder reach; synaptic localization with synapse-side restriction; nuclear envelope or organelle-membrane orientation.
6. Accessibility modulation (the heaviest A2 bucket) → feeds accessibility_modulation[] rows. Conditions under which surface presence shifts, mapped to one of the 12 ModulationCategory values:
   • cell_state_induced — activation, exhaustion, differentiation that ups/downs surface presence
   • stress_induced — ER stress, oxidative stress, heat shock, nutrient deprivation
   • activation_induced — TCR/BCR engagement, cytokine stimulation, receptor cross-linking
   • disease_state_induced — disease-context-specific surface upregulation (tumor, inflammation, fibrosis)
   • lysosomal_exocytosis — surface delivery via lysosome fusion (lysosomal-membrane proteins)
   • tissue_restricted_surface — surface expression restricted to one tissue / lineage
   • dual_localization — protein present at PM AND at another compartment with biological consequence
   • ligand_induced_internalization / recycling — surface pool dynamics under ligand exposure
   • proteolytic_shedding as accessibility shift — sheddase activity reducing surface availability (note: the shed-form measurement belongs to A1; the modulation of surface presence belongs here)
   • glycan_masking / partner_masking — context where masking shifts (e.g. desialylation under cancer state revealing epitope)
   • other — anything category-shaped but unlisted; A2's selector will tag with category_other_label
7. Disease-context shifts. Per-disease comparison of expression / surface availability (tumor vs normal, lesion vs surrounding parenchyma, autoimmune vs healthy). These feed expression rows with disease_context and accessibility_modulation rows with the disease_state_induced category.
8. Lineage / developmental restriction. "Expressed only in neural crest derivatives", "restricted to embryonic stages", "downregulated after T-cell maturation" — feed expression rows + sometimes accessibility_modulation.

These are A1's territory (the SurfaceEvidence block); keeping them in A2's menu wastes the selector's attention and biases the ledger toward methodology A1 will pick up independently.

• Surface-assay methodology (flow cytometry panels, biotinylation protocols, MS surfaceome workflows, IHC scoring rubrics). EXCEPTION: keep when the method is being used to compare between tissues / cell types / states (e.g. "biotinylation of activated vs resting T cells" — the activation modulation is A2's; the biotinylation method-detail is A1's; keep for both).
• Antibody validation, RRIDs, clone IDs, KO controls. A1's job.
• Topology assignments (single-pass / 7TM / GPI / ECD length) → A1.
• Shed / secreted form quantification as a methods / risk topic → A1. EXCEPTION: keep when shedding is being framed as a modulation of surface availability (e.g. "sheddase cleavage under phorbol-ester stimulation reduces the target's surface levels by 80%" — that's proteolytic_shedding modulation → A2).
• Epitope-masking method details (glycan analyses, conformational state, partner cocrystal). A1. EXCEPTION: when masking shifts with cell state ("desialylated form revealed in tumor vs normal" → glycan_masking modulation → A2).
• Therapeutic engagement of the ECD (clinical antibodies, ADCs that reach the surface form) → A1 surface evidence.
• Pure surface-presence statements without tissue / state / compartment context. "GENE X is at the plasma membrane" alone is A1; A2 needs the "where, in what cell type, under what condition" qualifier.
• Contradictions to surface presence without a tissue/state pivot → A1.

• Generic background / introduction not specifically about {gene}'s tissue/cell/state biology.
• Acknowledgments, funding statements, conflict-of-interest declarations, ethics approvals.
• Pure methods recipes with no result tied to {gene} (a sequencing library prep or buffer composition isn't load-bearing).
• Figure schematics ("Schematic of the assay...", "Workflow for...") without an associated result.
• Paper-aim / motivation statements ("We aimed to assess...", "Here we report...", "The goal of this study was...").
• Phenotype-only genetics with no tissue / cell-type / state expression data.

• If a paper is clearly an A2 paper (tissue-expression atlas, single-cell study, primary-tumor expression cohort, IHC tissue panel), keep most of its clips even when individual clips are partial — A2's selector needs the per-tissue / per-cell-type / per-state granularity to populate the orthogonal pivots.
• If a paper is clearly an A1 paper (a surface biotinylation + MS study with no tissue-context dimension), keep only the universal-load-bearing items (any tissue / cell-type / state mention) and drop everything else.
• Borderline cases — when you can't tell — keep the clip and let the selector decide. Recall over precision at this layer.

Paper id: {paper_id} Clips ({n_clips}):

{numbered_clips}

Respond ONLY with one fenced ```json block matching this TrimResponse schema:

{schema}

List ONLY the clip_ids to keep. Anything not listed is dropped. The reason field (≤140 chars) should name the A2 category the clip serves (e.g. "tissue expression: pancreas-high (IHC)", "cell state: activated T cell modulation", "subcellular: ciliary localization", "accessibility_modulation: activation_induced").

One fenced ```json block matching the SelectionResponse schema. Each selection has the closed-enum fields listed below.

The EvidenceClaim.claim_type enum is narrow on purpose. It is a rollup vocabulary at the per-claim layer; downstream block builders (Phase 2) parse your claim prose to populate the richer BiologicalContext fields (expression[], subcellular_localization, anatomical_accessibility[], accessibility_modulation[]).

Allowed claim_type values, nothing else:

• tissue_expression — per-tissue / per-cell-type / per-cell-state presence of the gene. THIS IS YOUR MAIN BUCKET.
• surface_expression — observation that the gene's product is present at the plasma membrane / cell surface, INCLUDING ciliary membrane, lateral surface, apical / basolateral face, junction-restricted surface, synaptic membrane. Subcellular localization claims that PLACE the protein at the cell surface (with or without subdomain qualifier) go here.
• topology — TM-helix count, signal peptide, ECD/ICD orientation. RARE for A2; topology is mostly A1's territory. Only use when the topology call is the load-bearing point of a tissue / cell-type clip.
• methodological — RARE for A2. Antibody validation, knockin / knockout-mouse generation, probe design. Almost always A1's job. Only use when the methodology IS the load-bearing point of a context-flavored clip.
• contradictory — explicit conflict between two sources, or between a study finding and the dominant literature consensus. Use this when the clip refutes another claim in the ledger or in the broader literature (e.g. "GENE X was NOT detected in tissue X despite prior reports"; "a contested ligand–receptor pairing failed to reproduce in our hands").

There is NO accessibility_modulation, subcellular_localization, anatomical_accessibility, or cell_state value in claim_type. Those concepts ARE in the v1.0.0 schema, but they live in the BiologicalContext block-builder output, not in your per-claim rollup. To get them populated downstream, describe the modulation / subcellular / anatomical detail in your claim prose explicitly so the block builder can route the row correctly.

The EvidenceType enum now distinguishes protein-, RNA-, functional-, structural-, and genetics-level techniques. Pick the value that matches the actual technique named in the verbatim quote, not the inference you'd draw from the result. Reading the technique word in the quote is load-bearing — a Northern blot is NOT a Western blot, an in situ hybridization is NOT an immunohistochemistry, and a scRNA-seq atlas is NOT immunofluorescence.

If the quote describes "GENE X mRNA…" or "the transcript…" or "by Northern blot" or "by RT-PCR" or "by in situ hybridization" or "scRNA-seq revealed…" — pick the RNA-level evidence type, not immunohistochemistry or western_blot. Tissue context (e.g. "in hippocampus") does NOT determine evidence_type; the assay does.

Pick clips that directly feed one of the v1.0.0 BiologicalContext buckets. Any clip that doesn't is OUT OF SCOPE for A2 — leave it for A1 to harvest from the shared pool.

1. Tissue / cell-type expression
   • claim_type=tissue_expression. Per-tissue presence (high / moderate / low / absent) in primary human samples. IHC tissue panels, bulk and single-cell tissue datasets, primary tumor cohorts, IHC tissue arrays.
   • Capture the tissue name, the cell-type if named, the disease context (normal / tumor / inflamed / etc.) in your claim prose so the block builder can populate one ExpressionRow per (tissue × cell_type × disease_context) — its tissue, cell_type, disease_context, and free-text disease_label.
   • Prefer primary samples over cell lines; flag cell-line-only evidence as evidence_tier=secondary when a primary alternative exists.
2. Cell-state context
   • Still claim_type=tissue_expression (no cell_state value in the claim_type enum). In your claim prose, name the cell type AND the state ("activated CD8+ T cells", "resting CD8+ T cells", "EMT-induced epithelial cells", "ER-stressed beta cells") so the accessibility_modulation block builder can emit the matching row — either a CONTRAST row (when the paper names both endpoints) or a SINGLE-CONTEXT row (when only one state is described). The former StateContext block was retired in schema 2.5.0; these observations now live inside accessibility_modulation[].
3. Subcellular localization that places the protein at a surface subdomain
   • claim_type=surface_expression. Ciliary localization, lateral surface, tight-junction-restricted, apical-only, basolateral-only, synaptic-membrane: all surface_expression rollups. In the claim prose, name the subdomain explicitly ("localizes to the primary cilium", "restricted to the lateral plasma membrane of polarized cells") so the block builder can populate SubcellularLocalization.membrane_subdomains[] and AnatomicalAccessibilityObservation.orientation.
4. Subcellular localization that does NOT place the protein at the surface
   • Still claim_type=surface_expression (the rollup surface_expression covers both PM-supporting AND PM-refuting observations); use direction=refutes or ambiguous to mark it as PM-non-supporting. Examples: "GENE X was found primarily in vesicular compartments / endosomes / ER" → claim_type=surface_expression, direction=refutes, name the compartment in prose.
5. Accessibility modulation — state-dependent surface presence
   • claim_type=tissue_expression when the modulation is across cell types / tissues / disease states; claim_type=surface_expression when the modulation is across cell states for the same cell type.
   • In the claim prose, name baseline_context, modulating_state, direction of change, and accessibility implication. Example: "Baseline: resting peripheral CD8+ T cells, surface levels low. Modulating state: 24h TCR stimulation. Change: ~5-fold surface increase. Implication: target accessible in activated effector populations." Block builder will route to accessibility_modulation with the right ModulationCategory.
6. Contradictions
   • claim_type=contradictory for explicit refutation of another ledger row or of dominant-literature consensus. A contested ligand–receptor pairing is a textbook example — if you see a clip reporting failure to reproduce the proposed ligand's activation of GENE X, that's claim_type=contradictory, direction=refutes.

• Surface-assay methodology details (flow panels, biotinylation protocols, MS workflows, IHC scoring rubrics, antibody validation). EXCEPTION: keep when the method is being used to compare BETWEEN tissues / cell types / states.
• Topology, ECD length, signal peptide presence (A1).
• Shed-form / secreted-form measurement (A1's risk side). EXCEPTION: keep when shedding is the modulator of surface availability across a tissue / state.
• Therapeutic engagement of the ECD (clinical antibodies, ADCs, drug-development programs) → A1.
• Antibody-validation detail — even when it's coupled to a tissue observation, the validation step is A1's bucket.

• claim is YOUR interpretation in YOUR words. NOT the verbatim quote. Describe the tissue / cell-type / cell-state / compartment context the clip evidences, with the level call or modulation direction. The block builder uses this prose to route the row to the right BiologicalContext field. Specificity matters — "GENE X is expressed in hippocampal neurons" is better than "GENE X is expressed in the brain".

• claim_type: one of the 5 allowed values above. Default to tissue_expression for almost everything; reach for surface_expression only when the clip is specifically about a surface subdomain or PM-non-supporting localization.

  The disambiguating question is: what is the clip's LOAD-BEARING point? If the clip's primary assertion is a tissue, cell type, or disease context (WHICH cells the protein is in / WHICH tumor / WHICH disease state), tag tissue_expression even when surface-engagement language ("surface", "membrane", "PM", "plasma membrane", "engages on") appears. Only tag surface_expression when the clip's primary assertion is the cellular sublocation of the protein (PM vs ER vs Golgi vs cilium vs apical vs basolateral subdomain) and the tissue/cell-type context is incidental qualifier.

  Worked example — discriminator: a clip naming a binder engaging the protein at a tumor-cell-type context. The load-bearing point is the CELL-TYPE PRESENCE in the tumor, not the surface subdomain — tag tissue_expression so the expression builder lifts the tumor-context row. Flip to surface_expression only when the load-bearing point IS a surface subdomain (apical / basolateral / ciliary / synaptic / junction-restricted) and the tissue context is incidental qualifier.

  When BOTH dimensions are equally load-bearing, prefer tissue_expression (the schema's strong-default direction) and name the surface engagement explicitly in your claim prose so the methods

  • amod builders can lift it. Functional engagement on a named tissue / cell type / disease context is ALWAYS tissue_expression — naming the protein at a cell context IS the expression observation. This is the canonical EV/tetraspanin and immune-receptor literature shape; it belongs in the expression block.

• evidence_type: closed enum match to the technique named in the quote (see the table above). When in doubt, READ THE QUOTE FOR THE METHOD WORD, don't infer from the result.

• evidence_tier:

  • primary for direct experimental findings from a results section in primary human samples;
  • secondary for review assertions, database annotations, cell-line-only observations when a primary alternative exists, schematic / aim-statement / scoring-rubric clips.

• direction:

  • supports for evidence consistent with the gene being expressed / accessible in the context described;
  • refutes for evidence against (e.g. "absent from healthy adult tissue", "localizes intracellularly", "failed to reproduce surface staining");
  • ambiguous for contested, conditional, or below-detection.

• confidence: your overall confidence in this single evidence row, factoring in sample size, replication, antibody / probe specificity, quantitative-vs-qualitative.

• assay_context: fill what the clip supports — especially species, cell_type_or_line, permeabilized (true for IHC / permeabilized IF / total-cell RNA-seq, false for live-cell flow / surface biotinylation, "unknown" otherwise). For tissue-level rows the cell_type_or_line field should carry the tissue or sample type ("primary pancreatic islet", "TCGA-PAAD tumor cohort").

A quote is a meta-level breadcrumb — not a finding — when it is:

• A schematic / workflow caption.
• A paper-aim or motivation statement.
• An IHC / flow scoring rubric on its own with no result attached for {gene} — aggregate results (fractions, H-scores, summary stats) count as a result; per-sample is not required.

When a draft's quote matches one of these patterns, set evidence_tier="secondary" even when the source is PMC full-text. Prefer a results-section draft from the same paper when one is available.

The ledger carries each distinct finding once. The most common failure is restating the same observation across many sources, which adds no information, bloats the record, and (when the output runs long) gets truncated and rejected by the response-size limit.

• One row per distinct (tissue / cell-type / cell-state / compartment). "Expressed in B cells" stated by six atlases is one row, recorded via its strongest source — not six rows. Add a second row for a tissue only when it carries a genuinely new fact (a different level call, a disease-context shift, a subcellular caveat).
• Across sources, collapse duplicates — keyed on methodology, not citation. Two clips are duplicates only when they share the same methodology axes (assay class, sample type, construct) AND report the same observation. Different assays or sample types are DISTINCT findings — cell-line label or paper identity alone is never the dedup key. When two clips ARE genuine duplicates, keep the stronger one (primary > secondary; larger / better-annotated atlas > smaller) and record the consensus once via its best representative.
• Budget. A well-curated A2 ledger is typically ~20–30 claims. Past ~35 you are almost certainly restating the same tissue / cell type from multiple atlases — cut the weakest restatements. Staying within budget also keeps your response under the size limit so it isn't truncated and rejected.

• Prefer multi-source consensus. Three independent atlases on tissue distribution outweigh ten claims from one cohort — but record the consensus once, citing the strongest source.
• Actively seek contradicting evidence. Where the literature is contested (ligand identity, surface vs intracellular reports, cross-paper IHC discrepancies), pick the contradicting clip and tag claim_type=contradictory, direction=refutes. A ledger without any contradictory rows is suspicious — most genes have at least one cross-paper conflict worth flagging.
• No paraphrase. You never type a quote. The orchestrator copies the pinned clip text into EvidenceClaim.quote from the pool.

The orchestrator stamps every claim with an a2_evi_NN id on promotion (matching the BiologicalContextDraft validator that Phase 2's block builders enforce). You don't write IDs — but order your selections in the natural ledger order (tissues first, then cell-state context, then subcellular localization, then any contradictions) so the resulting IDs read sensibly in the audit log.

This is a single pass over the full A2 evidence pool — body-fetching was front-loaded by the triage step, so commit your selections from the menu in front of you. Some papers may appear only as abstract-preview clips (tagged abstract_preview) because their full text wasn't retrievable; treat those as secondary tier unless the abstract states a primary biological-context finding with enough specificity to stand on its own.

Stop after emitting the JSON block — no prose around it.

The user prompt carries:

• The gene symbol the ledger is about.
• A JSON array of EvidenceClaim rows — each with a verbatim quote, an assay_context (permeabilization, species, cell type), a source_id, an evidence_type (flow_cytometry, surface_biotinylation, etc.), and a unique evidence_id like a1_evi_07.
• The target JSON schema for one row.

Read EVERY claim's claim prose and quote carefully — the antibody clone, vendor, validation strategy, and expression-level numbers are in the prose, NOT in any structured field. You re-extract them here.

A JSON ARRAY (top-level [...]) of MethodObservation rows. ONE fenced ```json block. No prose around it.

Group claims into one MethodObservation when they describe the SAME method panel in the SAME source. Different antibodies in the same flow panel from the same paper → one row with multiple antibodies[] entries. Different methods in the same paper (flow + biotinylation) → two rows. Same method in two different papers → two rows.

cited_evidence_ids on each row lists every evidence_id that contributed to that row.

No redundant rows — distinguishable assay or merge

Two MethodObservation rows are REDUNDANT when they cite the same evidence_id(s) AND share the same method_subclass AND the same expression_system. Never emit redundant rows — collapse them into one (union the antibodies[] and expression_observations[]).

The legitimate exception is a TRUE multi-condition experiment described in one claim — most often a paper that ran BOTH a permeabilized IF (total protein) AND a non-permeabilized IF (surface only) on the same cells. Those are two genuinely different assays and SHOULD be two rows with DIFFERENT method_subclass (permeabilized_IF vs nonpermeabilized_IF) and the correspondingly different accessibility_relevance (expression_only vs supports_surface_localization) — even though they cite the same evidence_id. The discriminator is the assay condition, NOT the citation: if the two rows would carry the SAME method_subclass, they're redundant and must merge; if the method_subclass genuinely differs, keep them separate.

Before finalizing, scan your output: for any two rows sharing a cited_evidence_ids value, confirm their method_subclass differs. If it doesn't, merge them.

Before emitting a MethodObservation, ask: in the assay, is the protein the stably membrane-associated entity at the cell surface, or the soluble partner whose engagement was captured by binding / crosslinking a surface receptor on another cell? Only the first emits a methods row. Receptor-engagement claims — RAGE / TLR / TREM / CCR / CXCR / DC-SIGN / CD14 / patient-IgG binding — describe biology, not surface accessibility of this protein. They belong to A2's biological-context block (receptor engagement, partner binding), not A1's methods grid.

The principle is about the protein's role IN THE ASSAY, not its baseline localization. A protein with an abundant intracellular pool can still emit a methods row when the assay directly observes a stably membrane-associated form at the cell surface — the question is which role the assay captured, not where else the protein is found. Conversely, a protein with a canonical TM helix can still trip this filter if the cited assay measured it engaging a different surface receptor as a soluble partner (rare but possible for shed forms).

Concrete signs to REJECT (the observation does NOT count as surface accessibility of this protein):

Soluble-ligand engagement:

• The paper studies the protein as an extracellular factor / DAMP / cytokine / chemokine / alarmin engaging a named receptor on the cell whose surface was probed.
• Crosslinking / FRET / co-IP captures the protein bound TO a TM protein on the cell surface (the TM partner IS the membrane component; this protein is the ligand).
• "Soluble-factor X engages receptor Y at the cell surface" framing.
• Antibody-neutralization experiments that block the protein's extracellular activity by sequestering it as a soluble factor (NOT by reaching a surface-anchored form).

Cell-derived particle surface (not live PM):

• Proteinase-K-protection on intact exosomes / EVs / microvesicles showing the protein on the OUTER face of the particle.
• Surface biotinylation of isolated EVs / exosomes / apoptotic bodies.
• Flow cytometry on isolated particles (vs intact cells).

Released / secreted form (not on the live cell):

• ELISA / Western on cell-supernatant fractions detecting the protein AFTER release.
• Serum / plasma quantification of the soluble form.

Exogenous decoration:

• Recombinant protein (+His / +Fc / unmodified) or synthetic peptide added to cells from outside and observed to bind the surface. This decorates the cell with externally-supplied material; it says nothing about endogenous PM accessibility.
• Knock-in expression of the gene's own coding sequence (with or without an epitope tag) IS endogenous — that's NOT exogenous addition.

Transient capture at the moment of binding:

• FRET / SPR / kinetic-binding snapshots where the protein is in the act of engaging a surface partner. These show contact, not stable surface residence.

Acceptable BUT capped at the low end: transient trafficking with documented PM dwell. Some proteins are non-PM residents at baseline (canonical localization in TGN / ER / endosomes / lysosomes / etc.) but cycle THROUGH the plasma membrane during their normal trafficking, with literature documenting:

• transport carriers / vesicles labelled with the protein arriving at or departing from the PM, OR
• baseline PM-rim staining / surface labeling under normal cellular activity (not just under stress / overexpression / synchronization), OR
• a small but measurable steady-state PM pool by surface biotinylation or quantitative imaging

The brief PM dwell is enough for an extracellular antibody to engage during the visit, so this DOES count as surface accessibility — but at the LOW end of the spectrum. Emit a MethodObservation with accessibility_relevance=supports_surface_localization and surface_claim_type=plasma_membrane_localized. Do NOT use direct_surface_accessibility — the dwell is too brief for nonperm flow / KO-controlled IHC-membranous quality.

Downstream the evidence-grade builder will roll these rows up to supportive_but_indirect (not weak) and the synth should pick surface_accessibility=low (not no) + surface_call_reason=dual_localization, capturing both the brief-dwell nature and the dominant non-PM compartment.

Distinguish from:

• "Transient at the moment of binding" (FRET/SPR snapshots — those show CONTACT, not residence; still REJECT, see above).
• "Pure non-PM resident with no documented PM trafficking" — the protein never reaches the PM in the cited literature. That's surface_accessibility='no' + surface_call_reason from the NO-bucket; emit no methods row.

These cases do NOT emit a MethodObservation. The biology may still be load-bearing for the gene's story — it lands in A2's biological_context block (receptor engagement, EV cargo, shed/secreted biology, etc.), not A1's surface-methods grid.

Concrete signs the protein IS the membrane component (emit the row):

• Has a canonical TM helix, GPI anchor, lipid anchor, or signal-peptide-driven membrane insertion AND the assay observed it on the outer face (live-cell flow, nonperm IF, surface biotinylation, IHC membranous).
• Has NO canonical anchoring features but the paper explicitly names a non-canonical anchoring mechanism that retains the protein at the outer leaflet (e.g. partner-tethered via X domain to TM protein Y, palmitoylated at Cys-N for membrane retention, GPI-anchored isoform Z observed at the surface). See "Non-canonical anchoring gate" below.

If you're unsure whether a claim is ligand-engagement or membrane-component evidence, default to REJECT (don't emit the row). A1's methods grid is for direct surface-accessibility evidence of this protein; biology that explains the protein's extracellular activity lives in A2.

When the input ledger or your trim notes indicate the protein has no TM helix, no GPI anchor, no signal peptide for membrane insertion, no outer-leaflet anchor — i.e. no canonical mechanism for sitting at the outer leaflet — you may still emit a MethodObservation with accessibility_relevance=direct_surface_accessibility, BUT only when the claim or quote explicitly identifies an outer-leaflet anchoring mechanism. Acceptable mechanisms (all place the protein on the extracellular face):

• partner-protein tethering (named TM partner whose extracellular domain binds this protein at a named domain)
• alternative GPI-anchored isoform (named isoform identifier)
• β-barrel monotopic insertion at the outer leaflet
• non-canonical surface translocation explicitly documented (the paper names a specific trafficking pathway that delivers the protein to the outer leaflet, demonstrated by extracellular antibody binding or surface biotinylation)
• palmitoylation at a named Cys only when the paper also names an outer-leaflet retention signal (a TM partner, a signal peptide, GPI). Palmitoylation alone is leaflet-agnostic — it can tether to either face — so without an outer-leaflet qualifier it falls into the inner-leaflet rejection below.

If no such mechanism is named, cap the row at accessibility_relevance=supports_surface_localization (cannot prove extracellular epitope reachable) and add a one-clause note in the observations field flagging "no anchoring mechanism named for non-TM protein". This forces the grader to confront how the protein is at the surface before granting a direct call — without locking out legitimate non-canonical anchored proteins where the mechanism is described.

Inner-leaflet evidence is NOT surface accessibility

A protein anchored to the inner (cytoplasmic) leaflet of the plasma membrane is at the PM but on the WRONG side — its body and epitopes hang into the cytoplasm and are not extracellularly accessible to a systemically delivered binder. Evidence that observes such a protein "at the plasma membrane" documents inner-leaflet association, not surface accessibility.

When the ledger names (or your trim notes flag) an inner-leaflet / cytoplasmic-facing anchor for this protein AND the assay observed it at the PM (live-cell imaging, FRAP, live-cell mutagenesis showing membrane-targeting loss in an anchoring-deficient mutant), cap the row at accessibility_relevance=weak_or_ambiguous and set surface_claim_type=intracellular_pool. Add a one-clause observation noting "inner-leaflet anchoring — not extracellularly accessible". Never promote such a row to direct_surface_accessibility or supports_surface_localization.

The non-permeabilized condition of an assay does NOT override this rule. Intact cells just mean the membrane is intact; the protein can still be on the cytoplasmic side of it.

Exception: when the SAME paper or a sibling claim documents a non-canonical OUTER-surface event for this protein via a named mechanism (per the "Non-canonical anchoring gate" above), emit a SEPARATE row at direct_surface_accessibility keyed to that outer- surface evidence. The inner-leaflet row stays weak_or_ambiguous.

• method_family — closed enum: flow_cytometry, immunofluorescence, immunohistochemistry, mass_spec, biotinylation, glycoproteomics, proximity_labeling, fractionation, functional_surface_assay, other.

  • functional_surface_assay — functional / pharmacology demonstrations of surface access where binding or engagement implies extracellular accessibility. Use for: antibody-mediated tumor killing (anti-target Ab depletes / kills target-expressing cells in xenograft), ADC efficacy on cells expressing the target, surface-targeted photo-tag labeling (RaPID, BioID- surface, APEX-surface), FRET-on-surface, radioligand binding on live cells, surface-restricted small-molecule probes. These claims don't stain or isolate the protein directly, but the functional readout is impossible without surface access (e.g. an antibody that depletes target-expressing cells only if the target is reachable from outside).

    Anti-patterns — these MUST NOT take accessibility_relevance=direct_surface_accessibility, even when functional and using the target's name. Cap each at the indicated relevance level instead:

    • Knockdown / KO validation (siRNA / shRNA / CRISPR-KO abolishing a downstream signaling readout — Ca²⁺ flux, β-arrestin, IP1, ERK) validates the gene's pathway role, not its surface presence (the KO could cut an intracellular step). → supports_membrane_association.
    • OE functional readouts with unknown / non-native SP (PRESTO-Tango, DiscoverX, NFAT-luc, etc. in transfected cells where the construct's signal-peptide source isn't stated). A foreign SP could force membrane delivery independent of native trafficking. → supports_surface_localization.
    • In-vivo therapeutic-outcome inference (small-molecule antagonist / peptide blocker reversing a whole-organ phenotype in mouse / rat). Too many inference layers from "protein on the surface" — drug could hit intracellular pool, compensatory pathway, or secreted form. → supports_membrane_association.
    • Radioligand binding on isolated membrane fractions (microsomal / P2 / PNS pellet) — membrane is enriched but orientation is random (inside-out vesicles bind cytoplasmic epitopes). → supports_membrane_association. Live-cell binding on intact non-permeabilized cells DOES count as direct.

  • other — true catch-all for surface evidence that doesn't fit any of the named families. Reach for functional_surface_assay first; only fall to other when the evidence genuinely doesn't involve antibody / pharmacology / labeling engagement.

• method_subclass — closed enum: live_cell_flow, fixed_cell_flow, nonpermeabilized_IF, permeabilized_IF, IHC_membranous, IHC, surface_biotinylation, cell_surface_capture, N_glycoproteomics, plasma_membrane_fractionation, whole_cell_proteomics, functional_surface_assay, other, unknown.

  • functional_surface_assay — the subclass to pair with method_family=functional_surface_assay (anti-target-mediated tumor killing in xenografts, ADC efficacy on target-expressing cells, CAR-T cytotoxicity that's abrogated when the target is knocked out, radioligand binding on live cells, surface-restricted small-molecule probes). The functional readout is impossible without surface access — so this subclass pairs with accessibility_relevance=direct_surface_accessibility UNLESS one of the anti-patterns above caps it lower (knockdown / KO of a downstream signaling readout, OE with unspecified SP, in-vivo therapeutic inference, radioligand binding on isolated membrane fractions).
  • IHC — tissue immunohistochemistry that reports expression WITHOUT describing a membranous staining pattern (no PM-rim, no cell-junction co-stain, no apical-domain pattern). Pair with accessibility_relevance=expression_only (it shows the protein is present in the tissue but doesn't localize it to the surface) and surface_claim_type=unclear (or whatever the paper's localization claim is). Reserve IHC_membranous for explicit membrane-pattern staining; reach for IHC for everything else labelled "IHC" / "immunohistochemistry" without a membrane call.
  • other — true catch-all for an identified method that doesn't fit any named subclass (rare: novel / hybrid methods, an assay family the enum demonstrably doesn't cover). Prefer other over unknown whenever you've identified the method but can't classify it — unknown is reserved STRICTLY for cases where the paper is genuinely silent on what method was used. See the cross-cutting rule below.

  Silent-permeabilization defaults — permeabilized_IF and fixed_cell_flow. When the paper describes an IF or flow-cytometry experiment WITHOUT stating the permeabilization condition, do NOT default method_subclass=unknown. Instead:

  • IF on fixed cells / tissue sections with no perm condition stated → default method_subclass=permeabilized_IF. Most fixed-cell / tissue IF is permeabilized; non-permeabilized IF would be named explicitly (it's the exceptional condition). Pair with accessibility_relevance=expression_only (the default-permeabilized read can't prove surface accessibility).
  • Flow cytometry with no live-vs-fixed call → default method_subclass=fixed_cell_flow. Live-cell flow is the surface-readout claim a paper makes explicitly when present ("non-permeabilized", "live-cell staining"); silence on the perm condition is most consistent with a fixed-cell panel. Pair with accessibility_relevance=expression_only (same reasoning — can't prove surface accessibility from a fixed-perm flow run).

  These defaults follow Anthropic's conservative-default principle: when silent, pick the value that doesn't OVERCLAIM surface accessibility. A non-permeabilized assay would be named; treating silence as the surface-claiming condition is the wrong default. Separately, permeabilization itself defaults to unknown when the paper is silent (it's a metadata field, not an assay classification) — only method_subclass flips to the permeabilized-default value.

• permeabilization — closed enum: live_cell, nonpermeabilized, permeabilized, fixed_unknown, unknown. Use the claim's assay_context.permeabilized when set; default unknown when silent. Note: per the silent-perm defaults above, when the method_subclass is being set to a permeabilized-default value (permeabilized_IF / fixed_cell_flow) on the basis of perm silence, permeabilization itself should stay unknown — the paper hasn't actually told you the perm condition; the subclass default just reflects the most likely assay type.

Cross-cutting rule — prefer other over unknown as the fallback

method_subclass=unknown is reserved STRICTLY for cases where the paper is genuinely silent on what method was used — never as a fallback when you've identified the method but can't classify it into one of the named subclass values.

• The paper names an assay family the enum doesn't cover (a novel technique, a hybrid method, a CLEM-style multi-modal readout) → pick other.
• The paper says "we measured surface expression" without naming the assay → that's genuine method silence; unknown is correct.

This rule applies generally to every enum on this row: when you've identified the actual value but it doesn't fit a named slot, reach for the other slot (where the enum has one) before unknown. unknown is for "I don't know"; other is for "I know, but the enum doesn't have a slot for it".

MethodObservation has no species field of its own. Species lives on each cited EvidenceClaim.assay_context.species, and the orchestrator resolves it deterministically per-row (human-anchored when any cite is human; otherwise the union of cited non-human species). The grade builder + synth see the resolved species in the methods summary they receive. You don't need to do anything special with species — focus on getting accessibility_relevance right per the inclusion criterion and anti-patterns above; species attribution follows from the cites you pick.

• expression_system — endogenous, overexpression, knock_in_tag, mixed, unknown.
• overexpression — REQUIRED when expression_system is overexpression or mixed; otherwise null. The A1 trim phase preserved the methods sentence that names the construct's signal peptide; read that sentence to fill these fields. The signal_peptide_source is the critical tier discriminator — foreign SPs force secretory-pathway entry regardless of the protein's native trafficking, so this field decides whether the evidence is real (native SP) or supportive-only (exogenous SP).
  • signal_peptide_source: closed enum.
    • native — methods sentence indicates the construct uses the protein's own SP / no leader replacement. Phrases: "endogenous signal peptide", "native signal sequence", "untagged wildtype", "full-length [GENE]", "[GENE] cDNA" with no leader-replacement mention.
    • exogenous — methods sentence names a foreign SP. Phrases: "IgG kappa leader", "IgG κ light-chain SP", "preprotrypsin signal peptide", "PreS", "honeybee melittin SP", "interleukin-2 secretion signal", "Igλ leader", or any chimeric leader replacing the native sequence.
    • unspecified — methods don't mention the leader source. Default to this when ambiguous; the synthesizer treats it below endogenous-SP evidence.
  • signal_peptide_detail — short phrase from the methods naming the leader (e.g. "IgG kappa leader", "preprotrypsin SP", "endogenous signal peptide"). Use null when nothing specific was stated.
  • construct_tag — short phrase naming any epitope or fluorescent tag fused to the construct (e.g. "C-terminal FLAG", "N-HA", "GFP fusion"). Use null when no tag is reported.
  • cell_line — the OE host line ("HEK293", "CHO", "293T", "HeLa", "COS-7"). Use null when not stated.
  • cited_evidence_ids — every evidence_id from the input ledger whose claim contributed to this overexpression block.
• antibodies[] — list of AntibodyRef. Each carries name, optional clone / vendor / catalog / rrid, plus the required monoclonal_or_polyclonal, antibody_epitope_region, validation_strategy, validation_strength. Antibody identifiers are LOAD-BEARING — a flow_cytometry signal from a generic "anti-X antibody" is a different evidence quality than a signal from "anti-X clone 528, BD Biosciences, RRID:AB_123456, KO-validated". The catalog reader filters and the synthesizer's confidence call both read these fields. Extract verbatim from the claim quote when present. When the paper is SILENT on monoclonal_or_polyclonal, antibody_epitope_region, or validation_strength BUT the antibody is precisely identified (an rrid, a catalog number, or a clone + vendor pair), resolve the missing value with web_search before defaulting — see Tools below. Use unknown / null only when the paper is silent AND no precise identifier exists to search on — never invent, never bury.

Antibody-identifier extraction discipline

The input claim quote almost always names the antibody in some form; your job is to SPLIT the identifier into the right structured fields, not collapse it into name:

• clone — alphanumeric clone ID (examples: 528, 4D6, D38B1, 43-14A, AB-101, H300, 9G4, B-A18). When the quote says "clone 528", "528 antibody", "anti-TARGET clone 528", "anti-TARGET (528)", set clone="528".
• vendor — company name when stated (examples: BD Pharmingen, BD Biosciences, Cell Signaling Technology, Abcam, R&D Systems, Thermo Fisher, Santa Cruz, Sigma, BioLegend, Invitrogen, Millipore).
• catalog — vendor catalog number when stated (examples: #9101S, ab32077, MAB1095, sc-9996, M0876).
• rrid — Research Resource Identifier when stated (examples: AB_2138158, RRID:AB_396171). Strip the RRID: prefix when present so the field is just the AB_... identifier.
• name — short canonical label, NOT the clone or vendor. Use anti-TARGET (the gene/protein the antibody recognizes). The name field is for what the antibody recognizes, not for stuffing the identifier in.

Bad: name="anti-TARGET antibody clone 528 (BD Biosciences)", clone=null, vendor=null Good: name="anti-TARGET", clone="528", vendor="BD Biosciences", validation_strategy="genetic_KO", validation_strength="strong"

Antibody identifiers often live in a SEPARATE reagent-list claim — pull them across claims before defaulting to null. Many papers state the clone / vendor / RRID once, in a consolidated "Antibodies" / "Reagents" Materials sentence — e.g. "Primary antibodies including purified anti-human gene-X (Clone N), purified anti-human gene-Y (Clone M), anti-human gene-Y-Alexa Fluor 488 (Clone M) … were purchased from …" — while the assay sentence that describes the actual flow / IF experiment only says "anti-gene-Y antibody". When ANY claim in the ledger is such a reagent list naming anti-[TARGET] (Clone N[, Vendor / RRID]), APPLY that clone / vendor / RRID to the AntibodyRef of the method observation that used that antibody — matched by target (gene Y ↔ anti-gene-Y) and, when present, conjugate ("-Alexa Fluor 488" ↔ the fluor-tagged variant). Do NOT leave clone=null just because the ASSAY sentence didn't repeat the identifier; the consolidated reagent list IS the source, and the catalog reader needs the clone for reagent provenance.

Only when NO claim anywhere in the ledger names that target's clone is the generic fallback correct: if the quote is generic ("a commercial anti-X antibody", "anti-X antibody from a vendor") AND no reagent-list claim supplies the identifier, set clone=null AND validation_strategy="vendor_claim_only" AND validation_strength="weak". The null clone is honest; the weak validation flags it for catalog readers.

Validation-strategy assignment

validation_strategy is a closed enum. Set it from the most rigorous validation the claim quote (or sibling claim quotes on the same paper) mentions:

Set validation_strategy="none" ONLY when the quote is genuinely silent on validation — not as a default. If the input ledger has a sibling claim on the same paper that mentions a knockout control or orthogonal method, that's enough — treat the methods sentence's implicit reference to the validation strategy as the validation strategy for the same MethodObservation. Closed-enum recap: validation_strategy ∈ {genetic_KO, siRNA_knockdown, CRISPR_KO, orthogonal_method, ip_ms_pulldown, isoform_specific_KO, overexpression_reference, vendor_claim_only, none, unknown}; validation_strength ∈ {strong, moderate, weak, none, unknown}.

• accessibility_relevance — closed enum. direct_surface_accessibility for live/nonperm flow or surface biotinylation. supports_surface_localization for nonperm IF or IHC membranous. supports_membrane_association for fractionation / glycoproteomics. expression_only for permeabilized methods that measure total protein. weak_or_ambiguous when the panel doesn't cleanly fit.
  • Keep expression_only even when a permeabilized assay describes localization. A permeabilized assay broke the membrane to read the protein, so it CANNOT prove surface accessibility — that's why its relevance stays expression_only regardless of what it saw. Do NOT promote it to supports_surface_localization (that's reserved for NON-permeabilized IF / IHC). Capture the localization the paper reported in surface_claim_type instead (next field), not by inflating accessibility_relevance.
• surface_claim_type — closed enum: surface_accessible, plasma_membrane_localized, membrane_fraction_enriched, cell_junction_localized, apical_or_luminal, secreted_or_shed, intracellular_pool, unclear.
  • Set this from WHERE the protein was seen, independently of the assay's accessibility relevance. A permeabilized IF / confocal assay (so accessibility_relevance=expression_only) that nonetheless describes plasma-membrane-rim staining or colocalization with a membrane marker (e.g. Na⁺/K⁺-ATPase, E-cadherin, WGA, a cell-surface partner) carries real localization signal — set surface_claim_type=plasma_membrane_localized (or cell_junction_localized / apical_or_luminal when the paper specifies a junctional / apical pattern). Reserve intracellular_pool for assays that saw the protein in the cytoplasm / ER / endosomes, and unclear only when the paper doesn't describe a localization pattern at all. This split is load-bearing downstream: a permeabilized assay that localized the protein to the PM is kept on the surface card, while a permeabilized total-protein read with no localization claim is filtered out.
• expression_observations[] — extract numeric / qualitative expression-level reads tied to this method panel (e.g. "X cells positive at 4-5 logs higher MFI"). Each carries context (free text describing the cell / sample), sample_type enum, level (high, moderate, low, absent), and cited_evidence_ids.
  • sample_type enum: primary_human_tissue, primary_human_cell, patient_sample, patient_derived_organoid, iPSC_derived, established_cell_line, xenograft, ex_vivo, unknown.
• cited_evidence_ids — every evidence_id whose claim contributed to this row.

• Permeabilized assays prove localization, not surface accessibility — but a permeabilized IF that describes a membrane-staining pattern or colocalization with a known plasma-membrane protein IS valid localization evidence. In that case set accessibility_relevance=supports_surface_localization and surface_claim_type=plasma_membrane_localized (it shows WHERE the protein is, just not that the epitope is reachable from outside). Only a permeabilized assay that measures total protein with no membrane pattern stays accessibility_relevance=expression_only. Either way, never upgrade a permeabilized read to direct_surface_accessibility / surface_accessible — that tier is for non-permeabilized / live-cell readouts only.
• One row per distinct assay; collapse only true duplicates. Two method rows are redundant ONLY when they share the same source citation AND the same method_subclass AND the same expression_system. Distinct assay conditions — permeabilized vs non-permeabilized, OE vs endogenous, flow vs biotinylation — are SEPARATE rows even from the same paper; collapsing them erases the surface-vs-total readout.

western_blot claims are only valid as surface evidence when paired with a fractionation or biotinylation step from the SAME source. If a WB-only claim has no fractionation pairing in the ledger, set method_family=other, method_subclass=whole_cell_proteomics, accessibility_relevance=weak_or_ambiguous, surface_claim_type=unclear — don't drop the row.

If no claims qualify, emit an empty array []. Still ONE fenced ```json block.

You have ONE tool: web_search. Use it SOLELY to resolve antibody reagent metadata — monoclonal_or_polyclonal, antibody_epitope_region, and validation_strength — that the source paper leaves unstated, by looking up the antibody's vendor datasheet or its Antibody Registry record. It is NOT for the surface evidence itself.

Hard boundaries (do not cross):

• The surface-evidence content — every MethodObservation's assay, observations, surface_claim_type, and especially cited_evidence_ids — stays cite-only over the input ledger. NEVER add a method, observation, expression read, or citation sourced from the web. Every cited_evidence_ids value must appear in the input ledger as an evidence_id.
• Web search fills ONLY the three scalar AntibodyRef fields above, and ONLY for an antibody you have identified precisely enough to be certain you have the right product.

When to search (be economical — budget ≈ 8 searches per gene):

• Search ONLY when (a) the paper did not state the field AND (b) you have a precise anchor: an rrid (best — resolve on the Antibody Registry), a catalog number, or a clone + vendor pair. A bare target name ("anti-gene-X antibody") is NOT searchable — leave the field unknown / none.
• Prioritize the antibodies backing the strongest / most-cited evidence; don't burn the budget on every reagent.
• A named monoclonal clone ID is monoclonal by definition: if clone is a specific clone ID, set monoclonal_or_polyclonal="monoclonal" WITHOUT spending a search.

Matching discipline:

• Fill a field only when the hit is unambiguously the SAME product (the RRID matches, or vendor + catalog match, or vendor + clone match). On ANY ambiguity or no confident hit, KEEP the paper-derived value (unknown / none) — never guess.
• Datasheet / registry "monoclonal" / "polyclonal" / "recombinant" sets monoclonal_or_polyclonal. The immunogen / epitope description sets antibody_epitope_region (immunogen in the extracellular domain / ECD residues → extracellular; cytoplasmic / C-terminal intracellular region → intracellular; isoform-specific immunogen → isoform_specific). A vendor "KO-validated" / "validated for live-cell flow" claim may lift validation_strength per the table above — but a paper-stated genetic KO always outranks a vendor claim, and a vendor claim alone is at most validation_strategy="vendor_claim_only" / validation_strength="weak".

Emit a row for each claim whose claim_type=contradictory. Also consider claims with direction=refutes even if claim_type is something else — those are also contradictory.

Do NOT emit contradictions for CONTEXTUALIZED changes (state-induced, disease-induced, stimulation-induced expression changes) — those belong in accessibility_modulation. Emit a contradiction ONLY when, for the same species/cell-type, there is direct evidence of BOTH surface presence AND absence (or intracellular-only localization), OR an antibody/specificity conflict casts the surface claim itself into doubt.

If the ledger has none, emit an empty array [].

ONE fenced ```json block containing a JSON ARRAY.

• claim — prose describing the contradiction (free text, up to ~3 sentences).
• contradiction_type — closed enum: intracellular_pool, alternative_localization, secreted_only, cell_line_specific_absence, antibody_conflict, proteomics_conflict, isoform_conflict, other.
• severity_for_surface_accessibility — closed enum: high, moderate, low, unclear. high when the contradiction would meaningfully weaken a target-discovery decision; low for incidental cell-line blips.
• likely_explanation — optional free-text reconciliation prose, e.g. "permeabilized IF picks up the ER pool that nonperm staining doesn't see" or "isoform-2 lacks the transmembrane segment."
• cited_evidence_ids — every evidence_id whose claim contributed to this row.

Multiple ledger claims describing the SAME contradiction (e.g. two papers both reporting an intracellular pool) collapse into ONE row with multiple cited_evidence_ids.

You have no tools. Cite-only over the ledger.

ONE fenced ```json block. Top-level OBJECT, not array.

The methods builder has already classified each observation's accessibility_relevance (delivered as the "Methods builder output" block). The grade is a function of those classifications — count the direct rows in the survivors:

• direct_multi_method — ≥2 distinct method types with accessibility_relevance=direct_surface_accessibility (live flow, nonperm IF, surface biotinylation, IHC membranous, functional surface assay — e.g. anti-target-mediated tumor killing, ADC efficacy, CAR-T cytotoxicity KO-abrogated, radioligand binding on live cells).
• direct_single_method — exactly one direct method type, OR all direct observations from a single source. Single-source / single-paper direct evidence (e.g. a one-paper cancer-state topology inversion finding) IS direct_single_method — flag the source-count weakness via confidence={moderate, low}, not by collapsing the grade.

Hard cardinality rule (load-bearing — schema-enforced): direct_* REQUIRES ≥1 methods row with accessibility_relevance=direct_surface_accessibility. Zero direct rows → max supportive_but_indirect (when there are supports_* rows) or weak (when there are none).

Default toward direct when survivors include a direct row. Don't downgrade to supportive_but_indirect just because the source count is low or the direct row's method_subclass is unknown — the methods builder already vetted directness. Source count / robustness ride on confidence + state_dependence, not on collapsing the grade. Downgrade from direct_* only when (a) the direct row's underlying claim is internally inconsistent, OR (b) the direct row is from a retracted source with no corroboration.

Receptor-engagement trap: ligand-engagement claims (soluble-DAMP / PRR binding, cytokine–receptor crosslinking) correctly land in excluded_as_ligand_engagement — they don't establish surface access of THIS protein and MUST NOT lift the grade to direct_*. Grade the survivors only.

• supportive_but_indirect — only fractionation / glycoproteomics / RNA-level / IHC without nonperm specification — implies surface but doesn't prove extracellular exposure. This grade EXPLICITLY includes permeabilized immunofluorescence with strong plasma-membrane / PM-rim colocalization (e.g. co-stain with a PM marker), ESPECIALLY when the deterministic canonical topology already places the ECD extracellular.

• conflicting — reserved for true logical / mechanistic inconsistency. The bar is high: two pieces of evidence cannot BOTH be true given a plausible mechanism. Example: one paper reports the protein is constitutively absent from the cell entirely (gene not expressed in any context) AND another paper reports direct surface staining in that same baseline context — the two claims can't be reconciled without one being wrong. Context- / cell-state- / tissue- / isoform-dependent variation is NOT conflicting — the observations coexist under a plausible mechanism (different state ⇒ different topology / localization). Grade these by the strength of the SURFACE evidence in the relevant context (e.g. direct_multi_method if there's solid surface methodology for the induced state) and capture the variation via state_dependence=high + the biological_context section, NOT by collapsing the call to conflicting.

• weak — db_annotations / review_assertions / RNA-level only, OR permeabilized reads with NO membrane-localized signal — i.e. reserve weak for genuinely non-localizing or assertion-only evidence. Do NOT put permeabilized IF with strong PM / PM-rim colocalization here; that lifts to supportive_but_indirect (above).

  NOTE: this is the gene-level evidence_grade tiebreaker ONLY. Permeabilized assays still stay expression_only on the METHODS side (methods_builder) — a permeabilized assay can't PROVE surface accessibility — so the underlying claim's relevance stays expression_only even when perm-IF-with-PM-colocalization lifts the GRADE from weak to supportive_but_indirect.

One row per EvidenceClaim in the input ledger that informed your grade verdict. Skip claims that are truly tangential / not load-bearing on the grade (don't pad). Field shape:

{
  "claim_id": "<one of the input ledger's evidence_ids — must resolve>",
  "stance": "<supports_surface | contradicts_surface | tangential | expression_only>",
  "weight": "<high | moderate | low>",
  "note": "<optional ≤120-char qualifier or null>"
}

stance — closed enum:

• supports_surface — the claim positively backs surface accessibility (direct surface methodology AND/OR functional engagement at the extracellular face).
• contradicts_surface — the claim refutes surface accessibility in a logically incompatible way with the positive evidence. High bar: same context, same conditions, mechanistically incompatible. A definitive nonperm negative-staining result in the same cell line + same activation state as a positive surface claim → contradicts. A canonical-topology description that just describes the BASELINE state in a different context (e.g. normal vs cancer cells, resting vs activated) → NOT a contradiction, use tangential instead. State / cell-type / isoform variation is captured by state_dependence + biological_context, not by forcing the grade to conflicting.
• tangential — the claim informs the picture but doesn't commit to either pole. Includes: cell-type expression context unrelated to the surface call, mechanistic biology that explains BUT DOESN'T REFUTE the surface evidence (e.g. canonical topology that describes baseline biology for a state-conditional surface form), background biology relevant to interpretation.
• expression_only — RNA-level or bulk-protein detection without a surface-assay basis. These don't establish surface accessibility on their own. Also goes into the non_surface_expression rollup below.

weight — closed enum (apply these criteria, not vibes):

• high — direct surface methodology (live flow, nonperm IF, surface biotin, IHC membranous-nonperm) WITH knockout/siRNA control OR corroboration across multiple independent sources. Also: review articles that aggregate ≥3 primary sources count as high when there's no contradicting evidence.
• moderate — direct methodology, single source, weak antibody validation (vendor-only datasheet, no KO control). OR a strong indirect signal (e.g. surface biotinylation MS in 2+ cell lines).
• low — indirect-only (fractionation, glycoproteomics without surface specification) OR a single mention without methodology detail OR review-level assertion with no primary citation traced.

Why emit stances before the rationale? So the structured per-claim call commits FIRST and the rationale then summarizes the stances. Writing the rationale first and back-filling stances to match leads to prose-driven post-hoc rationalization; the stance map should drive the prose, not the other way around.

Worked example — state-conditional surface form with a non-surface baseline (the canonical 5b.8 case). Apply the structure to whichever biology your gene's ledger surfaces — outer-leaflet inversion under transformation, stress-induced release of an intracellular pool, activation-induced display from an organellar reserve, etc. Pull the SPECIFIC baseline-state biology and induced-state mechanism from the gene's actual evidence ledger; the shape below is a template, not content.

"claim_stances": [
  {"claim_id": "a1_evi_01", "stance": "supports_surface",    "weight": "high",
   "note": "induced-state surface form, primary evidence in vitro + in vivo"},
  {"claim_id": "a1_evi_02", "stance": "supports_surface",    "weight": "high",
   "note": "antibody-mediated killing in xenografts (induced state)"},
  {"claim_id": "a1_evi_05", "stance": "tangential",          "weight": "high",
   "note": "canonical baseline topology — describes the baseline state, NOT a contradiction (different state)"},
  {"claim_id": "a1_evi_12", "stance": "supports_surface",    "weight": "high",
   "note": "non-permeabilized surface biotinylation"},
  {"claim_id": "a1_evi_15", "stance": "tangential",          "weight": "high",
   "note": "canonical baseline localization, resting state (not a refutation of induced-state surface form)"}
],

Canonical-baseline-localization claims describe the target's BASELINE state. They DON'T contradict the induced surface form — the two coexist under the state-conditional mechanism. Marking them contradicts_surface forces the grade to conflicting, which is wrong. They're tangential to the surface call (they inform the baseline picture that state_dependence=high captures) and the grade lands on direct_single_method — anchored on the induced-state-surface papers' direct methodology, with state-conditionality flagged separately on confidence / state_dependence.

Only mark a baseline-localization claim as contradicts_surface when it's incompatible with the surface-positive evidence under EVERY plausible mechanism (e.g. a definitive negative-staining result under the same conditions as a positive-staining claim).

Prose explaining the grade — what methods you saw, what sources, what contradictions if any. Soft target ≤800 chars (overshoots are accepted with a warning, prefer concision). The rationale should summarize the stance map you just emitted — name the high-weight supports + contradicts in plain language, then state the grade.

Citation discipline — inline cites on every specific claim

Every numbered item, named experiment, mechanism, or method-specific assertion in grade_rationale REQUIRES an inline (aN_evi_NN) cite immediately after the claim, drawn from claim_stances. The reader must be able to click straight to the source for every substantive claim. Loose summarizing prose ("the surface evidence is moderate overall") doesn't need a per-sentence cite; specific claims do.

A specific claim is anything that:

• enumerates separate experiments ((1), (2), (3) lists)
• names an experimental method (live-cell flow, surface biotinylation, crosslinking, photoaffinity labeling, knockin / knockout, cryo-EM, proteinase-K protection, ChIP, etc.)
• names a mechanism, observation, or result (e.g. "basolateral PM enrichment in polarized epithelial monolayers", "competitive displacement of a labelled ligand", "cargo-adaptor-mediated PM delivery from an internal pool")
• names a cell line, tissue, species, or assay condition

The schema enforces this: when grade_rationale contains structured-claim markers AND claim_stances has ≥2 rows AND zero inline cites are present, validation FAILS. Inline a cite per substantive claim, drawn from claim_stances — never leave a numbered item or named experiment uncited.

Per-claim specificity

Each substantive claim in the rationale should also name (in addition to its inline cite):

• the assay readout (what was measured — surface staining, flow MFI, crosslink band, localization pattern, structural complex)
• the cell type / species (human primary monocytes, polarized epithelial monolayers, patient-derived organoids, mouse primary cells of a defined lineage, etc.)
• the permeabilization status when relevant (live-cell, nonperm, permeabilized — say permeabilization unspecified rather than glossing it over)

A specific, citable, method-anchored claim looks like:

"live-cell flow on primary human cells of the relevant lineage showed surface staining lost in CRISPR-KO controls (nonperm, single source) (a1_evi_07)"

A vague claim that fails the discipline looks like:

"cargo-adaptor recycling delivers the protein to the PM" — no method, no perm status, no cell type, no cite.

Worked example — citation-disciplined rationale

Shape only; the prose is for a hypothetical type II single-pass receptor expressed on differentiated myeloid cells. Apply the SAME structure to whatever gene you're grading.

"grade_rationale": "Three direct lines of evidence support surface
exposure: (1) live-cell flow on primary human differentiated myeloid
cells with CRISPR-KO loss-of-signal control, replicated across two
independent donors (a1_evi_07); (2) non-permeabilized surface
biotinylation MS on the same primary cell population, peptide IDs
recovered above isotype background in 3/3 biological replicates
(a1_evi_12); (3) a cryo-EM structure of the ectodomain in complex
with a soluble extracellular partner, resolving the membrane-distal
half of the receptor (a1_evi_11). All three are independent method
classes (live flow, biotinylation MS, structural complex) on
consistent cell context, with the KO control on (1) ruling out
antibody cross-reactivity. Graded direct_multi_method."

Every numbered item carries its (aN_evi_NN) chip. The reader can click each to verify.

Each row is RNA / IHC / bulk-protein expression observation that does NOT establish surface accessibility on its own — it qualifies or contextualizes the surface claim rather than directly evidencing it. Source claims: those with evidence_type in {rt_qpcr, rna_seq, single_cell_rna_seq, in_situ_hybridization, northern_blot, microarray} OR claims with claim_type=tissue_expression that read tissue / cell line presence without nonperm surface assay.

Fields:

• context — free text describing the cell / tissue / sample.
• sample_type — closed enum: primary_human_tissue, primary_human_cell, patient_sample, patient_derived_organoid, iPSC_derived, established_cell_line, xenograft, ex_vivo, unknown.
• measurement_type — closed enum: RNA, bulk_protein, IHC_protein, single_cell_RNA, unknown.
  • rt_qpcr, rna_seq, northern_blot, microarray → RNA.
  • single_cell_rna_seq, in_situ_hybridization → single_cell_RNA.
  • western_blot (whole-lysate) → bulk_protein.
  • IHC without nonperm spec → IHC_protein.
• level — closed enum: high, moderate, low, absent.
• cited_evidence_ids — every evidence_id whose claim contributed.

Group claims describing the same context into one row.

Audit trail of A1 ledger claims that describe the protein as a soluble ligand engaging a surface receptor on another cell, NOT the protein being on the outer face of the plasma membrane that expresses it. These claims are real biology — receptor pharmacology, DAMP signaling, partner binding — but they're not surface-accessibility evidence for this protein, and the methods builder's inclusion criterion rejects them. Log them here so the reader can see "we filtered N claims as ligand- engagement; here's why" rather than wondering whether the agent missed papers.

Each row:

{
  "evidence_id": "<one of the input ledger's evidence_ids>",
  "reason": "<short why, ≤240 chars, name the receptor the protein was binding>"
}

When a claim belongs here:

• The protein is studied as an extracellular factor / DAMP / cytokine / chemokine / alarmin engaging a named receptor on another cell.
• Crosslinking / FRET / co-IP captures the protein bound to a TM partner on the cell surface, where the TM partner IS the membrane component and this protein is the soluble ligand.
• Antibody-neutralization assays where the antibody sequesters the soluble form of the protein (NOT a surface-anchored form).
• ELISA / Western on cell-supernatant or extracellular fractions detecting the protein after release.

Each excluded claim should ALSO appear in claim_stances with stance=tangential and a note explaining the exclusion — the two rows agree (stance tags WHY the claim doesn't count toward the grade; excluded_as_ligand_engagement tags it for the audit trail).

If the ledger has no ligand-engagement claims, emit "excluded_as_ligand_engagement": [].

If the ledger has no qualifying non-surface expression claims, emit "non_surface_expression": [].

You have no tools. Cite-only over the ledger.

Claims with claim_type=tissue_expression are the primary input. Read each claim prose + quote and extract: tissue, cell type (when named), present level, disease context, and any cell states.

Also consider claim_type=surface_expression claims as candidate expression rows — this is a safety net. Many clips carry BOTH dimensions: surface engagement AND a named tissue / cell type / disease context. When PTS A2 tags a dual-dimension clip as surface_expression (e.g. "ligand engages protein X on tumor macrophages"; "marker of activated effector cells in the tumor microenvironment"), the tissue dimension still belongs in this block. For each surface_expression claim whose claim prose or assay_context (especially cell_context.disease_state and cell_type_or_line) names a tissue / cell type / disease context, emit a corresponding ExpressionRow with disease_context derived from that prose. Pure subcellular-localization claims that name only a compartment (apical vs basolateral, PM vs ER vs Golgi, ciliary, etc.) without a tissue/cell-type/disease context do NOT yield an expression row.

Don't emit redundant rows. If the same (tissue × cell_type × disease_context) tuple is already covered by a tissue_expression claim, prefer the tissue_expression source — merge the surface_expression claim's evidence_id into the existing row's cited_evidence_ids rather than emitting a duplicate row. Emit a fresh row only when the surface_expression claim names a tuple no tissue_expression claim covers.

A2's deterministic kickoff casts a deliberately wide net to thicken your evidence pool. The normal_tissue_expression standing axis covers both the six-organ tox panel (liver / lung / kidney / intestine / heart / brain) AND broader surface-anchored expression coverage — per-cell-type and lineage-restricted descriptors, primary-tissue and organoid surface readouts. The axis deliberately EXCLUDES RNA-only sources (scRNA-seq, snRNA-seq, spatial transcriptomics, microarray) — surface-expression evidence is judged on the measurement TYPE (IHC / flow / surface-MS / etc.) carried by the protein-method categories, not on the consortium or brand that published the dataset.

Trust the ledger you receive — the axis has already produced claims by the time this builder runs. Your job is to collapse them into unique (tissue × cell_type × disease_context) rows, not to filter on source. An IHC read of "X-positive lymphocytes in lung" and a flow-cytometry read of "X+ memory CD8 T cells in tumor" can both contribute to the same ExpressionRow.

ONE fenced ```json block containing a JSON ARRAY. Empty [] is fine.

• tissue — free text (e.g. cerebellum, kidney cortex). Required.
• cell_type — the specific cell type when the source names one (e.g. Purkinje neurons, alpha cells); null for a tissue-level observation with no resolved cell type.
• present — closed enum: high, moderate, low, absent, mixed, unknown. Use mixed when sources disagree on the level for the same (tissue × cell_type × disease) — don't pick a winner.
• disease_context — closed enum: normal, tumor, tumor_adjacent, other_disease, mixed, unknown.
• disease_label — free-text specific disease when known (e.g. clear-cell renal carcinoma); null otherwise.
• cell_states — free-text list of cell states (activated, stress-induced). May be empty.
• cited_evidence_ids — every evidence_id whose claim contributed.

One row per (tissue × cell_type × disease_context), NOT one row per claim. Multiple claims reporting the same (tissue × cell_type × disease) collapse into one row carrying every contributing cited_evidence_id.

When multiple claims report the same (tissue × cell_type × disease) tuple with different cell states, MERGE their cell states into that row's cell_states list — emit ONE row per unique tuple, carrying all contributing states and all contributing evidence IDs. Do not emit duplicate rows that differ only in cell_states.

Always keep the normal-tissue baseline next to the disease row, even when the present level matches. The off-tumor baseline is load-bearing for toxicity: "high in tumor AND high in normal kidney" reads very differently from "high in tumor, absent in normal kidney", and dropping the matching-level normal row erases that read. Emit both the normal and the disease row.

Cite-only over the ledger.

ONE fenced ```json block containing a JSON OBJECT (NOT an array).

Claims with claim_type=surface_expression that describe the protein's compartment / subdomain (plasma membrane, cilium, endosome, lipid raft, tight junction, lateral membrane, etc.). Atlas-style db_annotation claims listing subcellular locations are also primary input.

• primary_compartment — closed enum: plasma_membrane, endosome, lysosome, ER, Golgi, mitochondrion, nucleus, cytosol, secreted, other. Default to plasma_membrane for surfaceome candidates UNLESS the ledger strongly indicates the dominant pool is elsewhere.
• rationale — prose (soft target ≤400 chars) explaining WHY this primary compartment, citing the specific methods + cell types that pinned it. Inline (a2_evi_NN) cites required for every method named. See the "Rationale discipline" section below.
• dual_localization — JSON ARRAY of DualLocalization rows. Each row:
  • compartment — SHORT canonical organelle name (e.g. endosome, cilium, Golgi). Validator-enforced: no parentheticals, no conditional clauses ("upon X", "under Y"), ≤40 chars.
  • fraction_estimate — float between 0 and 1, OR null when no quantitative estimate exists.
  • condition — short trigger / context phrase (≤80 chars), e.g. under stress, in polarized cells. The detailed WHY (assay, cell type, perm status, source) goes in rationale, NOT here. Put every condition HERE, never in compartment.
  • rationale — prose (soft target ≤300 chars) explaining why this compartment is a non-primary pool: what assay observed it, cell type, perm status, with inline (a2_evi_NN) cites. See "Rationale discipline" below.
  • cited_evidence_ids — list. Use this for non-primary compartments the protein is reported in.
• membrane_subdomains — JSON ARRAY of MembraneSubdomain rows. Each row:
  • subdomain — SHORT canonical microdomain name of the OUTER-leaflet plasma membrane. It MUST be one of this exact closed set (verbatim, lowercase, snake_case), using other when none fit: lipid_raft, tight_junction, primary_cilium, apical_membrane, basolateral_membrane, immune_synapse, focal_adhesion, caveolae, other. Do not invent capitalization / singularization variants (e.g. not "cilia" / "Primary Cilium" — use primary_cilium). Same name discipline as compartment. An inner-leaflet / cytoplasmic-face anchor is NOT surface-accessible — route it to dual_localization, not here (a non-canonical value emitted here is coerced to other, so put it in the right field instead).
  • rationale — one short line (soft target ≤200 chars) naming the evidence that assigned this microdomain (raft purification, cilium IF, polarized-epithelium IHC, immune-synapse cluster), cell type, and perm status, with inline (a2_evi_NN) cites.
  • cited_evidence_ids — list. ONLY for outer-leaflet / surface microdomains. Do NOT put whole compartments (endosome, lysosome, Golgi) here — those are dual_localization. Do NOT put the inner leaflet / cytoplasmic face of the plasma membrane here (e.g. myristoylated/palmitoylated inner-leaflet kinases): that is NOT surface-accessible, so it belongs in dual_localization with a compartment like inner leaflet of plasma membrane instead — never as a surface subdomain.

Every block you emit carries a rationale field — the top-level SubcellularLocalization.rationale, every DualLocalization.rationale, and every MembraneSubdomain.rationale. Treat them the same way the evidence_grade block treats grade_rationale: name the assay readout, cell type, and (where relevant) the permeabilization status, and inline-cite the supporting (a2_evi_NN) id immediately after each specific claim.

• rationale (top-level) — one short paragraph (soft target ≤400 chars). State the dominant pool and the methods that pinned it (immunofluorescence, fractionation, IHC, Atlas annotation, non-permeabilized flow, etc.). Inline cite per claim. State the permeabilization status when it materially constrains what the assay can prove ("non-permeabilized IF" — surface; "permeabilized IF with PM-rim co-stain" — localization, not surface).
• dual_localization[*].rationale — one short paragraph per row (soft target ≤300 chars). Why is this compartment a non-primary pool? What assay observed it? In which cell type / state? When the pool is state-conditional, name the trigger AND the trigger-specific assay (e.g. "stress-induced surface exposure measured by non-permeabilized flow on the activated lineage").
• membrane_subdomains[*].rationale — one short line per row (soft target ≤200 chars). Which evidence assigned this microdomain — raft purification by detergent-resistant membrane fractionation, cilium IF on a ciliated cell line, polarized-epithelium IHC on a tissue section, immune-synapse co-cluster imaging — and in what cell type?

A specific claim is anything that names a method, mechanism, cell type, or condition. Loose framing ("predominantly intracellular") doesn't need a per-sentence cite; specific claims do.

A good rationale (placeholder gene "gene X"):

"Non-permeabilized IF on intact polarized epithelial monolayers shows apical-membrane staining co-clustered with a canonical apical marker (a2_evi_06); subcellular fractionation of the same cell line enriches the protein in the PM fraction (a2_evi_07)."

A vague rationale that fails the discipline:

"Predominantly plasma membrane by literature."

When the ledger genuinely has no relevant data for a row (e.g. the primary compartment was assigned by deterministic Atlas vote), write the rationale as a one-line factual statement of the Atlas / DB source and leave cited_evidence_ids to point at the Atlas claim. An honest "no relevant data in the ledger beyond the Atlas annotation (a2_evi_NN)" beats invented prose.

For backward-compat with records that pre-date these rationale fields, the schema accepts an empty string at load time (legacy D1 rows + on-disk snapshots still validate). New annotator runs MUST fill every rationale — leaving them empty is a regression the audit checks for.

membrane_subdomains records the CELL-INTRINSIC microdomain — which face of which membrane / which microdomain the protein occupies — as a localization FACT, independent of tissue context or how a binder is delivered. You do NOT assess whether a systemically delivered binder can REACH that surface in a given organ; that is the anatomical_accessibility builder's job. So "localizes to the apical membrane" (a bare subdomain fact) → here; "apical / luminal in intestinal epithelium, so shielded from the blood" (orientation + tissue + reachability consequence) → anatomical_accessibility. The apical / basolateral / ciliary labels appear in both blocks on purpose: here they mark WHERE on the cell; there they mark what that orientation MEANS for binder access in a named tissue. When a claim has both, the bare subdomain fact is yours and the tissue-level reachability is anatomical_accessibility's (same evidence_id, different substance).

If the ledger has no compartment claims at all, emit primary_compartment="other", empty dual_localization, empty membrane_subdomains.

You have no tools. Cite-only over the ledger. Every cited_evidence_ids value must appear in the input ledger as an evidence_id.

Emit rows when the ledger contains DIRECT evidence that the protein itself is at a polarized surface or membrane subdomain — and that position has a binder-reachability consequence. Examples of qualifying direct evidence:

• Polarized epithelium IHC / staining that resolves apical vs basolateral vs luminal location (not just "membranous staining" in a polarized cell — the assay has to actually pick a side).
• Subdomain localization observed directly — ciliary membrane, synaptic cleft, immunological synapse, brush border, microvilli, podocyte foot processes, intercalated discs, axon initial segment, presynaptic / postsynaptic, focal adhesions, adherens junctions, desmosomes, caveolae, lipid rafts, tight-junction restricted.
• Luminal-vs-abluminal endothelial labeling, side-specific biotinylation, intravascular tracer studies, etc. — assays that specifically distinguish blood-facing from tissue-facing surfaces of the same cell.
• Observed cell-layer restriction within a polarized tissue that pins the protein to the blood- or lumen-facing layer via a paired- layer comparison — e.g. "positive in ductal luminal cells, negative in the surrounding myoepithelial layer". A single-layer "expressed in cell type X" claim, without the negative-layer comparator, does not qualify on its own.

Do NOT manufacture a row by combining tissue_expression evidence with textbook anatomy. Pattern to reject: "Protein is expressed in cell type X. Cell type X is anatomically positioned at Y. Therefore accessibility = Z." That is the tissues / cell_types builder's job (where the protein is expressed), followed by the biological context grade builder's reasoning (what reachability that implies). Every emitted row here must rest on evidence that directly observed the protein at a polarized surface, not on evidence that observed it in a cell type whose anatomical position you happen to know.

Patterns that do NOT qualify on their own:

• "Gene X is expressed on endothelial cells" → expression only; endothelium-is-blood-facing is textbook anatomy, not a subdomain observation. Emit nothing for this evidence.
• "Gene X protein detected on the basal-cell layer of an epithelium" → cell-type expression; the basal layer's anatomical position alone is not enough without a paired-layer comparator.
• "Complete-membrane staining in a polarized epithelium" → not side- resolved; "complete-membrane" includes BOTH apical and basolateral, so it does not pick a side.

These same evidence rows belong in tissues / cell_types / expression; let those builders carry them. The downstream biological-context-grade builder then reasons across the full set — expression + true anatomical-accessibility rows + barriers — to give the reader the reachability picture. Your job is the narrow, high-confidence anatomical layer: protein-specific polarization or subdomain restriction.

Boundary — you answer a TISSUE-scale, binder-delivery question. Given the protein's directly-observed orientation in a NAMED tissue, can a SYSTEMICALLY DELIVERED binder reach that surface? (blood- / interstitium-facing = favorable; luminal / apical-only / behind a barrier like the BBB / junction-restricted = restricted.) Each row carries a tissue context + orientation + accessibility_implication. You do NOT own the cell-intrinsic compartment / microdomain assignment per se — "the protein sits in lipid rafts / on the apical membrane" as a bare localization fact is the subcellular_localization builder's membrane_subdomains. Emit here ONLY when the claim ties a DIRECTLY OBSERVED orientation to a tissue context with a reachability consequence; a pure subdomain-localization fact with no tissue / accessibility framing belongs to subcellular_localization.

Many genes have no such evidence — emitting [] is normal and correct. Empty output is strongly preferred to inference-padded rows.

ONE fenced ```json block containing a JSON ARRAY.

• context — free text (e.g. intestinal epithelium, kidney proximal tubule, airway epithelium).
• orientation — closed enum: blood_interstitial_facing, luminal_facing, apical, basolateral, lateral, junction_restricted, ciliary, synaptic, matrix_facing, unknown.
• accessibility_implication — closed enum: favorable, restricted, context_dependent, unclear. For systemically delivered binders: basolateral and blood_interstitial_facing → favorable; apical, luminal_facing, junction_restricted, ciliary → restricted (BBB / tight-junction barrier blocks systemic access); synaptic, matrix_facing → context_dependent.
• rationale — prose ≤300 chars explaining WHY this orientation affects accessibility for systemic delivery. Lead with the directly-observed localization (the specific assay/finding that pinned the protein to this surface), then the reachability consequence. Do NOT lead with textbook tissue anatomy or with "the protein is expressed in cell type X" — that pattern is the warning sign the row shouldn't exist.
• cited_evidence_ids — every evidence_id whose claim contributed.

You have no tools. Cite-only over the ledger.

ONE fenced ```json block containing a JSON ARRAY. Empty [] is fine for genes with no documented modulation.

Read claims describing CHANGES in surface presence (induction by stress, activation, disease state; lysosomal exocytosis; restriction to a lineage; dual-localization with intracellular pool; polarized cells; post-translational shifts; developmental gating).

A2's deterministic kickoff includes a dedicated cell_state_modulation standing axis that pulls papers describing the protein in activation / stress / disease-state / tumor- microenvironment / EMT / senescence / differentiation contexts — both the contrast-shape papers ("activated vs resting") and the single-context-shape papers ("in drug-tolerant persister cells the surfaceome is remodeled"). Trust the ledger you receive — the axis has already produced claims by the time this builder runs. Your job is to convert each state-relevant claim into the right ROW SHAPE (below) and the right category, not to filter on retrieval provenance.

This builder emits two shapes of row, distinguished by whether the paper drew an explicit before→after contrast.

Shape A — CONTRAST row (baseline_context AND modulating_state both set): the ledger documents a real CHANGE in how much of the protein is on the cell surface / reachable by an extracellular binder, between two named states. Qualifying shifts: surface expression up or down, a surface fraction appearing or disappearing, trafficking to or from the plasma membrane, an epitope becoming masked / unmasked, or a polarity / compartment shift that moves the protein on or off the reachable surface. The change field MUST state that surface-level shift, and baseline_context → modulating_state MUST be the two states the shift occurs between. Use direction (increases / decreases / bidirectional / no_change / unclear) to encode the direction.

Shape B — SINGLE-CONTEXT row (baseline_context AND modulating_state both NULL): the ledger describes the protein's surface behaviour in ONE state without a comparison condition. Schema 2.5.0 merged the former cell_states[] block into this builder; these rows replace it. Use shape B when the paper says something like "in drug-tolerant persister cells, the target's surfaceome is remodeled" or "in a virally-transformed cell type, the target shows aberrant surface signaling" — the state matters and the protein behaves a certain way there, but there is no clean A→B contrast you could put into baseline_context and modulating_state without inventing the comparator. The change field carries the prose describing that single-state behaviour; direction is typically unclear; category still applies (the state IS state-induced / stress-induced / disease-state-induced even without a contrast pair).

Pick the right shape per row: if the paper draws the contrast, emit shape A. If it only describes one state, emit shape B. Never emit a row with one of (baseline_context, modulating_state) set and the other null — the schema validator rejects that as an under-specified contrast. If you have a baseline in mind, name it explicitly; if you don't, leave both null.

Do NOT emit a row for:

• A change in total or intracellular abundance, mRNA, signaling activity, phosphorylation, or downstream pathway that is NOT tied to a change in the surface-accessible pool.
• A bare statement that a state or cell type exists, with no documented surface relevance at all — neither a contrast (shape A) nor a single-context surface observation (shape B).

If the ledger has expression / biology context but no surface-accessibility shift AND no single-state surface observation, emit empty []. An over-broad row that just restates "this cell type has gene X" is worse than no row.

When two ledger claims describe SAME-TISSUE expression observations of this protein under different disease_contexts — one normal / healthy / non-tumor baseline, one tumor / tumor-adjacent modulating state — AND the present level differs by ≥1 enum step (absent → any present; low/moderate → high; etc.), you MUST emit ONE contrast (shape A) row capturing the modulation:

• category = cell_state_induced
• cell_state_trigger = oncogenic_transformation
• direction = increases when the tumor read is the higher level; decreases when the normal read is the higher level; bidirectional if the ledger documents both directions across multiple cancer sites.
• baseline_context = the normal observation's tissue / cell context (e.g. "normal colonic epithelium").
• modulating_state = the tumor observation's tissue / cell context (e.g. "colorectal carcinoma").
• cited_evidence_ids = BOTH the normal and tumor evidence_ids.

This rule lifts evidence that the model previously skipped because the paper did not draw the contrast in a single sentence — the contrast is across two separate ledger entries. The synthesizer's surface_call_ reason = cell_state_induced ↔ amod recall-check depends on these rows landing; missing them produces a false confidence-reasoning miss.

Apply the rule independently per tissue pair. A protein normal-vs- tumor-induced across breast / colon / lung emits THREE separate rows, not one merged row. Combine into one row ONLY when the tissues are part of the same anatomical-site family (e.g. "non-small-cell lung carcinoma" + "small-cell lung carcinoma" → one modulating_state = "lung carcinoma" row).

If a ## Candidate modulation rows derived from expression-level deltas section appears in the user prompt, it lists the pairs the deterministic detector found. Apply the qualifying-shift gate (level delta + surface relevance) to each candidate before emitting. False- positive candidates in that list are expected — skip any that don't clear the gate.

• category — 12-value enum. PICK ONE:
  • cell_state_induced — generic cell-state shift.
  • tissue_restricted_surface — surface presence restricted to one lineage / tissue type.
  • lysosomal_exocytosis — surface presence depends on lysosomal fusion / exocytosis.
  • dual_localization — splits between PM and another compartment.
  • stable_surface_attachment — baseline stable; no major modulation (use sparingly).
  • activation_induced — induced by immune / receptor activation (TCR, BCR, cytokine).
  • stress_induced — induced by ER / oxidative / heat / DNA-damage stress.
  • disease_state_induced — induced specifically in disease (tumor, autoimmune, infection).
  • polarization_dependent — depends on apical-basolateral polarity. Emit polarization_dependent ONLY when the paper describes an actual SHIFT or state-gated change in polarity. If the paper merely notes a static orientation (e.g. 'apical'), do NOT emit here — that belongs to anatomical_accessibility. If the paper reports BOTH a static orientation AND a state-dependent shift (e.g. apical in healthy epithelium, basolateral in cancer), emit a polarization_dependent row here for the shift AND let anatomical_accessibility capture the static orientation, both citing the same evidence_id.
  • post_translational_dependent — depends on PTM (cleavage, phosphorylation, glycosylation).
  • developmental_stage — gated by developmental window.
  • none — no modulation documented.
  • other — anything else (requires category_other_label).
  • unknown.
• category_other_label — REQUIRED iff category="other"; otherwise MUST be null.
• cell_state_trigger — closed enum (ER_stress, heat_shock, oxidative_stress, DNA_damage_response, apoptosis, necroptosis, oncogenic_transformation, infection_viral, infection_bacterial, immune_activation, antigen_stimulation, cytokine_stimulation, hypoxia, nutrient_deprivation, hyperthermia, mechanical_stress, other, unknown). MAY be set ONLY when category is in {cell_state_induced, stress_induced, activation_induced, disease_state_induced, lysosomal_exocytosis}. Otherwise MUST be null. Set it ONLY when the inducing state genuinely matches one of the listed cell-state mechanisms. The enum covers cell-state stressors (stress / oncogenic / infection / immune / metabolic / mechanical) — it is NOT a disease vocabulary. For a disease_state_induced row whose disease is NOT one of those mechanisms — e.g. a genetic, developmental, or neurodegenerative disease such as Familial Dysautonomia — leave cell_state_trigger null; the disease itself belongs in baseline_context / modulating_state (e.g. baseline "healthy iPSC-derived neurons" → modulating "FD-patient iPSC-derived neurons"). NEVER use oncogenic_transformation unless the modulating state is an actual cancer / malignant transformation — it is not a catch-all for "some disease". When no listed mechanism fits, prefer null over a wrong-but-plausible pick (other/unknown are last resorts).
• restricted_lineage — closed enum (germline_reproductive, embryonic_developmental, hematopoietic, neural, epithelial, endothelial, muscle, endocrine, specialized_somatic_other, other, unknown). MAY be set ONLY when category="tissue_restricted_surface". Otherwise MUST be null.
• dual_loc_partner_compartment — closed enum (ER, Golgi, endosome, lysosome, mitochondrion, nucleus, cytosol, secretory_vesicle, other, unknown). MAY be set ONLY when category="dual_localization". Otherwise MUST be null.
• baseline_context — OPTIONAL free text describing the baseline state where the protein is (or isn't) on the surface (e.g. resting CD4 T cell, unstressed HeLa, normal kidney epithelium). Set only on shape-A (contrast) rows; leave null on shape-B (single-context) rows.
• modulating_state — OPTIONAL free text describing the alternate state (e.g. TCR-stimulated CD4 T cell, thapsigargin-treated HeLa, tumor kidney epithelium). When set, MUST differ from baseline_context. Both baseline_context and modulating_state must be set TOGETHER (contrast row) or both null (single-context row); mixed is rejected.
• direction — closed enum (increases, decreases, bidirectional, no_change, unclear): the up/down direction of the surface pool from baseline to modulating state — independent of whether that is favorable or restricting for a binder.
• change — prose ≤300 chars describing what actually shifts.
• accessibility_implication — prose ≤300 chars describing what the shift means for binder access.
• cited_evidence_ids — every evidence_id whose claim contributed.

These rules are validator-enforced; violations cause schema validation to fail. Re-read this list before emitting EACH row:

1. category="other" ⇒ category_other_label is a non-empty string. Any other category ⇒ category_other_label is null.
2. cell_state_trigger may be non-null ONLY when category ∈ {cell_state_induced, stress_induced, activation_induced, disease_state_induced, lysosomal_exocytosis}.
3. restricted_lineage may be non-null ONLY when category == "tissue_restricted_surface".
4. dual_loc_partner_compartment may be non-null ONLY when category == "dual_localization".
5. baseline_context and modulating_state must BOTH be set (shape A — contrast row, two DIFFERENT states) or BOTH be null (shape B — single-context row). Mixed is rejected as an under-specified contrast. On a shape-A row, the two endpoints must name two DIFFERENT states ("cells express X" is not a contrast).
6. cell_state_trigger="oncogenic_transformation" requires a cancer / tumor context in baseline_context or modulating_state — use a different trigger for a non-cancer disease state. This rule only applies to shape-A (contrast) rows; on shape-B rows the trigger still has to match the category but the cancer-vocab check is moot (there's no contrast text to scan).

When in doubt, set the optional sub-field to null rather than risk a mispairing — empty rows still validate; mispaired rows fail.

You have no tools. Cite-only over the ledger.

ONE fenced ```json block. Top-level OBJECT, not array. No prose around it.

Judge across the four A2 axes — expression, subcellular_localization, anatomical_accessibility, accessibility_modulation (which now also carries cell-state context, since the former cell_states builder was merged into it) — on two dimensions: COVERAGE (how many axes have real evidence) and CONSISTENCY (whether the axes agree).

• rich — multiple A2 axes are well-populated from independent sources and paint a coherent, consistent picture: expression is mapped across ≥2 tissues/cell types, subcellular localization is pinned, AND at least one of {anatomical_accessibility, accessibility_modulation} carries real evidence. No unexplained internal contradiction.
• moderate — a usable picture exists but is partial: one or two axes are well-evidenced (e.g. expression + localization) while others are thin or absent, OR the evidence is solid but from a single source / single context. Includes the common case of "expression mapped + PM localization confirmed, but no state / anatomical / modulation data".
• sparse — only fragmentary A2 evidence: a single tissue mention, an RNA-only read, a bare localization assertion, or scattered claims that don't cohere into a context picture. The reader can't situate the surface call in a biological context from what's here.
• absent — effectively no usable A2 biological context: no expression mapping, no localization evidence, only db-assertion-level or off-target claims. Use when the ledger is empty or contributes nothing to the context picture.

Coverage vs. consistency — how to combine them. Coverage sets the ceiling (you can't be rich with one axis); consistency can lower it (broad coverage that internally disagrees without a plausible reconciling mechanism caps at moderate). Note in grade_rationale when consistency — not coverage — is what held the grade down.

State / context variation is NOT inconsistency (same rule as A1's conflicting). A protein expressed in cerebellum but absent from liver, or surface-exposed only in activated T cells, is a coherent context — that's exactly what the A2 axes are meant to capture, and rich state-dependence is a sign of a WELL-characterized context, not a contradictory one. Only treat the picture as inconsistent when two A2 claims cannot both be true under any plausible biology (e.g. one source reports the protein is exclusively nuclear while another reports plasma- membrane localization in the same baseline cell state, with no dual- localization mechanism offered).

Prose explaining the grade — which A2 axes had evidence, how many sources, whether the picture cohered. Soft target ≤300 chars (overshoots are accepted with a warning; prefer concision). Name the axes that carried the grade in plain language, then state the grade. If consistency (not coverage) capped the grade, say so. Cite only the A2 ledger.

A1's methods builder REJECTS claims describing the protein as a soluble ligand engaging a surface receptor on another cell (DAMP signaling, cytokine binding, alarmin engagement, secreted-factor receptor pharmacology — these don't establish surface accessibility of this protein). When the A2 ledger carries those claims — receptor engagement, named binding partners, signaling pathway activation downstream of binding — pull them into the A2 axes that cover them: typically accessibility_modulation (state-conditional receptor engagement), cell_types (which cells express the cognate receptor), or expression/subcellular_localization (where the secreted form accumulates). The A1 ledger may flag a claim as excluded_as_ligand_engagement; A2 should still claim that biology if it's load-bearing for the picture. Cite the A2 ledger directly — don't reference A1's excluded_as_ligand_engagement rows by id (those identifiers live on the A1 side).

Every evidence_id from the input A2 ledger that materially informed your grade. Don't pad with tangential ids; don't invent ids that aren't in the ledger (any unresolved id is scrubbed downstream). An empty ledger → "cited_evidence_ids": [] and grade absent.

You have no tools. Cite-only over the A2 ledger.

ONE fenced ```json block. Top-level OBJECT (AccessibilityRisks), not array. Keys: co_receptor_requirements, shed_form, secreted_form, restricted_subdomain, ecd_size_assessment, epitope_masking. Emit ALL six — never omit a sub-block. (Do NOT emit homo_oligomerization_prediction; the orchestrator attaches that deterministically from Schweke 2024 after you finish.)

Each sub-block carries severity + evidence_strength so speculative-but-severe is distinguishable from real-but-mild. When the ledger shows nothing for a risk, set present=false with severity="low" (or "unknown" if genuinely ambiguous) and evidence_strength="weak" — never drop the sub-block.

Every chip's rationale must be self-supporting, and every cited_evidence_ids entry must resolve to a REAL id in the merged ledger I hand you (a1_evi_* / a2_evi_*). You have no tools — cite-only over the merged ledger. If you cannot quote it from a ledger entry, you cannot claim it.

• Carry an inline (a1_evi_NN) / (a2_evi_NN) cite next to the specific claim each rationale makes, AND list those ids in that block's cited_evidence_ids.
• Cite the evidence FOR THAT specific claim — partner-dependency evidence for co_receptor_requirements; subdomain / polarity evidence for restricted_subdomain; serum/plasma soluble-form evidence for secreted_form; structural / complex evidence for epitope_masking. cited_evidence_ids is NOT a "related reading" list for the gene; a generic surface-expression paper attached to a partner-dependency claim reads to the reader as a wrong citation. Drop a clip from a block's cites even when it's about the same gene if it doesn't specifically bear on THAT block's claim — an empty but correct cite list beats a padded one.
• If a call is present=false with genuinely no relevant evidence, write the rationale as "no relevant data in the ledger" — never leave a rationale that alludes to evidence ("studies show broad distribution") without the inline cite. Invented or paraphrased ids fail the run.

Does a partner have to be present for the TARGET to reach the surface? This is the surface-expression axis ONLY — function-side dependency (does the partner have to be present for signaling?) is out of scope.

Fields: surface_expression_dependency (closed enum), partners (list of symbols), evidence_basis (closed enum), rationale (soft target ≤400 chars), cited_evidence_ids.

surface_expression_dependency — closed 4-value enum:

• required — the target does not reach the surface without the named partner (obligate heterodimer, obligate escort/chaperone for ER exit, obligate trafficking partner).
• modulatory — a partner increases or stabilizes surface presence but the target can still reach the surface without it.
• none — no co-receptor is needed; the target traffics to and resides at the surface on its own.
• unknown — the ledger has no co-receptor information at all.

Default to none, not unknown, on NEGATIVE calls. The catalog filter on this field is load-bearing — readers querying for "monovalent-binder-compatible targets" want none; unknown makes those targets invisible to the filter. When your rationale explicitly states no co-receptor is needed (e.g. a single-pass receptor that traffics independently, or a lipid-anchored protein whose membrane association is entirely lipid-modification-driven with no obligate partner), set none. Reserve unknown for cases where the ledger has no co-receptor information — NOT for cases where it has negative information.

evidence_basis — closed enum: co_expression_only (partner and target co-expressed, no causal test), trafficking (chaperone / ER-exit / surface-trafficking studies), knockout (partner KO/KD abolishes target surface presence), mixed.

Cite evidence that bears on the partner/co-receptor-dependency call ITSELF (chaperone/trafficking studies, partner co-expression that gates surface presence, or the explicit statement that membrane association is partner-independent), NOT generic surface-expression / signaling / disease papers that merely feature the protein.

A proteolytically shed soluble form OF THE TARGET. Fields: present, severity, evidence_strength, mechanism (free text or null), sheddase_if_known (free text or null), rationale (soft target ≤300 chars), cited_evidence_ids.

Set present=true only when the ledger documents the TARGET's own ectodomain being proteolytically released (e.g. a sheddase / matrix metalloproteinase cleaving the target to produce a soluble target ectodomain). A sheddase cleaving the target's ligand is IRRELEVANT here. Name the sheddase_if_known and mechanism when the ledger gives them. Grade severity by how much the shed form depletes surface target / acts as a decoy, not by the mere existence of a cleavage site.

Rationale discipline (mirrors the other risk chips). Use the same inline-cite shape as co_receptor_requirements.rationale and restricted_subdomain.rationale: name the assay (sheddase cleavage mapping, serum/supernatant ectodomain quantification, surface-density depletion after sheddase induction), cell type / sample, and any documented decoy behavior, with an inline (a1_evi_NN) / (a2_evi_NN) cite per claim. On a present=false call with no relevant ledger evidence, write "no relevant data in the ledger" — never leave a rationale that alludes to evidence without an inline cite.

This sub-block is for SOLUBLE protein free in supernatant / serum / plasma that can compete with the surface protein for circulating antibody — the antibody-decoy concern. Fields: present, severity, evidence_strength, ratio_to_membrane (float or null), source (closed enum or null), rationale (soft target ≤300 chars), cited_evidence_ids.

It is about the TARGET PROTEIN being soluble — NOT its ligand being shed. A sheddase releasing the target's cognate growth-factor ligands is the LIGAND becoming soluble, not the target. Only count evidence that THIS protein exists as a free soluble species: a proteolytically shed ectodomain OF THE TARGET, or a soluble / TM-less splice isoform OF THE TARGET. If the only "shedding" evidence is about the protein's ligand/agonist, do NOT cite it and do NOT set source="proteolytic".

Do NOT include protein inside extracellular vesicles, exosomes, microvesicles, or apoptotic bodies — those proteins are shielded inside a lipid bilayer, aren't accessible to circulating antibody, and are NOT a decoy. EV cargo is biology-context information (it belongs in the biological-context blocks, not here). If the ledger has ONLY EV-association evidence and no free-soluble evidence, set present=false.

source — closed enum: alternative_splicing, proteolytic, both, unknown.

Grade severity by DOCUMENTED decoy behavior, not by the mere existence of a soluble form:

• An annotated / predicted soluble splice isoform exists (e.g. a receptor's TM-less alternative isoform) but the ledger shows no evidence it actually circulates or competes for binder — this is the WEAK case: severity="low", evidence_strength="weak", source="alternative_splicing". (The orchestrator separately sets this floor deterministically from isoform topology; don't contradict it, but don't inflate it either.)
• The soluble form is DOCUMENTED to circulate — measured in serum / plasma, reported as a shed/soluble ectodomain at physiological levels, OR shown to bind / compete with a therapeutic antibody or ligand (a true decoy) — raise to severity="moderate" (or "high" when a paper explicitly ties it to reduced antibody efficacy / a clinical decoy effect), set evidence_strength to match the citation quality, and CITE the serum-level / competition papers. A receptor's serum-soluble ectodomain that competes with a therapeutic monoclonal is the canonical strong case.

So: name the soluble form, and if the ledger documents it circulating or out-competing a binder, the call is a real decoy risk — not a weak topology footnote. Cite that evidence. NOTE: when you set present=true, the record validator requires cited_evidence_ids to back it — a present=true with empty cites fails downstream.

Rationale discipline (mirrors the other risk chips). Use the same inline-cite shape as co_receptor_requirements.rationale and restricted_subdomain.rationale: name the assay (serum / plasma quantification, alternative-splicing isoform call, antibody-decoy competition assay), the sample / cell type, and the basis for the severity call, with inline (a1_evi_NN) / (a2_evi_NN) cites per claim. On a present=false call with no relevant ledger evidence, write "no relevant data in the ledger" — never leave a rationale that alludes to evidence without an inline cite. When the call rests entirely on the deterministic alternative-splicing floor (no ledger evidence of circulation / competition), state that explicitly so the reader can tell the floor apart from a literature-grounded decoy call.

Surface presence restricted to a membrane subdomain (apical, junctional, ciliary, etc.) that a systemic binder may not reach. Fields: present, domain (closed enum), severity, evidence_strength, rationale (soft target ≤300 chars), cited_evidence_ids.

domain — closed enum: apical, junctional, ciliary, synaptic, raft, basolateral, other, unknown.

Cite evidence about the actual spatial DISTRIBUTION (apical / basolateral / junctional / ciliary IF, polarized-epithelium staining), NOT papers that only establish the protein is surface-resident.

Cite evidence even on negative observations. When present=false (no restriction observed), cited_evidence_ids should still reference the evidence that DEMONSTRATES the broad distribution that rules out restriction — membrane-wide surface staining, distributed antibody-killing reactivity, surface biotinylation MS without subcellular fractionation specificity. Empty cited_evidence_ids is only appropriate when the ledger truly contains no relevant data — explicitly say "no relevant data in the ledger" in the rationale, and set domain="unknown".

ecd_accessibility_class is a closed enum (large / moderate / small / minimal / none) that the orchestrator computes deterministically AFTER you finish and OVERWRITES whatever you emit. It is the single source of truth, derived in code from deterministic_features.canonical_topology.ecd_length_residues. You have no judgment and no literature override on this field.

So do NOT make a real ECD-size judgment. Read ecd_length_residues from the deterministic-features summary in my message and apply these bands VERBATIM (no interpolation, no override). If you doubt the topology, do nothing about it here — your value is normalized away regardless.

• large — ecd_length_residues >= 200.
• moderate — 60 <= ecd_length_residues < 200.
• small — 30 <= ecd_length_residues < 60.
• minimal — ecd_length_residues < 30.
• none — ecd_length_residues == 0 (or the field is absent/null).

Fields: ecd_accessibility_class, rationale (soft target ≤300 chars), cited_evidence_ids. Write a short factual rationale stating the residue count and the band (e.g. "ECD ~310 residues -> large"); the orchestrator rewrites both class and rationale deterministically, so keep it minimal. Leave cited_evidence_ids empty — this is a deterministic topology call, not a literature-anchored one.

Whether the extracellular epitope is masked from binders. Fields: mechanism (LIST of closed-enum values), severity (closed enum), evidence_strength, rationale (soft target ≤400 chars), cited_evidence_ids.

mechanism is a list so multi-mechanism cases don't collapse to a single value. It spans three physically distinct masking axes, and the catalog filters on them, so pick the value(s) that name the actual cause:

• HOMO → oligomerization. The target's OWN homodimer / homo-oligomer interface buries epitope surface — tetraspanin and claudin cis-clustering, GPCR or receptor homodimers, self-associating ECDs. The protein masks itself; no second protein is involved. Use oligomerization, NOT conformational (a monomer closed/open state is a different mechanism) and NOT partner (a different protein).

  Treat homo-oligomerization as CORROBORATION, not a primary literature call. The orchestrator attaches the deterministic Schweke 2024 AF2 homo-oligomer prediction as its OWN structured chip (homo_oligomerization_prediction) after you finish — that is the authoritative AF2 signal, and you do not emit it. The deterministic-features summary in my message carries homo_oligomerization.is_homo_oligomer + stoichiometry for your awareness. Use them only to CORROBORATE an oligomerization mechanism the literature in the merged ledger already supports:

  • When the LEDGER documents a homodimer / homo-oligomer that buries extracellular epitope surface, emit oligomerization and cite the ledger evidence; the deterministic is_homo_oligomer=true is supporting context you can mention in the rationale (but the cite must be a real ledger id, not the deterministic block).
  • When the ledger has NO homo-oligomerization evidence, do NOT add oligomerization to the mechanism list on the strength of the deterministic prior alone — that signal already ships as its own chip. Adding it here would double-count and leave you with no ledger id to cite. stoichiometry = N is a useful severity scale when you DO have ledger support: a 2-mer buries less surface than a high-order complex.

• HETERO → partner. A different protein in a hetero-complex covers the epitope — e.g. a co-receptor sitting over the target's large extracellular loop in a constitutive complex. When you set partner, the masking protein is almost always one already named in co_receptor_requirements.partners; keep the two blocks consistent. Cite the structural / complex evidence (cryo-EM, co-crystal, pulldown), never a generic surface-expression paper.

• OTHER → glycan / conformational / cleaved. Glycocalyx or N-/O-glycan shielding of the epitope (glycan); intrinsic monomer closed/open occlusion (conformational); proteolytic removal of the epitope (cleaved).

• none — no masking documented; use it as the sole list entry when the epitope is unobstructed.

Multi-mechanism is allowed and common: a multi-pass tetraspanin can carry ["partner", "oligomerization"] when the ledger documents BOTH a co-receptor covering the large extracellular loop AND the target's own tetraspanin-microdomain clustering.

severity — closed enum: high, moderate, low, none. Grade by how constitutively the masking holds in the targetable state — a complex constitutive on the relevant cell is more consequential than an occasional or inducible one — and set evidence_strength to match the cited structural/complex evidence.

1. All six sub-blocks present.
2. Every cited_evidence_ids entry is a real a1_evi_* / a2_evi_* id from the merged ledger I gave you.
3. Each present=true (or non-none/unknown) call carries a cite that SPECIFICALLY backs it.
4. ecd_size_assessment.ecd_accessibility_class matches the deterministic residue band (it gets overwritten anyway, but emit it right).
5. oligomerization appears in epitope_masking.mechanism ONLY when the merged ledger documents it — not on the deterministic prior alone.
6. No homo_oligomerization_prediction key (orchestrator-only).

A single fenced JSON block: a SynthesizerDraft. The exact JSON schema is in your task message; follow it. You GENERATE three blocks (executive_summary, filters_llm, confidence + confidence_reasoning) and COPY THROUGH one frozen block (accessibility_risks):

• executive_summary — the 4-beat narrative one_paragraph (verdict → evidence → state-dependence → risk; see the dedicated section below for the full spec), accessibility_context_summary (ONE sentence naming WHEN and WHERE the protein is surface-accessible — the load-bearing §03 headline, not a generic blurb; e.g. "Surface-accessible on activated T cells and in tumor tissue, but intracellular in resting cells"), the closed-enum verdicts (surface_accessibility, evidence_grade_summary, confidence, state_dependence, subcategory, surface_call_reason), ≤3 headline_risks, and cited_evidence_ids from the merged ledger.
• accessibility_risks — PROVIDED, frozen. Copy it through verbatim (see "Accessibility risks — provided, not generated" below for the contract and the consumption rules).
• filters_llm — three rollups only: expression_level, expression_breadth, surface_specificity, each paired with a one-line *_rationale (expression_level_rationale, etc.) carrying inline (a1_evi_NN) / (a2_evi_NN) cites so each chip is self-auditable. The other 14 filter fields are orchestrator-derived; do not emit them here.
• confidence + confidence_reasoning (≤600 char; required non-empty when confidence ∈ {moderate, low}). Write this for the catalog reader (target-discovery analyst, biologist, BD reader), not for the pipeline. See the "confidence_reasoning — writing for the reader" section below for the prohibited-language list and worked example.

The risks builder runs in parallel over the merged A1+A2 ledger and emits all six risk sub-blocks — co_receptor_requirements, shed_form, secreted_form, restricted_subdomain, ecd_size_assessment, epitope_masking — each with severity, evidence_strength, rationale, and per-risk cited_evidence_ids. (The orchestrator overwrites ecd_size_assessment deterministically from the ECD residue count and attaches a deterministic homo_oligomerization_prediction chip.) The finished block arrives in your task message under "Accessibility risks (PROVIDED — frozen)". Treat it exactly like deterministic_features: a read-only input you reference and pass through, NOT something you author.

Your two jobs with it:

1. Copy it through verbatim into your output's accessibility_risks field. Do not regenerate, re-grade, re-cite, reword a rationale, flip a present flag, or change a severity. The orchestrator detects drift.
2. Consume it to drive the two top-line fields that depend on risks:
   • headline_risks — select from the FROZEN risk sub-blocks. A secreted_form with present=true + severity=high belongs; a low-severity restricted_subdomain doesn't. Read the provided block's flags — don't re-derive the risk picture from the raw ledger. See "Headline-risks selection discipline" below.
   • confidence — weigh the consequential provided risks alongside the evidence grade and state-dependence (a documented high-severity decoy or obligate co-receptor dependency is a confidence drag).

When confidence ∈ {moderate, low}, the validator requires non-empty confidence_reasoning prose. This prose is user-facing — it renders on the catalog gene page below the confidence chip, and gets read by target-discovery analysts, biologists, and BD readers who have never opened this prompt or the codebase. Write for them, not for the pipeline.

Target audience: a target-discovery analyst evaluating whether this protein is worth pursuing. They want to know:

• Why isn't the confidence higher? (what's the catch?)
• What would change the call? (what additional evidence would lift it?)
• What's the practical implication for the next experiment / decision?

Hard ban — treat as a syntactic filter

The following patterns MUST NOT appear in any freeform prose field you emit — confidence_reasoning, executive_summary.one_paragraph, executive_summary.accessibility_context_summary, surface_evidence.grade_rationale, biological_context.grade_rationale, and every per-block rationale or mechanism string. The viewer renders these fields directly to a target-discovery analyst; pipeline-namespace words leak the agent's internal structure and read as "weird" prose. (Historically the ban applied only to confidence_reasoning; in production records we caught leaks like "The entire A1 ledger consistently places BAX at intracellular compartments…" in executive_summary.one_paragraph and the per-block rationales, so the scope is now every prose field.)

Self-check before emitting every prose string: scan it for any of A1, A2, a1_evi_, a2_evi_, the merged ledger, the A1+A2 evidence, verdict=', accessibility=', state_dependence=', evidence_grade=', deep-dive, or triage called. If any appear, rewrite that sentence.

Required content: 2-3 sentences naming (a) why confidence is moderate or low — the specific weakness in the evidence, (b) what would lift it — the kind of follow-up that would make this a confident-high call.

Worked example — reader-facing shape (state-conditional cancer target):

"Confidence is moderate because the cancer-cell extracellular form comes from a single recent research cluster (two papers from the same group, PMID:XXXX and PMID:YYYY). The canonical baseline topology is well-established across decades of independent work. Lifting confidence would need a third independent group to confirm the cancer-state outer-leaflet exposure, ideally with a different methodology than the one in the existing reports."

The exemplar uses PMIDs (citable), names what alternative would corroborate, and says what would change the call. It does NOT mention A1, evidence_grade, or aN_evi_NN tokens.

You have no tools. Cite-only over the merged A1 + A2 evidence ledger in your task message. If you cannot quote it from the ledger, you cannot claim it. Every cited_evidence_ids value must resolve to an entry in the merged ledger (prefixes a1_evi_* and a2_evi_*). The orchestrator validates this at parse time — invented or paraphrased ids fail the run.

• surface_accessibility — pick the best-case state for a target-discovery reader, not the worst case. If accessibility is high in some state (cancer cells, activated immune cells, stressed cells, polarized epithelia), set surface_accessibility=high and use state_dependence={moderate, high} to flag the conditionality. Don't pick low just because normal-cell accessibility is low — that erases the targetable state. The reader filters on surface_accessibility to find candidates, and on state_dependence to understand whether targeting is constitutive or state-gated. Canonical case for an inner-leaflet kinase: a cancer-state outer-leaflet inverted form is high in cancer cells, so surface_accessibility=high + state_dependence=high (not surface_accessibility=low because normal-cell access is low). Reserve surface_accessibility=no for proteins where the deep-dive evidence does not surface a targetable state anywhere.

  high requires direct evidence in hand (topology prior alone is not enough — catalog readers treat high as "confidently surface-accessible"). THE BRACKET below is the single source of truth — other sections cross-reference it, never restate it:

  grade + confidence	max surface_accessibility
  direct_multi_method	high
  direct_single_method + confidence=high	high
  direct_single_method + confidence ∈ {moderate, low}	moderate
  supportive_but_indirect	moderate
  weak / conflicting	low

  state_dependence still flags conditionality when the bracket caps high → moderate, so canonical-receptor signal isn't lost.

• state_dependence — captures how much the targetable surface fraction VARIES by state (cell type, activation, cancer induction, stress, etc.). low means the surface form is essentially the same across the contexts the evidence covers; moderate / high mean the targetable state is state-conditional and the catalog reader should know that before scoping a campaign.

  state_dependence='low' is forbidden when the A2 biology shows state-conditional upregulation of the surface form. Specifically, pick at least moderate when ANY of:

  • filters_llm.induction_trigger != "none" (you set this — your own derivation of the dominant trigger bucket)
  • A2's accessibility_modulation contains ≥3 rows with direction="increases" (independent observations of the surface form going up under some state)
  • A2's accessibility_modulation contains any cell_state_trigger="oncogenic_transformation" row with direction="increases" AND surface_accessibility != "no" (the textbook tumor-induced overexpression case)

  Picking low in these situations erases the targetable state the catalog reader is filtering for. A protein with broad expression in normal tissue PLUS cancer-induced surface upregulation is exactly the case where the targetable signal is the delta — set state_dependence ∈ {moderate, high} so the reader sees it.

  The orchestrator validates this rule post-hoc; a violation raises rather than silently shipping a miscalibrated record.

• surface_call_reason — emit your own reason for the call, using the same closed 19-value enum as triage_record.reason. Re-derive it from the A1+A2 evidence ledger — do not blindly copy the triage's reason. The deep-dive's reason is more trustworthy than the triage's first-pass call.

  The enum partitions by which surface_accessibility verdict each reason supports — keep your reason consistent with your headline call:

  YES-bucket (use when surface_accessibility ∈ {high, moderate}): classical_surface_receptor, gpi_anchored, multipass_with_exposed_loops, extracellular_face_protein, stable_complex_partner, other.

  CONTEXTUAL-bucket (use when surface_accessibility ∈ {high, moderate} AND state_dependence ∈ {moderate, high} — the targetable state is conditional): cell_state_induced, tissue_restricted_surface, lysosomal_exocytosis, dual_localization, stable_surface_attachment, other. The first five mirror accessibility_modulation.category verbatim.

  Soft recall-check on scr ↔ amod.category coupling. If scr ∈ {lysosomal_exocytosis, dual_localization, tissue_restricted_surface}, the amod block should contain at least one row whose category is in the same finer-grained family — these three reasons name a specific accessibility mechanism, so the absence of a matching amod row is a recall miss worth flagging in confidence_reasoning. cell_state_induced is the umbrella category — finer-grained reasons like activation_induced, lysosomal_exocytosis, stress_induced, or disease_state_induced on an amod row are all valid backing for scr = cell_state_induced, and any non-empty amod block satisfies the check; only an empty amod block under scr = cell_state_induced is a soft signal to note. If no matching family row exists for any of these, note the recall miss in confidence_reasoning — the synth call still stands, but the catalog reader should know the amod backing is thin.

  NO-bucket (use when surface_accessibility ∈ {low, no}): cytoplasmic, nuclear, mitochondrial_internal, endomembrane_resident, nuclear_envelope, inner_leaflet_anchored, secreted_only, pmhc_only_intracellular, other. These are the deep-dive's best-guess for why the protein isn't surface- accessible — useful catalog signal even on negative calls.

  endomembrane_resident vs dual_localization — load-bearing distinction. Both apply to proteins whose canonical home is the endomembrane system (TGN / ER / endosomes / lysosomes / etc.). The test is whether the deep-dive evidence documents PM trafficking:

  • Literature shows transport carriers labelled with the protein arriving at the PM, baseline PM-rim staining under normal activity, or a measurable steady-state PM pool → pick dual_localization (CONTEXTUAL bucket) + surface_accessibility=low + state_dependence ∈ {moderate, high}. The brief PM dwell is enough for an antibody to engage; the targetable state IS the trafficking visit.
  • Literature treats the protein as never reaching the PM (no trafficking-to-PM evidence in the ledger, only intracellular compartment residence) → pick endomembrane_resident (NO bucket) + surface_accessibility ∈ {no, low}.

  Default to dual_localization whenever any trafficking-to-PM observation exists in A1's ledger, even when most of the evidence reads as intracellular — under-flagging the targetable PM-cycling state is the worse failure mode for the catalog reader.

  Often you'll confirm the triage's reason (canonical surface receptors stay classical_surface_receptor); sometimes you'll override (an inner-leaflet kinase's triage inner_leaflet_anchored becomes lysosomal_exocytosis when the deep-dive finds outer-leaflet inversion evidence, because the targetable state is the cancer-state-induced surface form). Confirm or override is the choice; just don't pass through without re-derivation.

• one_paragraph — ≤600 char (aim 500–580), the consultant-facing headline that opens every record. Write as a continuous narrative paragraph, NOT a bulleted or labeled list. The 4 beats below are structural — they fix the order and content — but the prefixes ("Risks:" / "State-dependence moderate:") never appear in the prose. The reader should read a paragraph, not a labeled list-in-prose-form.

  Beat order (structural; flow as one paragraph).

  1. Verdict beat (~100 char). Open with the surface_accessibility call + the architectural anchor + the gating qualifier baked into the sentence. Lead with the call adjective so the verdict is in the first 80 chars — never bury it in a subordinate clause. Shape examples (replace the parenthetical with THIS gene's actual architecture and gating):

     • "GENE X is constitutively surface-accessible as a (pan-tissue multi-pass receptor)."
     • "GENE Y is state-dependently surface-accessible in (cancer cells only) — a (non-surface-baseline protein with a state-conditional surface form)."

     Verdict-beat tone — match the evidence, not the topology prior. The opener tracks THE BRACKET row (above) — same row, fixed opener phrasing:

     bracket row	opener
     direct_* + confidence=high	"is constitutively / state-dependently surface-accessible"
     direct_single_method + confidence ∈ {moderate, low}	"is likely surface-accessible"
     supportive_but_indirect	"has supportive but indirect surface evidence"
     weak	"has weak surface evidence; topology suggests but direct readouts are missing"

     The opener is the FIRST thing the reader sees — overclaiming misleads a campaign-scoping reader.

  2. Evidence beat (~150 char). Flow into the evidence: use evidence_grade vocabulary for compression ("direct multi-method support", "supportive but indirect evidence") rather than reciting four method names. Name the 2–3 strongest evidence classes with inline cites. Example: "Direct multi-method support: live-cell flow with CRISPR-KO controls (a1_evi_04), surface biotinylation–MS (a1_evi_08), and LEL-blocking functional assays (a1_evi_13)."

     Methods-citing discipline. Read surface_evidence.methods[]. A cite can be labeled "Direct surface evidence" / "Direct multi/single-method support" ONLY if its accessibility_relevance = direct_surface_accessibility. Anything else (supports_membrane_association, supports_surface_localization, expression_only, weak_or_ambiguous, excluded_as_ligand_engagement) is supportive / indirect — never "direct".

     Cap the Evidence beat at the direct rows. When evidence_grade=direct_single_method, name the ONE direct method

     • its anchor cite and stop. When direct_multi_method, name the 2-3 direct methods. Do NOT list 3-6 indirect cites after the direct line — that contradicts the "single/multi" framing and the reader gets the indirect rows from the methods table anyway. Hard cap: ≤2 cites total in the Evidence beat when evidence_grade ≤ direct_single_method. A single indirect cite is allowed ONLY when it changes the picture (e.g. the ONLY in-vivo readout); even then, cap at one.

     Shapes:

     • direct_single → "Direct surface evidence is single-method: live-cell flow on [cell line] ([cite])."
     • direct_multi → "Direct multi-method support: live-cell flow with KO controls ([cite]) and surface biotinylation ([cite])."
     • supportive_but_indirect / weak → lead with the grade vocabulary, name the strongest 1-2 lines of support: "Supportive but indirect: perm IF with PM-rim colocalization ([cite])."

  3. State-dependence beat (~150 char). Continue the paragraph with state context. Embed the state_dependence value in flowing prose, never as a labeled prefix:

     • low → "Surface presence is constitutive across pan-tissue baseline with no significant state-modulation."
     • moderate → "Surface levels are moderately state-modulated, upregulated in X (a2_evi_NN) and downregulated in Y (a2_evi_NN)."
     • high → "Surface presence is strictly state-gated, requiring X and absent in Y (a2_evi_NN)."

  4. Risk beat (~150 char). Close with a single narrative sentence that:

     • Names the principal binder-engineering caveat when one fires — with the severity adjective and the structural locus, using the headline_risks enum content implicitly: epitope_masked → "epitope masking at the LEL interface"; secreted_form → "a soluble decoy pool"; restricted_subdomain → "restricted distribution at the ciliary membrane"; AND
     • Frames any meaningful positive nulls as risk-RULE-OUTS, not risk items. Use "rule out / are absent / are not documented" framing so a clean negative cannot be misread as a flagged concern. Examples:
       • "Moderate epitope masking at the LEL homodimer interface (a1_evi_21) is the principal binder-engineering caveat; the absence of a shed or secreted form and ≤28% paralog identity rule out decoy and cross-reactivity concerns."
       • "A dominant free-soluble pool is the principal antibody- decoy risk (a1_evi_18)."
       • When no risk fires at all: "No binder-engineering caveats emerged — no shed or secreted form, no co-receptor requirement, no restricted subdomain, and low paralog cross-reactivity."

  Citation syntax (load-bearing for the viewer). Inline cites are bare tokens in parentheses: write (a1_evi_04), never (`a1_evi_04`). The viewer's linkifier matches the bare aN_evi_NN token; backticks around it leak into the rendered chip as stray characters. Same rule applies to expression_level_rationale, expression_breadth_rationale, surface_specificity_rationale, has_known_ligand_rationale, accessibility_context_summary, and every other rationale field. The token shape is (a1_evi_NN) or (a1_evi_NN, a1_evi_MM) for lists — no backticks anywhere.

  Authoritativeness rules.

  • Lead with the call adjective (high / moderate / low / no) — never "appears to be" / "seems to" / "likely is". You have evidence; speak from it.
  • Use evidence_grade vocabulary for compression: "direct multi-method" encodes more than listing four method names.
  • Embed the state_dependence enum value in flowing prose so the rendered chip and the paragraph use the same vocabulary.
  • Frame absences as rule-outs, not as items in a risk list. "The absence of a shed form rules out a decoy concern" is correct; "Risks: ... no shed form ..." misreads as a flagged risk.
  • Cite ≥3 evidence_ids inline (≥1 per evidence-bearing beat) so each load-bearing claim has a trust anchor. Cluster cites at clause ends.
  • No marketing prose: drop "compelling target", "billion-dollar market", "promising candidate". State the biology; let the reader judge.
  • Do NOT exceed 600 chars — authoritative writing earns shorter, not longer. If you're at 620 chars, the right move is to compress the evidence beat (use evidence_grade language) rather than truncate the state-dependence or risk beat.

  Worked exemplars — two archetypes covering the rule-out vs risk-firing risk-beat patterns. Anchor on the SHAPE not the gene-specific content.

  Constitutive accessibility, state-modulated, risks ruled out (~580 char):

  "Gene X is constitutively surface-accessible as a pan-tissue multi-pass tetraspanin. Direct multi-method support: live-cell flow with CRISPR-KO controls (a1_evi_04), surface biotinylation–MS (a1_evi_08), and ECD-blocking functional assays (a1_evi_13). Surface levels are moderately state-modulated, upregulated in selected hematologic malignancies and viral-transformed B-cell contexts (a2_evi_15) and downregulated on activated lymphocytes and a hepatotropic-virus-replicating hepatocyte state (a2_evi_20). Moderate epitope masking at the homodimer interface (a1_evi_21) is the principal binder-engineering caveat; the absence of a shed or secreted form and low paralog identity rule out decoy and cross-reactivity concerns."

  State-gated accessibility, decoy risk firing (~545 char):

  "Gene Z is state-dependently surface-accessible despite a non-surface baseline localization. Multiple methods (a1_evi_04, a1_evi_07, a1_evi_11) document an extracellular pool that engages receptors on responder cells. Surface presence is strictly state-gated, requiring a stress / damage / activation trigger, with the implicated cell states driving the accessible pool (a2_evi_06, a2_evi_12). A dominant free-soluble pool — the protein released as a soluble factor under the same trigger — is the principal antibody-decoy risk (a1_evi_18)."

  Authoritativeness note on these exemplars. Each archetype is a SHAPE template, not a content template. The abstract mechanism phrases are intentional — do NOT paste them into every record. Pull the SPECIFIC mechanism, cell state, and trigger from THIS gene's evidence ledger; the exemplar only fixes the narrative arc (verdict → evidence → state → risk) and the character budgets.

• accessibility_context_summary — ONE sentence (≤240 chars) stating when and where the protein is surface-accessible, synthesized over the A2 biological_context block (accessibility_modulation + subcellular_localization + anatomical_accessibility). It is the headline behind the §03 "Localization & accessibility context" summary and the §01 signal panel, so keep it to the accessibility condition — the gating state / lineage / tissue and what becomes reachable — and do NOT restate one_paragraph. Examples (shape only — pull the SPECIFIC gating state, tissue, and mechanism from THIS gene's A2 ledger): "Surface-accessible only on cancer cells, where the state- conditional anchoring mechanism brings the protein to the outer membrane."; for a canonical receptor: "Constitutively surface- accessible across normal and tumor tissue; not state-gated." Leave it null only when A2 produced no localization / modulation context at all.

• evidence_grade_summary rolls up A1's evidence_grade — it should track it unless a major A2 contradiction (e.g. dominant secreted form) drags the integrated verdict down. State the rollup logic in confidence_reasoning only when you depart from A1's grade. Weight A1's methods[].validation_strength explicitly when reasoning about the rollup: methods with validation_strength="strong" (paper-level genetic_KO, CRISPR_KO, isoform_specific_KO validation) carry the surface call and earn the direct_* grades; methods with validation_strength="weak" (vendor_claim_only — no paper-level KO / siRNA / orthogonal-method corroboration) corroborate but should not, on their own, support a direct_multi_method grade. The schema enforces this cross-block cardinality at record assembly: a direct_multi_method grade with methods=[] (or fewer than 2 direct_surface_accessibility entries from distinct sources) is rejected.

• headline_risks (≤3) selects the consequential sub-blocks of the PROVIDED, frozen accessibility_risks block. See "Headline-risks selection discipline" below for the closed enum values, the anti-other rule, and the historically-misused patterns.

• confidence weighs three things: A1's evidence_grade, the count and severity of A1's contradicting_evidence, and A2's state_dependence — plus the consequential risks in the PROVIDED accessibility_risks block (a documented high-severity decoy / obligate co-receptor dependency is a drag). A direct_multi_method block with no contradictions, low state dependence, and no severe risks is high; conflicting + state-dependent is low.

• triage_signal disagreement. The task message carries the upstream triage verdict. If triage_signal="unlikely" and you call surface_accessibility="high" (or any cross-agent disagreement), you must justify it in confidence_reasoning — the assembled record's validator rejects an empty reasoning under that conflict.

When present in your task message, the Triage prior block carries the full TriageRecord (verdict + reason taxonomy + verdict_reasoning prose + key_uncertainty + confidence) rather than just the rolled-up triage_signal. The triage agent already wrote prose about cell-state / disease / lineage context — read it before deciding your state_dependence and confidence calls:

• reason is structured (stable_surface_marker, cell_state_induced, tissue_restricted_surface, lysosomal_exocytosis, dual_localization, stable_surface_attachment, not_at_surface, unknown). The first five mirror accessibility_modulation.category verbatim. If triage said cell_state_induced and A2's accessibility_modulation is empty, that's a recall miss to flag in confidence_reasoning.
• verdict_reasoning is the triage agent's prose justification (≤800 char). When you're justifying a triage_signal disagreement, the triage prose often contains the exact cell-line / paralog / state condition that explains the disagreement — quote it in confidence_reasoning if useful (substring quotes are fine, you're citing your own task input, not the evidence ledger).
• key_uncertainty is what triage itself flagged as the unresolved question. If A1+A2 resolved it, that's a confidence bump worth stating. If A1+A2 didn't, your confidence should not exceed triage's own confidence on this gene.

The Triage prior block may be absent (no triage run for this gene). Don't fabricate a verdict; just lean on A1+A2 alone.

headline_risks is SELECTED from the PROVIDED, frozen accessibility_risks block — you read its sub-blocks' present / severity flags to pick the consequential ones. You are not authoring the underlying risk call here, only choosing which of the already-made calls to surface as a headline.

The headline_risks enum has five values: shed_form, secreted_form, co_receptor, epitope_masked, isoform_decoy.

Each remaining value names a load-bearing risk that the reader can't easily reconstruct from another structured field. Pick at most three. The post-design-review trim dropped six values that were redundant with existing structured signals — DO NOT try to use them. They are either listed in the structured field where the catalog already filters on them, or moved entirely.

Removed values + where the signal lives now. Read this before you reach for a value you remember from earlier:

• ecd_too_small → READ filters.ecd_accessibility_class (small / minimal / none). The catalog filter is the canonical source; the headline flag was a binary restatement of the same field.
• restricted_subdomain → READ accessibility_risks.restricted_subdomain.present. The headline flag was a direct copy.
• antibody_validation_weak → READ surface_evidence.evidence_grade (weak / conflicting); per-antibody detail lives in AntibodyRef.validation_strategy / validation_strength. The reader sees both surfaces; don't double-encode.
• low_endogenous_expression → DERIVED in the orchestrator from filters.expression_level ∈ {low, absent}; emitted as filters.low_endogenous_expression. The catalog filters on the derived bool, the headline list can't drift from it. Don't try to put it back in headline_risks.
• ligand_unknown → NOT a risk; it's an orphan-receptor status flag. Set filters_llm.has_known_ligand=False for orphan GPCRs / NHRs / kinases, and explain in has_known_ligand_rationale. The catalog treats has_known_ligand=False as a target-tractability signal.
• other → forbidden. If the risk you have in mind doesn't map to one of the five named values, raise it in one_paragraph and let the reader see the prose. The headline list is for catalog filters the reader can scan; other makes it unfilterable.

Three patterns the headline_risks list HISTORICALLY misused (the audit found these in committed samples):

• "EV-associated decoy pool" → DO NOT map to secreted_form. Proteins inside extracellular vesicles / exosomes / microvesicles are shielded by the EV lipid bilayer — they aren't accessible to circulating antibody, so they aren't a decoy in the monovalent-binder sense. EV association is biology-context information, NOT an accessibility risk. Capture it in biological_context.subcellular_localization.dual_localization (compartment=exosome / EV) and/or accessibility_modulation (category=secretion_via_EV if you need to flag a state-conditional EV-trafficking pattern). Reserve secreted_form for FREE-SOLUBLE protein (no lipid shield) in supernatant / serum / plasma that genuinely competes with surface protein for antibody binding.
• "intracellular pool with stress-induced surface" → NOT a headline risk per se; the state-dependence lives in executive_summary.state_dependence. Set state_dependence ∈ {moderate, high} and let that carry the signal — don't squeeze it into headline_risks.
• "antibody cross-reactivity with paralog" → this is what evidence_grade is for. Set evidence_grade to weak or conflicting and explain in grade_rationale; the catalog will filter on evidence_grade directly.

Audit guard. If your evidence_grade ∈ {weak, conflicting} AND your headline_risks is empty, that's fine — the evidence-grade filter already signals the weakness; you don't need to also pick a headline risk just to "say something." If the protein has a genuine shape-specific risk (e.g. it's mostly shed, or co-receptor- dependent), pick the named value for it. Otherwise leave the list empty.

The surface_accessibility enum has five values: high, moderate, low, uncertain, no.

• "no" is for confident negative calls — the deep dive's evidence says this protein is NOT meaningfully at the surface (the triage's verdict="no" end of the scale, extended to the deep-dive verdict). Pick when the literature directly contradicts surface presentation OR when the canonical localization (cytoplasmic / mitochondrial / nuclear) is corroborated by multiple methods AND no ectopic-surface evidence surfaced. A conservative read of an inner-leaflet kinase might land here, for example — the dominant population is cytoplasmic-side and the ectopic-surface story is method-specific.
• "uncertain" is for absence-of-signal cases — neither direction has enough evidence. Use this when you genuinely can't tell.
• Don't pick "no" just because the evidence is weak; that's what confidence="low" + evidence_grade="weak" are for.

The schema splits taxonomy into two orthogonal axes, aligned with SURFACE-Bind (Balbi et al. 2026 PNAS, doi:10.1073/pnas.2506269123). Set BOTH on every record.

subcategory = architecture — how the protein sits in the membrane. Closed enum:

• single_pass_T1 — Type I single-pass (N-term out, C-term in; classical receptor topology). Classical single-pass receptors with cleaved N-terminal signal peptides.
• single_pass_T2 — Type II single-pass (N-term in, C-term out). Surface aminopeptidases and dipeptidyl peptidases, syndecans.
• multi_pass — generic multi-pass (≥2 TM, not 7TM and not tetraspanin). SLC family, claudins, ABC transporters, aquaporins.
• GPCR — seven-pass heptahelical receptor architecture. Kept as a common-name shortcut because 7TM is essentially synonymous with GPCR in practice.
• GPI_anchored — post-translational GPI lipid anchor; no TM span. Complement regulators, prion-class proteins, glypicans.
• tetraspanin — four-pass with large EC2 loop (~80-100 residues).
• other — soluble-cytoplasmic with ectopic-surface story (ER chaperones moonlighting on the cell surface, inner-leaflet kinases, cytoskeletal proteins); inner-leaflet lipid-anchored; or genuinely-uncategorized topology.

llm_family = function — what the protein does, your high-level call. Mirrors SURFACE-Bind's four main classes. (The orchestrator separately attaches deterministic, curator-assigned family tags — hgnc_gene_groups and uniprot_family — alongside this; you do not emit those.) Closed enum:

• receptor — signaling receptors (GPCRs / RTKs / cytokine receptors / integrins / immunoreceptors / NHRs).
• enzyme — surface-exposed catalytic activity. Aminopeptidases, dipeptidyl peptidases, ectonucleotidases, ADP-ribosyl cyclases, surface peptidases, sheddases, matrix metalloproteinases, ectophosphodiesterases. Inner-leaflet kinases count as enzyme by protein identity, regardless of whether the ectopic-surface story is moderate — the catalog filters on what the protein IS, not just where it lives.
• transporter — SLCs, ABC transporters, ion channels, aquaporins, pumps. Subsumes the dropped ion_channel and transporter Subcategory values.
• miscellaneous — adhesion molecules, junction proteins (claudins, occludin, cadherins), tetraspanins, scaffolds (PDZ proteins), structural / cytoskeletal, chaperones, prion-class. Default when none of the above fit cleanly.

A given gene carries one value from EACH axis. Examples:

• A classical single-pass receptor: subcategory=single_pass_T1, llm_family=receptor.
• A 7TM signaling receptor: subcategory=GPCR, llm_family=receptor.
• A Type-II single-pass surface peptidase: subcategory=single_pass_T2, llm_family=enzyme.
• A tetraspanin: subcategory=tetraspanin, llm_family=miscellaneous.
• A multi-pass solute carrier: subcategory=multi_pass, llm_family=transporter.
• An ER chaperone with an ectopic-surface story: subcategory=other, llm_family=miscellaneous.
• An inner-leaflet kinase with an outer-leaflet inversion story: subcategory=other, llm_family=enzyme (kinase by identity).

filters_llm.has_known_ligand is a bool tracking whether the gene has a validated, endogenous binding partner — natural biology, NOT therapeutics. has_known_ligand_rationale (≤300 char) is mandatory and non-empty when has_known_ligand=True (orchestrator rejects placeholders).

The "ligand" here means an endogenous biological binding partner — a natural agonist, cognate receptor, physiological cargo, native substrate, or constitutive heterodimer partner produced by human biology. Therapeutic engagement is a SEPARATE concept and MUST NOT flip this flag to True.

What counts as a known endogenous ligand (→ True):

• Validated natural agonist for a GPCR (chemokines for chemokine receptors, neurotransmitters for neurotransmitter receptors, etc.).
• Cognate receptor for a ligand-class protein (the gene's natural binding partner in physiology).
• Documented constitutive heterodimer / cargo partner (e.g. invariant chain for MHC II, β-microglobulin-equivalent partners).
• A natural binding partner with multiple independent biochemical characterizations (binding affinity, structural data, functional consequence). Proposed-but-widely-cited endogenous ligands count.

What DOES NOT count (→ False if these are all the gene has):

• Therapeutic antibodies, ADCs, CAR-T, bispecifics (clinical or investigational, named or unnamed). Engineered binding ≠ endogenous biology. The catalog reader treats has_known_ligand=True as "natural biology gives you a ready binding pocket / signaling pathway"; therapeutic agents fail that test by definition.
• Small-molecule drugs, blockers, agonists, antagonists — even widely-used pharmacology. Endogenous biology is the bar; pharmacology is not.
• Tool compounds, fluorescent probes, biotinylated binders, photoaffinity ligands. Reagents, not biology.
• Patient autoantibodies / disease-state autoreactive antibodies. Biology of the disease, not endogenous receptor-ligand pairing.

Orphan-class call (→ False): When the only documented binding is therapeutic / pharmacological / investigational, OR when the gene is an orphan GPCR / NHR / RTK with no deorphanized endogenous ligand, set has_known_ligand=False. The rationale names the orphan status

• what's been TRIED but not confirmed (if anything).

Worked example — orphan-like rationale shape:

"Orphan-like: no validated endogenous ligand reported. The clinical ADCs ([drug-1] and [drug-2]) are therapeutic antibody-conjugates, not endogenous biology. The receptor's natural binding partner remains unidentified."

In the True direction, the rationale names the agonist + binding evidence. In the False direction (orphan GPCR with proposed-but- unconfirmed ligand candidate, or tumor antigen with therapeutic binder only), the rationale names the orphan status + what's been tried.

A proposed but widely-cited endogenous ligand stays True if the proposal has multiple independent characterizations; flip to False if the literature explicitly calls the gene orphan / deorphanization-pending.

This applies to the cites YOU author — executive_summary.cited_evidence_ids and the three filters_llm *_rationale inline cites. (The per-risk cited_evidence_ids inside accessibility_risks are authored by the risks builder; you copy them through unchanged — don't add, drop, or re-anchor them.)

Pull cited_evidence_ids from the ledger entries that backed the A1/A2 claim you are integrating. The same a1_evi_* id A1 used inside its methods[].cited_evidence_ids is the one you cite here. Do not paraphrase ledger quotes back into the body of your output — your prose synthesizes, the ledger carries the verbatim text.

Cite only evidence that SPECIFICALLY supports the claim it is attached to. A field's cited_evidence_ids is NOT a "related reading" list for the gene or the section — every id must directly back THAT field's specific assertion. Each cite is rendered next to the claim in the viewer, so an over-broad id reads to the reader as a wrong citation: an expression_level_rationale should cite the expression-level evidence, not a generic surface paper. If a clip in hand doesn't specifically bear on a field's claim, drop it from that field's cites even when it's about the same gene — an empty but correct cite list beats a padded one.

A1's surface_evidence and A2's biological_context are inputs, not outputs — do not rewrite them. accessibility_risks is PROVIDED by the risks builder — copy it through verbatim, never regenerate or edit it (see "Accessibility risks — provided, not generated"). deterministic_features is orchestrator-only; the same goes for the 13 deterministic filter fields (everything in Filters outside the four rollups in filters_llm).

{{ gene_block }}

{{ triage_record }}

{{ deterministic_features }}

{{ a1_output }}

{{ a2_output }}

Produce executive_summary, filters, accessibility_risks, confidence, and confidence_reasoning. Cite only from the merged ledger above. No tool calls available — every claim must be ledger-grounded.