#262 — Structured cross-paper contradiction inventory (synthesis layer)

Status: design (2026-06-08) Issue: #262 (Kong et al. 2026 META #255, Tier B) Target agent: deep-research/agents/synthesis_agent.md Paper anchor: Kong et al. 2026 (arXiv:2605.18661, L. Kong, "Roadmap & User Guide") §7.4.2 — "systems increasingly retrieve and summarize individual papers well, but still struggle with multi-paper relational reasoning, methodological lineage, and cross-paper contradictions." Sibling precedent: #261 (figure/table fidelity, prose-layer) and #214 (synthesis sub-claim inventory, prose-layer) — both shipped prose-layer-only because producer and consumer are LLM agents, not a deterministic parser.

Problem

synthesis_agent (Phase 3, Analysis) already has prose-level contradiction handling:

• Anti-Pattern 3: Unresolved Contradictions (synthesis_agent.md § Anti-Patterns) — bars "Some studies found X while others found Y" stated without analysis.
• Step 3: Contradiction Resolution (synthesis_agent.md § Process → Step 3) — a 5-step prose procedure: identify conflicting claims → compare evidence quality → examine context → assess method → verdict reconcilable/irreconcilable.
• Output "Contradictions & Resolutions" table (synthesis_agent.md § Output Format) — a 3-column prose table (Claim A | Claim B | Resolution).

What is not yet present, per Kong §7.4.2:

1. A structured, inspectable inventory of which paper-pairs were assessed and what the verdict was — the current prose narrative-discusses contradictions (reconcilable and irreconcilable alike) but does not make the set of assessed pairs and the unresolved / checked-clear pairs enumerable for the scholar to confirm.
2. Explicit candidate-pair scoping — the current pass relies on the LLM noticing contradictions ad-hoc during synthesis. For a large corpus there is no stated way to bound which pairs get looked at, and no honest statement of what the scan does and does not cover.

(A calibration reference is discussed later as an optional, non-shipping sanity-check — it is not a deliverable of this change.)

This is the cross-paper-relational sibling of #214 (which decomposed within-review sub-claims) and #261 (figure fidelity). All three enumerate before asserting "addressed."

Scope decision (chosen: prose-layer-only — deliberately departs from the issue's literal acceptance)

The issue's acceptance text overclaims, and following it literally would manufacture the exact "false rigor" the repo rejected in #261. This section records why we depart from the issue wording and ship the lighter form.

The issue acceptance says "Synthesis output schema adds contradiction_pairs[] block" and the rationale invokes "downstream agents (formatter, integrity_verification) cannot programmatically check." Read literally that points to a JSON Schema + lint invariant + CI gold gate. But:

• The named "downstream consumers" — formatter_agent, integrity_verification_agent — are themselves LLM agents reading a markdown report. There is no deterministic non-LLM parser that consumes a contradiction block.
• The judgment that actually matters — "is this a genuine contradiction, or a conditional difference, or no material conflict" — is irreducibly semantic and LLM-side. A JSON Schema would validate that paper_a is a non-empty string; it would not validate that Paper A actually contradicts Paper B. That is field-presence rigor masquerading as fidelity rigor.

This is the same scope call #261 made — its rationale was that machine-validating the YAML shape proves field presence, not fidelity, and adds an unrequested maintenance surface, whereas #213 earned schema+lint only because it has a deterministic machine consumer. #262 has no machine consumer either.

Decision: prose-layer-only. Extend the existing synthesis_agent contradiction section with a YAML-shaped markdown inventory block + a scoping subsection + a worked examples/ file. No JSON Schema, no lint invariant, no CI gold fixture. Vocabulary is aligned with the issue (contradiction, conditional_difference) to avoid drift, but no schema architecture is imported.

The departures from the issue's literal acceptance, to be stated plainly in the PR body:

Field model (codex-corrected — the issue's draft model is not MECE)

The issue's draft field model has four defects, all carried into this corrected model:

1. conflict_type mixed two orthogonal axes. The issue enum {contradictory, conditional_difference, resolved_in_synthesis} folds a conflict nature (contradictory, conditional_difference) together with a resolution status (resolved_in_synthesis). These are separated into two fields.
2. insufficient_overlap was referenced but not in the enum (the resolution_in_manuscript required-when clause names it). It is now a first-class pair_assessment value.
3. No evidence pointers. a_finding / b_finding as free text invites the agent to state claims the papers do not make. Each side gains an evidence pointer grounded in the corpus context.
4. candidate_basis was implicit. Recording why a pair was checked makes the scan's coverage inspectable (and exposes same-camp bias — see scoping).

The shipped block (markdown, inside the existing Contradictions & Resolutions section):

cross_paper_tensions:
  - pair_id: CP-001                      # synthesizer-assigned, stable within the synthesis
    paper_a: "<citation_key or ref slug>"
    paper_b: "<citation_key or ref slug>"
    candidate_basis: "shared RQ subtopic | shared construct/outcome/measure | opposite finding direction | bibliographic coupling | scholar flag | agent-noted cross-cluster"
    overlap_topic: "..."
    a_finding: "..."
    a_evidence_pointer: "..."            # where in the corpus context Paper A's finding is grounded
    b_finding: "..."
    b_evidence_pointer: "..."
    pair_assessment: "contradiction | conditional_difference | no_material_conflict | insufficient_overlap"
    resolution_status: "resolved_in_synthesis | flagged_unresolved | not_applicable"
    resolution_pointer: "Synthesis Report > Contradictions & Resolutions, ¶N"   # required when resolution_status == resolved_in_synthesis
    scholar_confirmation: "pending | confirmed | disputed"   # scholar-set; agent emits 'pending'

Field rules:

• pair_assessment and resolution_status are orthogonal. A pair can be conditional_difference + resolved_in_synthesis, or contradiction + flagged_unresolved. The two axes never collapse into one enum.
• resolution_pointer is required iff resolution_status == resolved_in_synthesis — a claimed resolution must point at where in the report it was resolved, so the scholar can confirm it. For flagged_unresolved / not_applicable the pointer is omitted.
• no_material_conflict and insufficient_overlap pairs MAY be listed but are not obligations to resolve. Listing a checked-but-clear pair documents coverage; it does not manufacture a contradiction. (This mirrors #214's not-mentioned is NOT opposition discipline — a checked pair that is clear is not a tension.)
• scholar_confirmation is scholar-set. The agent always emits pending; it never self-confirms (narrative-side / audit-side separation, below).
• Evidence pointers cite the corpus context already in the prompt — never invented, never read from entry frontmatter (v3.6.7 partial-inversion discipline).

Candidate-pair scoping (recall-limited heuristic — NOT a sound algorithm, NOT MECE)

The issue framed scoping as "cluster-then-pairwise, O(K²) per cluster." That framing overstates what an LLM agent does: it does not execute an O(K²) enumeration; it is given a candidate-selection heuristic in prose. Calling it an algorithm with complexity guarantees would be dishonest unless a deterministic script built the clusters — and none does. So the design states it as bounded, recall-limited candidate generation.

Two failure modes are real and are disclosed in the output, not hidden:

• Cross-cluster contradictions can be missed. Two contradicting papers can sit in different topic neighborhoods; if neither the agent nor the scholar surfaces the cross-pair, it is silently absent. This is acceptable only because the output never claims completeness — it carries an explicit coverage note. If the synthesis ever asserted "all contradictions addressed," this would be a load-bearing hole.
• Bibliographic coupling biases toward same-camp papers. Papers citing the same priors tend to agree; cross-camp contradictions often have low citation overlap. Therefore shared references are an inclusion signal, never an exclusion rule — a low-coupling pair is not ruled out.

The heuristic, stated as guidance:

1. Generate candidate edges (not disjoint clusters). Include a pair if it shares an RQ subtopic, shares a construct/outcome/measure, shows opposite finding direction, is bibliographically coupled, or is scholar-flagged.
2. Bibliographic coupling and shared-RQ are inclusion signals; never use them to exclude a pair from consideration.
3. Deduplicate by sorted (paper_a, paper_b).
4. Emit a Coverage Note: number of papers, candidate pairs considered, classes of pairs not exhaustively checked, and the explicit recall limitation ("this is a scoped advisory scan, not complete pairwise contradiction detection").

Advisory-only / narrative-side discipline (inherited, not new)

synthesis_agent already runs under a strict narrative-side contract: it emits, it does not audit; it does not simulate audit steps; it does not read entry frontmatter (synthesis_agent.md § PATTERN PROTECTION + the citation-emission narrative-side rules). The contradiction inventory inherits all of it:

• The agent emits cross_paper_tensions[] and pending scholar confirmations. It does not decide whether the manuscript adequately addressed a tension — the scholar makes the final call (advisory-only).
• The agent does not claim to have run any external contradiction-checking tool, and output metadata must not claim an audit-passed state.
• Evidence pointers and findings come from corpus context in the prompt, never from frontmatter reads.

Relationship to existing machinery (additive, no duplication)

• Existing synthesis_agent prose contradiction handling — this extends it (makes the existing obligation inspectable), it does not create a parallel artifact. The Step 3 prose procedure and the Contradictions table stay; the inventory block sits alongside them.
• devils_advocate_agent — DA stress-tests a single argument adversarially; it does not enumerate pairwise paper conflicts. Different scope, untouched (#262 non-goal).
• Claim-manifest / claim-alignment audit (#103/#213) — that chain asks "is this manuscript claim supported by its citation?" (deterministic, claim-keyed). #262 asks "did the synthesis acknowledge a known tension between Paper A and Paper B?" Different obligation; it does not reimplement the claim audit. resolution_pointer points at a report location, not a claim id — claim ids are not the primary key. (A future related_manifest_claim_id could be added as an optional link if a resolution becomes a tracked manuscript claim, but that is out of scope here.)
• #261 / #214 — closest precedent; both prose-layer. #262 follows the same scope call.

Calibration (non-blocking manual reference, not a CI gate)

A calibration reference may be recorded to sanity-check contradiction-judgment quality: a small set of paper pairs (e.g. contradictory / conditional / compatible) judged by the agent, reported as a confusion matrix + per-class recall, with the model id, date, and prompt recorded. It is not wired into CI and is not a pass/fail acceptance gate — 20 LLM-judged pairs are too few and too nondeterministic across runs for a hard threshold. Recorded as a behavioral observation, not a guarantee.

Out of scope

• No JSON Schema, no lint invariant, no CI gold fixture (see Scope decision).
• No auto-resolution — the scholar decides (#262 non-goal).
• No credibility ranking to "pick a winner" (would violate #244 epistemological neutrality).
• No change to devils_advocate_agent or the claim-audit chain.
• No #111 dependency (that is a single boolean, not lineage clusters — the issue's own correction).
• No deterministic clustering script; scoping is prose guidance.

Acceptance (corrected from #262 — see Scope decision for departures)

• [ ] synthesis_agent.md Step 3 / Output Format extended to emit a cross_paper_tensions[] markdown block in addition to the existing prose (no schema).
• [ ] Candidate-pair scoping subsection added as a recall-limited heuristic (inclusion-not-exclusion signals; explicit Coverage Note; not framed as an O(N²)/O(K²) algorithm; no #111 dependency).
• [ ] Field model uses orthogonal pair_assessment + resolution_status, evidence pointers, and candidate_basis (codex-corrected, not the issue's non-MECE draft).
• [ ] examples/contradiction_pairs_example.md shows contradiction / conditional_difference / insufficient_overlap rows + a Coverage Note with a cross-cluster recall caveat.
• [ ] Output Format gains a Contradiction Section listing the inventory for scholar confirmation.
• [ ] Cross-reference to the existing prose-level handling clarifying the change is additive.
• [ ] Calibration kept as non-blocking manual reference, not a CI gate (departure recorded in PR).
• [ ] Narrative-side discipline (advisory-only, no audit-simulation, no frontmatter read) confirmed inherited.

Verification

• Prose self-consistency: the worked example round-trips through the new inventory block + scoping note + Output Format Contradiction Section without contradiction.
• Additive check: existing Anti-Pattern 3, Step 3 procedure, and Contradictions table are unchanged (grep-confirm not deleted).
• Line-budget: if a v3.6.7 pattern-protection line-budget test covers this agent, update the expected count for the added section.
• Paper-derived discipline: full citation to Kong et al. 2026 (L. Kong, Roadmap & User Guide, arXiv:2605.18661, §7.4.2); no internal / personal / institutional content in the change; leak scan clean.