State and Persistence

An intelligence record is reproducible only when it preserves the complete basis of the judgment, not merely the selected action.

flowchart LR
    C[Candidate cohort and fingerprints] --> D[Decision record]
    E[Evidence references and posture] --> D
    P[Policy, metrics, weights, thresholds] --> D
    S[Scenario outcomes and uncertainty] --> D
    X[Contradictions, falsifiers, refusals] --> D
    D --> R[Review packet]
    D --> L[Later outcome and learning record]

Durable decision state

• candidate identities, fingerprints, lifecycle states, cohort membership, exclusions, and quality signals;
• evidence references, provenance, freshness, gaps, contradiction state, and readiness result;
• policy identity, metric catalog, weights, thresholds, scenario assumptions, and evaluation time;
• per-scenario action, confidence, hypothesis status, and unresolved questions;
• ranking, ties, robustness, stability, sensitivity, drift, and provenance reports;
• recommendation or refusal, reasons, downgrade chain, gate result, escalation flags, and human-review requirement;
• advisory or enforced mode, including promoter, policy identifier, and rationale;
• review packets, belief audits, challenge results, and later learning records.

Storage boundary

Candidate stores and artifact records support package workflows, but service persistence remains a runtime concern. Knowledge evidence is referenced rather than copied into an intelligence-owned source of truth. Lab outcomes can be linked for learning without becoming retroactive inputs to an older recommendation.

Decision records should be append-only in meaning: corrections or new evidence create a superseding evaluation. Keeping old policy and evidence snapshots allows reviewers to distinguish a changed world from a changed model.

Record supersession explicitly

Supersession cause	New record must identify	Historical record retains
corrected candidate data	corrected fields, source, affected candidates, and validation	original cohort, exclusions, and scores
changed evidence	prior and current Knowledge bundle identities plus relationship changes	evidence snapshot and sufficiency state used at decision time
changed policy	prior and current policy identities, changed constraints or weights, and approval	original objectives, thresholds, tie-breakers, and rationale
expanded scenario burden	added or removed scenarios, falsifiers, and sensitivity ranges	outcomes under the earlier challenge envelope
human override	accountable actor, authority, rationale, conditions, and expiry	computed recommendation and required review state
observed consequence	linked Lab consequence and declared learning rule	recommendation as issued before the outcome existed

A “latest” projection may select the active record, but it must remain possible to reconstruct the chain and compare like-for-like dimensions. Deleting an earlier recommendation destroys evidence about calibration, policy drift, and the cost of past decisions.