RC Rick Collettesecurity · AI · music · people
Pattern

Evidence Bundle Artifact

pattern-evidence-bundle-artifact · canon/patterns/evidence-bundle-artifact/unit.md

Evidence Bundle Artifact

Evidence that grounds a consequential decision should be packaged as a durable artifact: source ids, hashes, freshness state, approval state, snippets, sensitivity labels, signatures, and export formats. The bundle carries source snapshots and the preserved derivation; the decision record links to it.

Problem

Systems often have evidence at decision time but cannot produce it later. Citations point to changed pages, snippets disappear, freshness is forgotten, and approval state lives somewhere else. The decision may have been grounded, but the grounding cannot be inspected. Worse, if the citations were added after generation, the bundle would only preserve citation theater.

Context

Use this pattern when recommendations, approvals, enforcement actions, delegated actions, or AI-assisted decisions depend on evidence that must remain explainable after the decision is made.

Forces

keep grounding future decisions.

moment.

behavior.

observed outcome is not lost.

and retention rules.

  • Sources change over time.
  • Freshness and approval state matter as much as relevance.
  • Parser success, crawl drift, and harvest manifests decide whether a source is healthy enough to
  • Retrieval is an admission boundary, not a neutral transport detail.
  • Admission needs to be recorded as a decision, not assumed from fetch success.
  • Provenance determines whether an evidence artifact can be trusted after it leaves the retrieval
  • Signed artifacts need an admission gate; signature metadata is not useful if consumers ignore it.
  • Signature metadata only helps if the verifier knows which trust root applies.
  • Unsigned artifact exceptions must be captured as policy evidence, not lost as local installer
  • Execution outcome may need a result manifest, not only a pre-execution evidence bundle, so the
  • Reviewers need readable evidence while systems need structured evidence.
  • Sensitive evidence needs sensitivity labels, access rules,
  • Retention rules decide whether the bundle remains replayable long enough to support review.
  • The evidence package must connect to the decision without replacing the decision record.

Solution

Create a durable evidence bundle artifact for each consequential grounded decision. Include source ids, titles, URLs, hashes, fetch dates, freshness status, licenses, approval states, content snippets, sensitivity labels, support/contradiction role, and signatures or tamper evidence. Preserve source-health signals such as parser success, crawl drift, harvest status, and corpus approval when they affect admissibility or confidence.

Export the bundle in machine-readable and human-readable forms when appropriate. Link it to the decision record, approval, policy result, and action.

Consequences

Reviewers can inspect the exact sources, verify hashes, confirm freshness and approval state, and decide whether the conclusion was grounded. The decision record remains the lifecycle owner; the evidence bundle carries the detailed preserved derivation.

Failure Modes

  • Citations are stored without source hashes or versions.
  • Freshness and approval state are checked during retrieval but not preserved.
  • The evidence bundle cannot be connected to the decision record or approval.
  • The same underlying finding cannot be recognized across repeated retrieval or scan runs.
  • Sensitive evidence is exported without sensitivity labels, retention policy, or access rules.
  • Counter-evidence is omitted.

Proof Points

fetch dates, licenses, approval states, snippets, signatures, and replay correlation.

lifecycle before, during, and after governed action.

  • AISDR — evidence bundles package cited sources, hashes,
  • EDDA — evidence is part of the delegated-action

Full source pattern: evidence-bundle-artifact.md.

Incoming References

Case Study 1
Pattern 1
Reference Architecture 1