← Về thư mục
📄 / / academic-research-skills / docs / design / 2026-05-29-kong-259-domain-evidence-profiles-spec.md

Kong #259 — Discipline-Relative Domain Evidence Profiles

Status: prompt + PCR field + standalone reference-doc design (R7 PCR-native rewrite) Date: 2026-05-29 Issue: #259 (Tier A, [ACCEPT-MEDIUM]; attaches to #244 epic alongside #246) Scope: Let a scholar select a domain_evidence_profile at paper intake so that literature_strategist_agent adjusts its post-retrieval screening gates to the scholar's discipline, instead of applying a single Western evidence-based-medicine (EBM) pyramid to every field. Advisory only, scholar-selected, no auto-detect, override first-class.

Carrier (R7, the load-bearing rewrite decision): the active profile lives as a single-value field on the Paper Configuration Record (PCR) — exactly like Style Profile (intake_agent.md:238). It is NOT carried by the Material Passport, and it is NOT a numbered handoff Schema. Rationale below ("Why PCR, not the Material Passport").

Single consumer = literature_strategist_agent. The profile is produced by academic-paper's intake_agent (Phase 0) and consumed only by literature_strategist_agent (Phase 1) — both in academic-paper Stage 2, intake before literature_strategist. source_verification_agent is NOT a consumer (it runs in deep-research Stage 1, upstream of intake; serving it would require a deep-research-stage profile — deferred, see Out of scope).

Anchor: Kong et al. 2026 (arXiv:2605.18661) §7.4.6 (kong2026_full.txt:L2213-L2227, PDF p.39): "Extending AI-assisted research to chemistry, biology, medicine, materials science, physics, and social science requires more than retraining on domain papers. These fields differ in evidence standards, experimental infrastructure, safety constraints, data availability, and community norms."

Why PCR, not the Material Passport (R7 root-cause correction)

The first six review rounds (R1–R6) all tried to carry the profile on the Material Passport and could never pass the gate. The R7 audit (opus subagent carrier-lifecycle trace + codex gpt-5.5 xhigh whole-pipeline scan + scholar's stated design intent) found that the carrier choice itself was the root cause, not any of the individual symptoms the six rounds kept patching.

The scholar's design intent (highest authority): the Material Passport is an opt-in side input — it lets a user wire their own KB/database in to bring literature into the pipeline. A core feature must NOT depend on the passport; with no passport, the pipeline must still run. Repo proof: shared/handoff_schemas.md:505 (literature_corpus is Optional, "ARS does not produce these entries itself"), :576 ("input port for user-owned literature"), :580 ("corpus-first, search-fills-gap" — present is used, absent searches normally).

The decisive precedent (Style Profile): schema text once called Style Profile "carried by the Material Passport," but that is a simplification, not the runtime mechanism. In practice intake_agent writes it to the PCR (intake_agent.md:192, :196, :238) and draft_writer_agent reads it from the PCR (draft_writer_agent.md:46) — never the passport. Style Profile only also appears on the passport because it has a cross-skill consumer living in deep-research. The domain profile's only consumer (literature_strategist_agent) is in academic-paper, in the same run as the producer — there is no cross-skill consumer, so there is no reason to involve the passport at all.

What "profile selection is a PCR concern" buys (every following item follows from the carrier choice):

Form of the PCR field. A new PCR row, mirroring Style Profile exactly:

| **Domain Evidence Profile** | [effective_value]                          |   # ship-ready selection
| **Domain Evidence Profile** | unknown_user_defined (requested: clinical) |   # reserved-fallback: effective + the scholar's request, parenthesized

The row stores the resolved effective profile (one of the 4 ship-ready enum values) — this is the only value the consumer reads. When the scholar requested a reserved profile, the row appends the request in parentheses (unknown_user_defined (requested: <reserved>)) so the scholar's intent is visibly acknowledged in the user-facing PCR table rather than looking like the system ignored their choice (the parenthetical is display-only; the consumer parses only the leading effective token). For a ship-ready selection there is no parenthetical (effective == request, so nothing is hidden). The PCR holds the currently effective profile — the active state, not a history. Override is "re-run intake, the row is overwritten" (same lifecycle as every other PCR row). There is no append-only selections[] ledger (R7 design decision: the PCR has no history mechanism — Style Profile does not have one either — and adding one would be unrequested abstraction for a single Phase-1 consumer; see "Override" for what is and is not retained).

Boundary

This patch is the profile-selection layer. It is a checklist/search adjustment, not an autonomous grading or domain-detection layer.

Relationship to the existing PCR Discipline field (R-3 reconcile — "each manages its own")

The PCR already has a Discipline field (intake_agent.md:228), and literature_strategist_agent already uses it to pick databases and terminology — its ### Step 2: Database Selection table maps Discipline → primary databases (literature_strategist_agent.md:46-55: Education → ERIC, CS → IEEE Xplore, etc.). The new Domain Evidence Profile is a second discipline signal, so the two must be reconciled.

Resolution: separation of duties, no override (maintainer-approved 2026-05-30).

Profile → implied-discipline map (so "implies a discipline" is deterministic, not guesswork):

Profile Implied discipline(s) for the mismatch check
cs_ml CS / Engineering / Technology
humanities_interpretive Humanities
general_social_science Social Science or Policy (matches either — no mismatch warning if Discipline is either)
unknown_user_defined none — the neutral default never triggers a mismatch check

The mismatch advisory fires only when the PCR Discipline falls outside the profile's implied-discipline set. unknown_user_defined and an absent profile never warn. The check is a string-category comparison against this table, not an inference.

This keeps the change minimal (no edit to the Step 2 database table) and consistent with the advisory-only / no-unrequested-abstraction discipline.

Profiles

Ship-ready enum (exactly 4)

general_social_science | cs_ml | humanities_interpretive | unknown_user_defined

unknown_user_defined is the default and the neutral fallback. The other three are real, populated profiles.

Reserved (documented, NOT in enum)

clinical | wet_lab | materials_physics | legal_case_based | education

Reserved profiles are named in the reference doc with a "not in enum yet — selecting this falls back to unknown_user_defined" note. They ship per demand in later releases. They are deliberately kept OUT of the enum so that intake cannot select a profile whose checklist does not exist (which would manufacture false rigor).

Design decisions (settled in brainstorm 2026-05-29, carrier revised R7 2026-05-30)

These resolutions deliberately diverge from three of the issue's literal acceptance criteria; §"Acceptance mapping" records each divergence and its reason.

# Decision Rationale
Q1 carrier The active profile is a single-value PCR field (Domain Evidence Profile row), produced by intake_agent, consumed by literature_strategist_agent. Not the Material Passport; not a numbered handoff Schema. (R7 root-cause rewrite — supersedes the R1–R6 Material-Passport / Schema-13 design.) The passport is an opt-in side input the pipeline must run without; binding a core feature to it was the root cause of six failed rounds. The profile's only consumer is in the same skill as the producer, so no passport carrier is needed. Mirrors the real Style Profile mechanism (PCR row, intake_agent.md:238 / draft_writer_agent.md:46). Avoids the Schema-13 collision with the sprint-contract schema (shared/contracts/README.md:10).
Q2 no append-only ledger The PCR holds only the currently effective profile (single value). No selections[] event ledger. Override = re-run intake overwrites the row. (R7.) The PCR is a single-state snapshot table with no history mechanism (intake_agent.md:221-239); Style Profile is stored the same single-value way. A per-entry append-only ledger would be unrequested abstraction for one Phase-1 consumer, and it is what forced the R1–R6 "last-entry resolution," "malformed-ledger carve-out," and reset/resume-ordering complexity. Dropping it dissolves all three.
Q3 discipline reconcile Separation of duties: Discipline keeps database/terminology; the profile owns evidence admissibility only; mismatch warns, neither overrides. (R7.) Discipline already drives the Step 2 database table (literature_strategist_agent.md:46-55); letting the profile touch it would contradict the advisory-only contract and risk regressing database selection. Minimal-edit, no override.
Q4 naming reconciliation In the new profile reference doc, the profile-centric table carries forward the substance of the field-centric Field-Specific Adjustments rows from source_quality_hierarchy.md. The deep-research file's own table is left untouched. (R7 — changed from R1–R6, which replaced the deep-research table in place.) source_quality_hierarchy.md is read by deep-research's source_verification_agent + bibliography (deep-research/SKILL.md:392, source_verification_agent.md:38). Editing it would leak #259 into deep-research (R-5). Instead the profile table lives in academic-paper's own reference and copies the field guidance; the shared file is not modified, so deep-research is unaffected.
Q5 #246 forward-dependency Add a forward-reference note in the profile reference doc; do NOT pull aggregation into #259, do NOT ship placeholder logic. #246 (grade aggregation) is unbuilt and the maintainer is doing #259 first. A forward note avoids a dangling cross-link without scope-creeping #259 into #246.
Q6 profile granularity Add a single academic-paper/references/domain_evidence_profiles.md file with a 4-row structured table. Do NOT create a deep-research/references/domain_profiles/ directory of per-profile files. 4 profiles (3 real, 1 neutral fallback), advisory-only, no aggregation, do not justify a directory of mini-specs. A single dedicated file keeps the agent reading one place while giving enough structure to consume reliably. Split into files only when profiles grow long, are independently maintained, or are reused by tooling.

Change set (3 product/prompt edits + test/CI)

Product/prompt edits (the user-facing behavior change): B (intake produces the PCR row), C (new academic-paper profile reference doc), E (literature_strategist consumes the PCR row). There is no Change A (no Schema), no Change D (no second consumer).

Test/CI edits (required, separate from the 3 product edits — see Test strategy for content): a new scripts/check_domain_evidence_profile.py lint, a new scripts/test_check_domain_evidence_profile.py mutation suite, and a one-line wiring into .github/workflows/spec-consistency.yml. These are implementation artifacts the plan MUST deliver; they are not counted among the 3 product edits because they ship no runtime behavior, but an implementer must not skip them.

Round history: R1–R6 carried the profile on the Material Passport as "Schema 13" with up to 6 product edits (orchestrator producer + two consumers + in-place edit of the deep-research reference file). The R7 PCR-native rewrite collapses the product surface to 3 edits with no passport, no Schema number, and no edit to any deep-research file.

B. Intake production — academic-paper/agents/intake_agent.md

intake_agent is the producer of the PCR Domain Evidence Profile row. Add a profile-selection step as a new intake step under the heading ### Step 12: Domain Evidence Profile — same ###-under-## Interview Protocol level as the existing ### Step N steps (intake_agent.md:110-204), numbered after the current Step 11 Funding — and add the | **Domain Evidence Profile** | ... | row to the Paper Configuration Record output table (after the Style Profile row).

Profile-value rules (prose validation, mirroring the PCR's other single-value fields — NO JSON Schema file):

C. Profile definitions — new file academic-paper/references/domain_evidence_profiles.md

A new reference doc in academic-paper's own references directory (the consumer, literature_strategist_agent, lives in academic-paper). No deep-research file is touched — this is the R-5 isolation. Contents:

  1. Carry forward the substance of the existing ## Field-Specific Adjustments table from source_quality_hierarchy.md (lines ~132-143; 6 rows — Medicine/Health, Education, Social Science, Policy, Humanities, Technology). The source table is read, not edited; its guidance is copied into the new profile-centric framing so no per-field guidance is silently dropped. Specifically:
  2. general_social_science, cs_ml, humanities_interpretive profile rows absorb the substance of the Social Science / Technology / Humanities rows respectively. The Policy row's substance is folded into the general_social_science profile row as well (Policy has no dedicated profile; its expert-panel/context-dependent guidance is merged into general_social_science's notes).
  3. The Medicine/Health and Education rows map to reserved profiles (clinical, education). Their existing adjustment text is preserved verbatim in a legacy mapping note within this new file (not deleted, not moved out of the deep-research file — copied).
  4. Preservation is historical / non-normative. The legacy note records what the field-centric table says, for reference and for the eventual clinical / education profiles. It does NOT change runtime behavior: until those reserved profiles ship, Medicine/Health and Education runs use the neutral unknown_user_defined pyramid, exactly as every other unmapped selection does. Label the note "historical reference — non-normative; current behavior for these domains is neutral unknown_user_defined until the clinical / education profile ships."

Legacy mapping (old field label → disposition in this new file): - Social Science → substance folded into general_social_science profile row (normative, via the profile) - Technology → substance folded into cs_ml profile row (normative, via the profile) - Humanities → substance folded into humanities_interpretive profile row (normative, via the profile) - Policy → substance folded into general_social_science profile row (normative, via the profile; no dedicated Policy profile) - Medicine/Health → text preserved verbatim in the historical/non-normative legacy note; runtime behavior = neutral unknown_user_defined until the reserved clinical profile ships - Education → text preserved verbatim in the historical/non-normative legacy note; runtime behavior = neutral unknown_user_defined until the reserved education profile ships

  1. A ## Domain Evidence Profiles section with a 4-row table. Columns: Profile / Standard evidence types / Common provenance requirements / Critical gaps to surface / Reserved-note. (The gaps column is named "Critical gaps to surface" — NOT "disqualifying" — because the profile is advisory: it tells the agent what weaknesses to flag, it never disqualifies a source or changes the grade.) Plus a separate short list of the 5 reserved profiles with the "not in enum" note.

  2. A #246 forward-reference note: "Discipline-relative grade aggregation (how these evidence expectations roll up into an Overall Grade) is tracked separately in #246 and is not yet implemented; until then the A-F Overall Grade lookup in source_quality_hierarchy.md applies unchanged."

  3. An advisory-only statement at the top of the file: the profile changes which evidence types the literature screening admits; it never changes the A-F Overall Grade and never blocks ship.

Why the deep-research file is not edited (R-5): source_quality_hierarchy.md is a deep-research reference consumed by source_verification_agent and bibliography for the evidence pyramid + grading rubric. #259's only consumer is in academic-paper. Copying the field guidance into a new academic-paper file (rather than replacing the deep-research table in place) keeps deep-research's behavior byte-identical and contains #259 entirely within academic-paper.

No dual-table read by the consumer (the key fact): literature_strategist_agent does not read source_quality_hierarchy.md at all — it has its own screening tree, quick-assessment checklist, and quality gates (literature_strategist_agent.md:413/436/468) and does not reference the deep-research evidence-pyramid file (verified: no source_quality_hierarchy mention in the consumer). So there is no "agent reads both the old and the new field table" collision: the consumer reads only the new domain_evidence_profiles.md, the old Field-Specific Adjustments table is read only by deep-research's agents, and the two never meet in one context. The new profile table is seeded from the old table's substance (a one-time authoring copy), not read alongside it at runtime.

E. literature_strategist consumption — academic-paper/agents/literature_strategist_agent.md

literature_strategist_agent is the sole consumer of the profile. It resolves the active profile from the PCR Domain Evidence Profile row (NOT the passport, NOT a ledger). A profile must influence the post-retrieval screening gates (it does not touch Step 2 database selection — that stays Discipline-driven per the reconcile rule). Graceful-fallback cases, none of which block:

There is no HANDOFF_INCOMPLETE concern: the profile is a PCR field, not a handoff object, so the general handoff-validation convention (shared/handoff_schemas.md:7-9) never applies to it. A profile defect simply falls back to neutral with the advisory.

Loosen-only / additive contract (realizes INVARIANT 5). Every sub-edit below is monotonic admit-only: under a non-neutral profile a screening gate may admit an evidence type it would otherwise wrongly exclude, but it MUST NOT exclude, down-rank, or fail any source that the neutral gate currently admits. Where a gate combines neutral and profile criteria, combine by OR (union of admissible), never by replacement.

# Monotonic admit-only resolution (pseudocode — the load-bearing definition)
profile = resolve_from_PCR_row()            # one of 4 enum values; absent/non-enum -> unknown_user_defined (+ advisory on non-enum)

# The neutral screening tree (literature_strategist :413) has SEVERAL branches. The profile
# loosens ONLY the evidence-type / publication-type / currency / provenance branches. It does
# NOT touch the universal-quality branches, which apply to every source regardless of profile:
UNIVERSAL_GATES = [relevance_to_RQ,          # abstract must address >=1 aspect of the RQ (:421-423)
                   methodology_not_fatally_flawed,   # no obvious design flaw / fatal methodology (:425-428)
                   not_predatory_or_fabricated]      # predatory-publisher / fabrication checks
PROFILE_LOOSENABLE_GATES = [peer_review_requirement,  # "is it peer-reviewed?" (:415-419)
                            publication_type,          # preprint / proceedings / archival / primary-source admissibility
                            currency_window,           # ">= 50% in last 5 years" gate (:480)
                            provenance_expectation]     # field-appropriate provenance

def admit(source):
    if not all(g(source) for g in UNIVERSAL_GATES):
        return False                          # profile can NEVER bypass relevance / methodology / predatory
    if neutral_passes(source, PROFILE_LOOSENABLE_GATES):
        return True                           # never tighten: anything neutral admits on these gates stays admitted
    if profile != unknown_user_defined and profile_admits(source, profile, PROFILE_LOOSENABLE_GATES):
        return True                           # only loosen: profile ADDs admit paths on the loosenable gates only
    return False
# The OR (union of neutral + profile) is scoped to PROFILE_LOOSENABLE_GATES, applied AFTER the
# universal gates pass. Under unknown_user_defined, admit() == the full neutral tree exactly.
# A cs_ml preprint that is off-topic, fatally flawed, or predatory is still EXCLUDED — the profile
# only forgives "not peer-reviewed / is a preprint", not "fails core quality".

The sub-edits, all reverting to current behavior under fallback cases (a)/(b)/(c). Each loosens only the peer-review/publication-type/currency/provenance gate it names; none touches the universal relevance / methodology / predatory gates:

  1. Screening decision tree (### Literature Screening Decision Tree, ~line 413): the current tree excludes non-peer-reviewed unless gov/white-paper gray literature, which would drop cs_ml preprints and humanities_interpretive archival/primary sources. Add a profile-aware branch on the "is it peer-reviewed?" node only so the profile's standard evidence types are additionally admissible (preprints includable under cs_ml; primary/archival sources includable under humanities_interpretive), tagged by type rather than excluded. The downstream relevance ("abstract addresses the RQ?") and methodology ("no obvious design flaw?") nodes still apply unchanged — a profile-admitted preprint must still pass them. The branch only adds an admit path at the peer-review node; nothing the neutral tree admits becomes excluded, and no universal-quality node is bypassed.
  2. Quality quick-assessment checklist (### Literature Quality Quick Assessment Checklist, ~line 436): the existing 5-item quick-assessment total score penalizes preprints/archival via the Journal-ranking item (the other four items — methodological rigor, relevance, citation count, data quality — are universal-quality and stay in force). Under a non-neutral profile, a source passes if it meets the neutral quick-assessment outcome OR meets the profile's evidence-type expectations on the journal-ranking item only (union scoped to the publication-type axis, not replacement — a source that passes the neutral score must still pass even if it does not match the profile; and a source the profile would admit on publication type must still clear the methodological-rigor / relevance items, which are universal).
  3. Quality gates (## Quality Gates, ~line 468): the >= 70% peer-reviewed (line 479) and >= 50% currency (line 480) pass criteria contradict humanities_interpretive (primary/older canonical texts) and cs_ml (preprints). Make these two corpus-ratio gates profile-relative in the loosening direction only: under a non-neutral profile the admissible set expands (preprints count toward the cs_ml peer-reviewed-equivalent ratio; canonical texts do not count against humanities_interpretive currency). The thresholds are never raised, and a corpus that passes the neutral gate always passes the profile-relative gate. The other quality-gate rows (source count, annotation completeness, matrix coverage, research-gap count) are not evidence-type gates and are untouched.
  4. Quality-criteria checklist — the mirror constraints (## Quality Criteria, literature_strategist_agent.md:558-559): the same two ratios are restated here in prose — "Source quality distribution: majority should be peer-reviewed" and "Recency: >50% of sources from last 5 years (unless historical topic)". These are a mirror of the Quality Gates row and MUST receive the identical profile-relative loosening, or a profile-admitted cs_ml preprint / humanities_interpretive canonical text would pass the gate at line 479-480 only to be re-tightened by this duplicate at 558-559. Apply the same union/loosen-only rule: under a non-neutral profile, "majority peer-reviewed" counts the profile's peer-reviewed-equivalent types, and the recency criterion does not penalize the profile's canonical/older sources. (This is the mirrored-artifact catch — both the gate and its restated criterion move together; changing one without the other reintroduces the contradiction the gate fix was meant to remove.)

Search-string enrichment (optional, additive, NOT a database change): under a non-neutral profile, Step 3 (Search String Construction, ~lines 57-61) MAY add profile-relevant search strings (e.g. cs_ml → also include preprint-server-style query terms). This is purely additive to the search strings; it does not change the Step 2 database table (which stays Discipline-driven). Never drops a search string the neutral strategy would have used.

Upstream filter carve-out (Step 3 search filters + Step 7 publication-type inclusion criteria, literature_strategist_agent.md:57-70): the Step 3 Filters: peer-reviewed, [year range] line and the Step-7-style "Peer-reviewed journals / books / conference proceedings (include) vs blog posts / news (exclude)" publication-type rows are not the same corpus-ratio mirror as sub-edits 3/4, but they are upstream hard filters that could exclude profile-admissible evidence (preprints, primary sources) before it ever reaches the screening tree. Under a non-neutral profile these MUST receive the same loosen-only treatment: the profile's standard evidence types are added to the includable set (peer-reviewed filter relaxed to peer-reviewed-equivalent for cs_ml; year-range relaxed for humanities_interpretive canonical texts), never tightened, never dropping anything the neutral filter would have kept. (Same monotonic rule as the gates — this just names the upstream filters so an implementer does not leave them as hard excludes that starve the admit paths sub-edits 1-4 open downstream.)

Mismatch + enrichment interaction (the [PROFILE-DISCIPLINE-MISMATCH] corner — addresses the "search can't feed the admit path" concern): when Discipline and the profile disagree (e.g. Discipline: Education, profile cs_ml), the Step 2 databases stay Discipline-keyed (Education databases), so a profile-opened admit path (e.g. "preprints admissible") may have little to admit if those databases surface few preprints. This is a known advisory-only limitation, not a failure: the corpus is not bricked (peer-reviewed Education sources still flow through the unchanged universal + neutral gates exactly as today), it is merely not enriched with the profile's preferred evidence type. Step-3 search-string enrichment partially mitigates by adding profile-relevant query terms even under a mismatched discipline. The [PROFILE-DISCIPLINE-MISMATCH] advisory tells the scholar both signals are active in their own lanes and that aligning Discipline to the profile's field (re-run intake) would let database selection feed the opened admit path. Nothing blocks; the scholar decides.

This composes with the existing ## Distributional Skew Advisory (Kong #257) section without contradicting it (skew advisory measures distribution; profile sets admissibility — orthogonal).

Missing-profile gate (R-4 — explicit per-path handling)

The resolution rule ("read the PCR row; absent ⇒ neutral") must be correct across every path that reaches literature_strategist_agent, not just a clean intake. Each path below is accounted for — some apply the profile, some fall back to neutral — and every one is intended behavior, not a defect:

Path to literature_strategist Does academic-paper intake run? Profile state Result
Standard academic-paper full with intake yes (full Step 12) PCR row set by Step 12 profile applies
deep-research → academic-paper handoff (with or without a bibliography) yes — intake_agent runs, skipping only redundant questions (intake_agent.md:37-50, deep-research/SKILL.md:352), not Step 12 PCR row set by Step 12 (intake asks the profile; inferred discipline MAY pre-fill the suggested default, scholar confirms) profile applies — literature_strategist_agent still runs even when a bibliography is handed in: it "goes directly to Phase B (full-text assessment), skipping Phase A" search (literature_strategist_agent.md:550), so the profile has a live consumer screening the handed-in sources
Scholar explicitly skips Phase 1 entirely (academic-paper/SKILL.md:139, e.g. mid-entry with a finished draft, no literature screening at all) intake runs, but Step 12 takes the Phase-1-fully-skipped carve-out (Change B): no placebo prompt PCR row recorded unknown_user_defined + [NO-PROFILE-NEUTRAL] advisory explaining no literature screening will occur, so a profile would have no consumer neutral, surfaced — the scholar is told why a profile is not asked, and that running Phase 1 would let one apply. NOT a placebo.
plan → full transition (literature_strategist runs from plan artifacts) plan intake is the simplified 3-question intake → Step 12 exempt → no row no row neutral fallback (see note below)
Mid-entry pipeline start (e.g. Stage 2.5 with existing draft) no — intake skipped entirely no row neutral fallback
Re-entry after reject/restructure or resume-from-checkpoint no new intake (resumes existing PCR) PCR row carries whatever the prior intake set (or absent if intake never ran) profile applies if the row was set, else neutral fallback
Standalone deep-research run n/a — never reaches academic-paper intake or literature_strategist n/a profile not in play

Correction vs R1–R6 framing: the deep-research → academic-paper handoff is the common path; it does run intake_agent (skipping only redundant questions) AND literature_strategist_agent still runs (Phase-B screening over the handed-in bibliography, literature_strategist_agent.md:550), so Step 12 runs, the profile IS asked, and it has a live consumer — a profile-applies path regardless of whether a bibliography was handed in. The genuine always-absent-row paths are plan → full and true mid-entry (intake never sets the row). Resume is row-based, not always-absent — it keeps whatever the prior intake wrote, so it is a missing-profile path only when the prior intake left the row absent. The only neutral-by-carve-out path is the scholar explicitly skipping Phase 1 entirely (no literature screening at all), where Step 12 records neutral + [NO-PROFILE-NEUTRAL] rather than prompting. Resolve by reading the row; the table classifies paths only to confirm each was considered.

plan → full note (a deliberate gap, surfaced not silent): because plan-mode intake is the simplified 3-question intake that omits Step 12, a scholar who goes plan→full is never asked for a profile and silently gets neutral evidence standards. This is acceptable for plan mode's lightweight purpose, but to avoid a silent quality degradation, literature_strategist_agent MUST, when it runs with no PCR profile row in a full-mode run, emit a one-line [NO-PROFILE-NEUTRAL] advisory noting that neutral (single-pyramid) standards are in effect and the scholar may re-run intake to select a domain profile. (This makes "absent ⇒ neutral" observable rather than invisible — it never blocks.)

The gate is therefore the single rule "resolve from the PCR row; absent ⇒ neutral + [NO-PROFILE-NEUTRAL] advisory in full mode" (Change E cases (a)) applied uniformly. The table exists so a reviewer can confirm each run shape was considered; there is no run shape where a missing profile breaks the pipeline, and no run shape where a missing profile degrades quality invisibly.

Acceptance mapping (incl. deliberate divergences)

Issue acceptance criterion How this spec satisfies it Divergence
#1 Intake adds domain_evidence_profile with enum = 4 ship-ready values + default unknown_user_defined Change B (intake Step 12 + PCR row) aligned
#2 deep-research/references/domain_profiles/ directory + 4 profile md files Changed to a single academic-paper/references/domain_evidence_profiles.md with a 4-row table (Change C) DIVERGES — maintainer-approved (Q6): single dedicated file, not a directory; and it lives in academic-paper (the consumer's skill), not deep-research, to avoid leaking #259 into the shared deep-research reference (R-5). To be noted on #259 at close.
#3 Reserved profile list documented with "not in enum yet" note Change B (profile-value rules) + Change C (reserved-note column + reserved list) aligned
#4 source_verification_agent.md + literature_strategist_agent.md consume the profile Only literature_strategist_agent consumes it (Change E). source_verification_agent is dropped from the consumer set. DIVERGES — maintainer-approved: source_verification_agent runs in deep-research (Stage 1), upstream of the producer (intake, Stage 2), and serving it would require a deep-research-stage profile. Its domain-aware verification is deferred (see Out of scope). To be noted on #259 at close.
#5 #246 cross-links to relevant profiles Changed to a forward-reference note (Change C.3), because #246 is not yet implemented DIVERGES — maintainer-approved (Q5): forward-reference note instead of a live cross-link, to avoid a dangling reference. The live cross-link lands when #246 ships.
#6 User can override profile mid-pipeline, recorded in Material Passport Override = re-run intake overwrites the PCR Domain Evidence Profile row (Change B); active = the row's current value (Change E) DIVERGES — maintainer-approved (R7 Q1/Q2): recorded in the PCR, not the Material Passport, and as the current effective value, not an append-only history. The passport is an opt-in side input the pipeline must run without, so a core feature cannot be carried there; and the PCR has no history mechanism, so override is overwrite (the scholar's latest choice is authoritative). If a future release needs an override audit trail, it can add a history block then — out of scope now. To be noted on #259 at close.

INVARIANTS

  1. Enum cardinality. The ship-ready enum (the PCR row's stored effective values) is exactly 4 (general_social_science, cs_ml, humanities_interpretive, unknown_user_defined); the 5 reserved values (clinical, wet_lab, materials_physics, legal_case_based, education) are NOT valid stored effective values (they may appear only as a scholar request, which resolves to unknown_user_defined).
  2. Reserved fallback. A request ∈ reserved → the PCR row stores unknown_user_defined + intake surfaces an explicit advisory naming the requested reserved value. No reserved profile ever silently activates a checklist.
  3. Single-value PCR carrier (no ledger). The active profile is the current value of the PCR Domain Evidence Profile row — a single state, not a history. Override overwrites the row (same lifecycle as Style Profile and every other PCR row). There is no append-only selections[] ledger and no "last-entry" resolution rule. Safe because the sole consumer runs once in Phase 1 and the revision loop (8→5→6, academic-paper/SKILL.md:270) does not re-run Phase 1, so the consumer never observes a stale-vs-current ambiguity. (Supersedes the R1–R6 passport-ledger design.)
  4. Graceful fallback. Absent PCR row / row = unknown_user_defined / row value not in the 4 enum (incl. hallucinated or reserved-as-effective) → the consumer applies the current neutral single-pyramid behavior, unchanged, with no block; the non-enum case also emits [PROFILE-UNRESOLVED]. The profile is a PCR field, not a handoff object, so the HANDOFF_INCOMPLETE convention never applies — a profile defect never blocks.
  5. Advisory only. A profile never changes the A-F Overall Grade and never blocks manuscript ship. In literature_strategist, profile-relative screening gates are monotonic admit-only: they only admit evidence types the neutral gates would wrongly exclude (combining by OR / union), and never exclude, down-rank, or fail any source the neutral gate currently admits. The profile never touches Step 2 database selection.
  6. No auto-SELECT. No agent activates a domain_evidence_profile without scholar confirmation. The producer is intake_agent; discipline inference (deep-research handoff or Step 1 topic interview) MAY suggest a default, but the scholar confirms.
  7. No discipline loses existing guidance. Carrying forward the 6-row field-centric table into the new profile reference MUST account for every row: Social Science / Technology / Humanities / Policy fold their substance into ship-ready profile rows (Policy → general_social_science); Medicine/Health + Education are preserved verbatim in a historical, non-normative legacy note whose runtime behavior is neutral unknown_user_defined until their reserved profiles ship. No row is dropped, and the preserved note is non-normative (no "applies until" normative claim). Scope clarification (Education does not regress): "existing guidance" here means the documentation in the deep-research Field-Specific Adjustments table. That table is consumed by deep-research's source_verification/bibliography, not by literature_strategist_agent — the academic-paper literature screening already runs on its own tree/checklist/gates and never used the deep-research field table. So for an Education paper, literature_strategist's screening behavior is identical before and after #259: #259 only adds (advisory, monotonic) admit paths under ship-ready profiles, and Education maps to a not-yet-shipped reserved profile, so it stays on exactly the neutral behavior it has today. Nothing Education-specific is taken away, because nothing Education-specific was in the consumer to begin with. The deep-research table that is preserved (copied into the legacy note) keeps serving its actual consumers unchanged (INVARIANT 9).
  8. #246 boundary. #259 references grade aggregation as a forward dependency only; it ships no aggregation logic and no placeholder aggregation code.
  9. Deep-research reference untouched (R-5). source_quality_hierarchy.md (and every other deep-research file) is not modified by #259. The profile-centric table lives only in the new academic-paper/references/domain_evidence_profiles.md; the field guidance is copied into it, not moved. deep-research's source_verification_agent + bibliography behavior stays byte-identical.
  10. Discipline reconcile — separation of duties. The PCR Discipline field continues to drive database/terminology (Step 2); the Domain Evidence Profile drives evidence admissibility only. Neither overrides the other. A mismatch emits [PROFILE-DISCIPLINE-MISMATCH] and both signals proceed in their own lanes; nothing is blocked.
  11. Row-based resolution ⇒ absent row ⇒ neutral (visible). The active profile is resolved by reading the PCR row, never by classifying the entry path. An absent row ⇒ neutral unknown_user_defined fallback (INVARIANT 4 case (a)) + [NO-PROFILE-NEUTRAL] in full mode so neutral is visible. Paths that leave the row absent: plan → full (simplified intake omits Step 12) and true mid-entry start. Resume-from-checkpoint preserves the existing row — it is neutral only if the prior intake left it absent, otherwise the resumed profile applies (so resume is NOT an unconditional missing-profile path). The deep-research → academic-paper handoff is also not a missing-profile path — intake sets the row and literature_strategist_agent still runs (Phase-B screening of the handed-in bibliography), so the profile applies whether or not a bibliography accompanied the handoff. The only carve-out is the scholar explicitly skipping Phase 1 entirely, where Step 12 records neutral + [NO-PROFILE-NEUTRAL] (no placebo, Change B). The Missing-profile gate table enumerates every path and which side it falls on. No path degrades quality invisibly.
  12. Request/effective coherence. The scholar's request ∈ {4 ship-ready ∪ 5 reserved}; if ship-ready, the stored effective value == the request; if reserved, the stored effective value == unknown_user_defined with an advisory. No other combination is valid — a stored effective value can never silently differ from the scholar's confirmed ship-ready selection.

Test strategy

Prose contract (the profile is a PCR field, not a JSON schema — mirrors how Style Profile is validated by prose, no JSON Schema file). The checker is honest about its reach: a markdown structural checker verifies documentation surface (presence/shape of required text), NOT runtime semantics. Several INVARIANTS are semantic; the split below makes that explicit.

scripts/check_domain_evidence_profile.py — documentation-surface checks (deterministic):

  1. intake_agent.md documents the 4 effective enum values + the 5 reserved values; the PCR output table contains a Domain Evidence Profile row; the request/effective coherence rule, the reserved-fallback advisory text, the reserved-fallback display form (unknown_user_defined (requested: <reserved>)), the ### Step 12: Domain Evidence Profile heading, the Phase-1-fully-skipped carve-out, and the [PROFILE-OVERRIDE-NO-RESCREEN] advisory (emitted when an override lands after Phase 1 has already run OR was explicitly skipped — the checker requires both halves of that condition, and a mutation fixture that drops the "or was skipped" half must FAIL) are present (INVARIANTS 1, 2, 12).
  2. academic-paper/references/domain_evidence_profiles.md exists and has a ## Domain Evidence Profiles section with a 4-profile table whose gaps column is "Critical gaps to surface" (NOT "disqualifying"), plus the 5 reserved names with the "not in enum" note (INVARIANTS 1, 2, 5).
  3. The consumer agent (literature_strategist_agent), in its ### Domain Evidence Profile Resolution block, contains a "resolve domain_evidence_profile from the PCR row" instruction AND the graceful-fallback cases (absent / unknown_user_defined / non-enum incl. hallucinated) AND the three advisory tags it (the consumer) is specified to emit — [NO-PROFILE-NEUTRAL] (absent-row, full mode), [PROFILE-UNRESOLVED] (non-enum), [PROFILE-DISCIPLINE-MISMATCH] (reconcile) — plus the universal-gate carve-out language ("relevance / methodology / predatory still apply") so the monotonic contract is documented (INVARIANTS 4, 5, 10). It also confirms source_verification_agent is NOT given a profile-resolution step. (The fourth tag, [PROFILE-OVERRIDE-NO-RESCREEN], is emitted by intake_agent, not the consumer — checked in item 1.)
  4. domain_evidence_profiles.md carries the advisory-only statement + the #246 forward-reference note (INVARIANTS 5, 8).
  5. The legacy field guidance is present in domain_evidence_profiles.md, labeled non-normative AND containing the preserved Medicine/Health + Education text; the general_social_science profile row references the folded Policy substance (INVARIANTS 7).
  6. No deep-research file is modified by #259: source_quality_hierarchy.md does NOT contain a ## Domain Evidence Profiles section and its ## Field-Specific Adjustments table is unchanged (a guard asserting the profile table did NOT leak into the shared file) (INVARIANT 9).
  7. The PCR Domain Evidence Profile row is named in intake_agent.md as the single carrier (INVARIANT 3). Carrier-regression guard (scoped by concrete heading delimiters to be deterministic): the implementation MUST place the profile instructions under two named headings — ### Step 12: Domain Evidence Profile in intake_agent.md (matching the existing ### Step N level under ## Interview Protocol) and ### Domain Evidence Profile Resolution in literature_strategist_agent.md. The guard scans only the byte range from each of those headings to the next same-or-higher-level heading, and asserts that within those ranges the profile is NOT stored on or resolved from a Schema 13 number, a selections[] ledger, or the Material Passport. It does not scan the whole repo, the whole agent file, or this spec's prose: the spec deliberately cites Schema 13 / selections[] / Material Passport as the superseded R1–R6 design, and unrelated repo sections legitimately use Material-Passport / Schema language for other features (e.g. shared/handoff_schemas.md, the corpus blocks of the same agent files). The named-heading range makes the scope a deterministic delimiter rather than a fuzzy "instruction block," so the guard neither false-fails on contrast prose nor misses a regression. (Requiring the exact headings is itself one of the documentation-surface checks.)

Negative fixtures + mutation test (scripts/test_check_domain_evidence_profile.py): deliberately (a) add a 5th effective enum value, (b) rename the gaps column back to "disqualifying", (c) strip one graceful-fallback case, (d) delete the preserved Medicine/Education text, (e) remove the Policy fold reference, (f) add a ## Domain Evidence Profiles section to source_quality_hierarchy.md (must FAIL — the R-5 leak guard), (g) inside the Step 12 / profile-resolution instruction blocks, store/resolve the profile via selections[] / Schema 13 / Material Passport (must FAIL — the carrier-regression guard, scoped to instruction blocks so the spec's own historical-contrast mentions do NOT trip it) — each must make the checker FAIL (so it cannot trivially accept-all). A companion positive test asserts the guard does not fire on the historical-contrast paragraphs (lock the scope). Wired into .github/workflows/spec-consistency.yml.

Out of the checker's reach — relies on plan-stage review + worked example: INVARIANT 2's runtime reserved-fallback behavior; INVARIANT 5's no-grade/no-block runtime behavior and its monotonic admit-only ("never tighten") half (a markdown surface checker cannot verify that a gate only loosens); INVARIANT 6's no-auto-SELECT; INVARIANT 8's no-aggregation-logic; INVARIANT 10's mismatch-advisory runtime behavior; INVARIANT 11's no-intake-no-profile runtime path. These are agent-prompt semantics. The implementation plan MUST include worked examples exercising (i) a reserved-fallback, (ii) a scholar override at intake, and (iii) a discipline mismatch; the dual-track reviewers verify the prompt text — the linter does not claim to enforce these.

Out of scope (forward work)