Failure Recovery

An intelligence result is recoverable only when its decision path can be reconstructed. Restoring the previous rank or recommendation without the same evidence cohort, policy lineage, contradictions, and uncertainty would create a new opaque decision rather than recover the old one.

flowchart TD
    D[Unexpected decision] --> E{Evidence complete and current?}
    E -->|no| H[Hold or refuse; repair evidence]
    E -->|yes| P{Policy fingerprint unchanged?}
    P -->|no| N[Evaluate as a new policy decision]
    P -->|yes| R[Replay ranking and scenarios]
    R --> C{Contradictions or instability?}
    C -->|yes| X[Escalate for review]
    C -->|no| V[Validate brief and recommendation]

Preserve the decision context

Before rerunning, retain candidate records, evidence references, freshness assessment, contradiction report, factor definitions and weights, policy lineage or fingerprint, scenario inputs, refusal thresholds, and the original review artifacts. Record package and contract versions. These elements explain why a result was produced; a final score does not.

Classify the failure before changing inputs:

Finding	Recovery response
Missing, stale, or contradictory evidence	Repair or refresh the evidence set; keep the prior recommendation on hold.
Invalid factor, weight, or policy lineage	Correct the policy and issue a new decision context.
Deterministic replay produces different ordering	Treat as an evaluator defect or version drift and escalate.
Scenario actions disagree materially	Preserve the spread and require review; do not average away the disagreement.
Strong claim fails support or localization thresholds	Keep the structured refusal and request the named next check.
Review-board decision lacks rationale or evidence references	Return it for completion rather than inferring approval.

Replay and compare

Replay with the exact evidence identifiers and policy configuration. Compare candidate order, per-factor contributions, confidence, evidence posture, scenario actions, refusals, unresolved questions, and follow-up actions. If the evidence or policy changed, compare the two decisions as separate attributable evaluations and explain each movement.

Do not overwrite review history or convert hold, redesign, escalation, or a refused claim into advance to unblock downstream work. Learning records may inform future posture, but they must not rewrite the evidence and rationale of a past decision.

Recovery is complete when the result is reproducible from a named evidence set and policy, every high-confidence claim has a belief-audit row, contradictions remain visible, and the review packet explains whether the recommendation was restored, deliberately changed, or still refused.