2.7 KiB
Evidence-grounded diagnosis decision v2
The decision is whether an adequately supported cause justifies a bounded, verifiable change, not whether every conceivable cause has been excluded. There remains at least one targeted alternative countercheck; unlimited investigation of theoretical or newly proposed hypotheses is not required.
Evidence IDs are usable across hypotheses, including root-cause support and counterchecks. The record's hypothesis ID is provenance only. Technical checks still require actual referenced findings; a schema-only capture is insufficient for causal support or a meaningful countercheck. The evaluator must explain which prerequisite/prediction a finding tests and respect its coverage. Absence within a complete filtered snapshot is not absence of all possible runtime paths.
New diagnosis contexts freeze evaluation_protocol=2. diagnosis_v2 and diagnosis_early_v2 require decision_basis: causal_explanation, scope_limits, sufficient_for_change and contradictions. Each material contradiction cites known evidence and records its resolution and reason. A confirmed output cannot claim sufficient support while marking a material contradiction unresolved. Unfinished alternatives require explicit remaining_uncertainty; another supported hypothesis is not automatically a contradiction or an obligation to research it.
The deterministic validator checks structure, identities, references, nonempty findings, consistency and declared unresolved contradictions. Semantic relevance, whether findings are actually contradictory, adequacy of a countercheck, and whether prose overstates evidence remain LLM judgments. The structure does not prove that a model has detected every contradiction or told the truth. Prompts explicitly reject unjustified universal exclusion and require scoped claims.
Cause confidence and implementation readiness remain separate. Change Request and output checks are unchanged; missing implementation data must not be invented. This does not add dynamic hypothesis creation or change investigation budgets.
Every new diagnosis retry includes prior failure codes, call references and correction guidance from the same review namespace. It never overwrites the frozen base context. The precise request is reproducible after a restart and when validating accepted output. A pending historical call retains its old prompt/hash; after a historical rejection the next fresh attempt uses v2 with feedback. Completed historical diagnoses remain replayable.
The current live case is not rewritten or auto-approved by this change. Its stored response can be structurally tested offline under shared-evidence rules; that is not a new semantic review of its reasoning. No extra summary LLM is used.