feat(analysis): bind prediction-contradiction refusals to an analysis-run profile

[seonghobae]

Consolidation status

fold_into_landing_vehicle — prediction_contradiction_v1 / tepp.prediction_contradiction.v1 is a Validation / Analysis Run profile and changes the same shared integration surface as #416 (Cargo.lock, crates/analysis_engine/Cargo.toml, crates/analysis_engine/src/lib.rs, DOCUMENTATION.md, docs/TRACEABILITY.md, and docs/adr/README.md). It remains a Draft fold child of #416, not an independent protected-main landing authority.

Current exact child head: e07b2ff9f78ef456ff911b8643710af20921fe54. Current surviving #416 head: 0b7155cc238defb1e55129ff3000658f04b343cf. Retargeting is non-destructive; unique source/tests/doctoring remain here until a real conflict-resolving fold verifies inheritance on the surviving head.

Scientific finding and repair

The initial profile incorrectly treated a deliberately mixed four-class fixture as a production-data prerequisite: assignment_count >= 4 and nonzero covered, partial-overlap, adjacent, and contradictory counts were required during execution/artifact validation. Those Allen support classes are observations, not design strata. A real historical census may legitimately contain only covered predictions or only contradictions; absence of a relation class must remain a truthful zero count.

RED a2892b6ad4f632882f63da39eed2d4706ddf9213 adds covered-only and contradiction-only censuses. Causal source repair a64020152300232cb3214a66b45d97225b6d2b5b removes the artificial minimum/four-class gate while preserving nonempty admitted evidence, MAX_EVIDENCE_UNITS, exact class-count sums, refused_promotion_count + covered_count == assignment_count, cutoff-before-duplicate admission, digest validation, and the fixed claim boundary. 6b0c8de64f41bc11f8bf908e0f9cbe854c1e213c removes predecessor tests that encoded the rejected fixture-as-validity rule.

Fresh docs-to-code review found ADR 0079 still contradicted that repaired source: it remained Accepted on a conflicted fold child and still said a mixed four-class census was required. Commit e07b2ff9f78ef456ff911b8643710af20921fe54 repairs the ADR to Proposed implementation lineage under ADR 0002/0016/0022, removes the false class-strata prerequisite, records sparse-census RED/repair lineage, and leaves canonical ADR identity/status normalization to #437. The shared ADR index remains part of the #416 conflict-resolving fold surface rather than independent architecture authority on this child.

The four-class mixed fixture remains useful coverage because it exercises every PredictionContradictionError mapping; it is not a scientific eligibility condition.

Preserved profile evidence

• Profile/contract prediction_contradiction_v1; schema tepp.prediction_contradiction.v1.
• Inference status unmatched_prediction_is_not_observed.
• Binds existing protected-main prediction_contradiction::refuse_promotion semantics rather than duplicating interval truth.
• Covered pairs may authorize promotion; partial overlap, adjacency, and disjoint predictions stay hypothetical/refused as appropriate.
• Request cutoff equality is typed through KnowledgeCutoff::instant().
• AvailableTime <= knowledge_cutoff admission occurs before duplicate-identity checks, so future-unavailable rows cannot perturb historical identity/count results.
• MAX_EVIDENCE_UNITS bounds execution admission and artifact validation.
• contradiction_agreement_rate remains library-side; inspect payloads remain metric-free.
• Not GPU, not MCMC, not topic birth/split/merge, and not implemented-main.

Exact-head merge authority

Only the eventual surviving #416/successor exact head counts after the fold. Child-head checks/reviews do not transfer. Reacquire all required workflows and qualifying independent review on the eventual surviving head. No self-approval, force push, destructive rebase, routine admin bypass, or child-head evidence transfer.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}