Repository navigation
scientific(relation): exclude unpromoted inferred transitions from numerical estimation #673
Description
Problem
At b85c5896beb24d1b8bb415689c1e6cf4a277a2dc, ReferenceTopicInput::new(...) collects every RelationEdge whose kind is a transition and whose endpoints are in the admitted corpus. collect_transition_pairs(...) does not inspect RelationEvidenceStatus.
That crosses an existing owner boundary. relation_graph::RelationEvidenceStatus explicitly distinguishes Observed from Inferred so estimators and audits do not treat LLM/model/heuristic proposals as raw documentary evidence; Inferred is documented as not yet promoted. Today an inferred TransitionsTo edge therefore enters transition_pairs and changes the relational penalty, eta gradient, GGN precision/diagonal variance, and emitted sequence-edge/lineage counts exactly like observed evidence.
This is scientific activation, not a presentation-only defect. It lets unpromoted model/heuristic output alter numerical estimation and uncertainty.
Required repair
topic_measurementnumerical admission must use only directly observed transition edges.- Inferred transition edges may remain in
relation_graphfor provenance/review but must not enter the reference estimator'stransition_pairsuntil some explicit owner promotion changes their evidence status. - A graph containing only inferred transition edges must fail
ReferenceTopicInput::new(...)withInvalidModelInputbecause there is no numerically admissible transition evidence. - A mixed graph must behave as though only its observed transition edges were supplied; no silent promotion or status mutation.
- Keep relation kind/time validation in
relation_graph; do not duplicate source truth or introduce an LLM/model-specific exception intopic_measurement. - Update rustdoc/CHANGELOG and trace the exact RED→repair commits.
Non-goals
No change to relation-edge persistence/provenance, no automatic promotion of inferred edges, no tolerance/estimator hyperparameter change, no scanner/gate change, and no claim that observed status alone solves #658/#671 source/event-time provenance.
This issue is independent of #671: #671 authenticates source/event-time coordinates; this issue prevents already-typed but unpromoted inferred relation evidence from becoming numerical authority.
Activity
test(relation): reject inferred-only numerical transitions (#673)
fix(relation): keep inferred edges out of numerical transitions (#673)
docs(changelog): record observed transition numerical authority (#673)
test(relation): prove inferred transitions stay provenance-only (#673)
seonghobae commentedon Sep 22, 2026
Ordinary-forward implementation on #639:
400f64b2c1cfff106bd7d318355132ff1a98f42f: public integration contract builds a valid inferred-onlyTransitionsTograph and requiresReferenceTopicInput::new(...)to fail withInvalidModelInput. On predecessorb85c5896...,collect_transition_pairs(...)accepts that inferred edge, so the contract is causally red. Its hosted workflows were superseded/cancelled before execution; no hosted RED is claimed.55274ed1914ccc9fcd49edec5aae7b77eff497cd:topic_measurement::referencenow admits transition pairs only when the edge is both a transition kind andRelationEvidenceStatus::Observed; rustdoc now matches ADR 0012's existing "explicit observed predecessor/successor relations only" estimand. No relation graph persistence, estimator arithmetic, tolerance, or hyperparameter changed.29d97647bb1a358d3379f9e3f7401b9ac6291e10.205df30282c6f7b52809e819a1421873d1e81f8d: deterministic reference fits for an observed graph and the same graph plus a distinct inferred cross-cluster transition must be exactly identical, while inferred-only still fails admission.The issue remains open until exact-head hosted gates provide current GREEN evidence. This repair does not solve #658/#671 source/event-time provenance; it only prevents unpromoted inferred relation evidence from becoming numerical authority.
seonghobae commentedon Sep 22, 2026
Follow-on owner-authority gap: #675. #673 now correctly prevents
RelationEvidenceStatus::Inferredfrom becoming numerical authority, butRelationEdge::new(...)still lets any public caller supplyRelationEvidenceStatus::Observeddirectly. #675 therefore keeps #673's estimator filter but requires final Observed activation to be backed by an owner-issued Evidence/Temporal Semantics receipt after #527/#658 settle. Do not weaken #673 or add local booleans/digests intopic_measurement.