Failure Recovery

Knowledge recovery protects meaning across evidence ingestion, identity resolution, graph construction, conflict reconciliation, and decision briefing. The safe objective is not to restore the largest bundle; it is to recover the largest bundle whose records remain attributable, valid, and honestly bounded.

flowchart LR
    S[Source records] --> I[Normalize with ingestion report]
    I -->|rejected or duplicate| Q[Quarantine and reconcile]
    I -->|accepted| G[Validate evidence graph]
    G -->|broken references| Q
    G --> C[Cluster and resolve conflicts]
    C -->|recommended hold| H[Curator review]
    C --> B[Rebuild state index and decision brief]
    H --> B

Recover from the earliest failed boundary

For ingestion failures, preserve the source record and IngestionReport. Reconcile input count with accepted, skipped duplicate, and rejected counts; inspect duplicate identifiers, rejection reasons, and accepted fingerprints. Correct the source or normalization mapping, then ingest into a copy of the last known-good bundle. Do not silently replace an existing evidence identifier with new content.

For graph failures, validate that claim-to-evidence links reference existing records, relation direction is permitted, and required context is present. Repair the source record or explicit edge. Creating placeholder evidence merely to satisfy graph connectivity converts an integrity failure into unsupported knowledge.

For reconciliation failures, retain all conflicting records and the trust assessment used to group them. A conflict cluster that recommends a hold stays on hold until a curator records a defensible resolution. Do not delete adverse evidence, select the most convenient source, or average incompatible contexts.

Rebuild derived views

After source records are corrected, rebuild the evidence-state index, quality audit, conflict clusters, critical-claim provenance lines, reference disagreement report, and decision brief. Derived trust, triangulation, readiness, and biological-grounding states must be recomputed; editing their serialized values would disconnect them from the bundle.

Compare record identifiers, accepted fingerprints, graph issue counts, conflict membership, freshness, trust and triangulation scores, gaps, recommendation, and provenance with the prior state. Schema compatibility must also pass before a historical bundle is loaded or migrated.

Recovery is complete when counts reconcile, accepted evidence has stable fingerprints and provenance, the graph validates, unresolved conflicts remain visible, and the new brief can explain every recommendation change. Preserve the failed bundle and report as historical evidence instead of rewriting them.