cross_paper_tensions[] for a literature synthesis (#262)A documentation example (not an executable fixture) showing how the synthesis_agent emits a cross_paper_tensions[] inventory in its Step 3b / Contradictions & Resolutions section, and how a scholar reads it. The corpus is a fictional review of "remote-work productivity", with six papers.
Reference: deep-research/agents/synthesis_agent.md (Step 3b — Cross-Paper Tension Inventory). The inventory is a prose contract — there is no JSON Schema or lint behind it (mirroring the #214 / #261 prose-layer decision, not the #213 schema-layer one): the producer (synthesis_agent) and the readers (the scholar, plus the downstream report/integrity LLM agents) all read prose, so machine-validating the YAML shape would prove field presence, not contradiction fidelity.
cross_paper_tensions:
# Case 1 — genuine contradiction, left unresolved (flagged for the scholar)
- pair_id: CP-001
paper_a: "alvarez2024"
paper_b: "becker2023"
candidate_basis: "shared RQ subtopic; opposite finding direction"
overlap_topic: "effect of fully-remote work on individual task productivity"
a_finding: "Fully-remote engineers showed a 13% productivity gain vs. office baseline."
a_evidence_pointer: "Alvarez 2024, Results Table 3 (n=412, within-subject)"
b_finding: "Fully-remote staff showed an 8% productivity decline vs. office baseline."
b_evidence_pointer: "Becker 2023, §5.2 (n=1,030, cross-sectional survey)"
pair_assessment: "contradiction"
resolution_status: "flagged_unresolved"
scholar_confirmation: "pending"
# Case 2 — conditional difference (not a true contradiction), resolved in synthesis
- pair_id: CP-002
paper_a: "alvarez2024"
paper_b: "okafor2025"
candidate_basis: "shared construct/outcome/measure; bibliographic coupling"
overlap_topic: "remote-work productivity by task type"
a_finding: "Net productivity gain for fully-remote engineers."
a_evidence_pointer: "Alvarez 2024, Results Table 3"
b_finding: "Productivity loss for fully-remote roles requiring frequent synchronous coordination."
b_evidence_pointer: "Okafor 2025, §4.3 (moderation by coordination intensity)"
pair_assessment: "conditional_difference"
resolution_status: "resolved_in_synthesis"
resolution_pointer: "Synthesis Report > Contradictions & Resolutions, ¶2"
scholar_confirmation: "pending"
# Case 3 — checked but no material conflict (documents coverage, not an obligation)
- pair_id: CP-003
paper_a: "becker2023"
paper_b: "dubois2024"
candidate_basis: "shared RQ subtopic"
overlap_topic: "self-reported remote-work satisfaction"
a_finding: "Higher satisfaction among remote staff."
a_evidence_pointer: "Becker 2023, §6.1"
b_finding: "Higher satisfaction among remote staff (consistent magnitude)."
b_evidence_pointer: "Dubois 2024, Fig. 2"
pair_assessment: "no_material_conflict"
resolution_status: "not_applicable"
scholar_confirmation: "pending"
# Case 4 — topics touch but evidence does not actually meet (insufficient overlap)
- pair_id: CP-004
paper_a: "okafor2025"
paper_b: "dubois2024"
candidate_basis: "shared RQ subtopic"
overlap_topic: "remote work and productivity"
a_finding: "Productivity moderated by coordination intensity (objective task metrics)."
a_evidence_pointer: "Okafor 2025, §4.3"
b_finding: "Satisfaction outcomes only; no productivity measure reported."
b_evidence_pointer: "Dubois 2024, §2 (scope statement)"
pair_assessment: "insufficient_overlap"
resolution_status: "not_applicable"
scholar_confirmation: "pending"
# Case 5 — real conditional difference the synthesis noticed but did NOT yet reconcile
- pair_id: CP-005
paper_a: "becker2023"
paper_b: "nguyen2022"
candidate_basis: "shared construct/outcome/measure; opposite finding direction"
overlap_topic: "productivity effect of remote vs. hybrid schedules"
a_finding: "Fully-remote staff showed an 8% productivity decline."
a_evidence_pointer: "Becker 2023, §5.2"
b_finding: "Hybrid-schedule staff showed a 5% productivity gain; the schedule type is the apparent moderator."
b_evidence_pointer: "Nguyen 2022, §4.1 (remote vs. hybrid split)"
pair_assessment: "conditional_difference"
resolution_status: "flagged_unresolved"
scholar_confirmation: "pending"
Coverage Note: 6 papers in corpus (alvarez2024, becker2023, okafor2025, dubois2024, nguyen2022, silva2025); 5 candidate pairs considered (basis among the candidate-edge signals: shared RQ subtopic / shared construct / opposite finding direction / bibliographic coupling). This is a scoped advisory scan, not complete pairwise contradiction detection — the full corpus has 15 possible pairs; 10 were not exhaustively checked. In particular, silva2025 (a manager-perception study) sits in a different topic neighborhood and was not paired against the productivity studies, and most silva2025/nguyen2022 cross-pairs were not enumerated; a genuine tension among the unchecked pairs may exist and is not claimed absent. Bibliographic coupling was used as an inclusion signal only — low-coupling cross-camp pairs were not excluded on that basis. Scholar confirms each resolution_pointer and may flag additional cross-pairs.
CP-001 (genuine contradiction, unresolved)The two papers speak to the same outcome (individual task productivity under fully-remote work) and report opposite directions (+13% vs. −8%). The orthogonal axes carry distinct information: pair_assessment: contradiction says the conflict is real (not a definitional or scope artifact), while resolution_status: flagged_unresolved says the synthesis did not reconcile it. Because it is unresolved, there is no resolution_pointer — the agent must not invent one. The evidence pointers expose the most likely lever: Alvarez is within-subject (n=412), Becker is a cross-sectional survey (n=1,030); a scholar can see the design difference without re-reading both papers. scholar_confirmation: pending until the scholar decides whether to reconcile, present both, or weight by design.
CP-002 (conditional difference, resolved)This is the axis-orthogonality payoff. Naively this looks like another contradiction (gain vs. loss), but pair_assessment: conditional_difference records that the divergence resolves on a moderator (coordination intensity) — Okafor's loss is for coordination-heavy roles, which is compatible with Alvarez's net gain. resolution_status: resolved_in_synthesis + a required resolution_pointer send the scholar to the exact paragraph where the synthesis explains the moderator. Had the issue's original single conflict_type enum been used, "conditional_difference" and "resolved" would have been forced into one value and the scholar could not tell whether the difference was real but reconciled from whether it was never a real conflict.
CP-003 (no material conflict)A pair can be listed without being a contradiction. Becker and Dubois both find higher remote satisfaction at consistent magnitude. Recording it as no_material_conflict / not_applicable documents that the pair was checked (coverage), without manufacturing a tension. This mirrors #214's discipline that a checked-but-clear position is not opposition.
CP-004 (insufficient overlap)The topics touch ("remote work") but the evidence does not meet: Okafor reports productivity, Dubois reports only satisfaction. insufficient_overlap records that the pair was considered and set aside because there is no shared measured construct to contradict — distinct from no_material_conflict (where they do measure the same thing and agree). Without this value the agent would be pushed to force a verdict on a non-comparison.
CP-005 (conditional difference, unresolved)The third legal pairing on the orthogonal axes. Like CP-002 the divergence is a conditional_difference (schedule type — remote vs. hybrid — is the apparent moderator, not a true contradiction), but unlike CP-002 the synthesis noticed it without yet reconciling it: resolution_status: flagged_unresolved, so there is no resolution_pointer (omitted, exactly as the field rule requires for non-resolved status). CP-002 (conditional_difference + resolved_in_synthesis) and CP-005 (conditional_difference + flagged_unresolved) together show that the resolution axis moves independently of the assessment axis — the same conflict nature can be resolved or still open.
The note is not boilerplate — it is the honesty mechanism that makes the recall limitation acceptable. It states the denominator (15 possible pairs), what was skipped (10), and names the paper in another neighborhood that was not paired. This is why the inventory may ship as a scoped scan: it never says "all contradictions addressed." Remove the note and the same block would be an overclaim.
devils_advocate_agent — stress-tests a single argument adversarially; it does not enumerate pairwise paper conflicts. Different scope; untouched.