Agent Assure Evidence Diff

Same approval. Broken evidence trail. CI blocked.

The candidate agent preserved the visible decision, but lost the material-claim evidence link required for governance review.

Gate classificationCI Gate Blocked Candidate RegressionCore thesisOutput equivalence is not process equivalence
Decision outputapprove → approvepreservedrecommendation and outcome stayed aligned
Decision-field delta0 decision fields changedno driftrecommendation/outcome across 1 case
Evidence invariantclaim-duration → nonefailed invariantcontrol: material_claims_have_evidence
CI gateblocked as designedblockedcandidate release stopped before merge

Executive TL;DR

  • Decision drift: none observed in recommendation/outcome fields.
  • Process regression: claim-duration lost its required evidence link (material_claims_have_evidence).
  • Process scope: 1 of 1 cases; findings or missing links by case; 1 missing material-claim link of 1 baseline claim link
  • Release action: candidate regression caught and blocked in CI before release.

Trust Boundary

Local deterministic evidence for human review.

This report is not a compliance attestation. This artifact does not certify safety.

Agent Assure Evidence Diff · generated locally · evidence packet available1 / 5

The diff that matters

Process Evidence Diff

Same final approval, different governed path. Agent Assure compares the process evidence around the decision, not just the visible answer.

Baseline Process
claim-durationevidence-duration
material claim has evidence
Candidate Process
claim-durationnone
material claim evidence link missing
CI Gate
control responseblocked
candidate blocked

Key Finding

1 material claim link regression

The candidate preserved the visible decision fields but dropped a required material-claim evidence link. This is a link-level regression: the source evidence may still be present for another claim.

CaseMaterial claimBaseline linkCandidate linkReason
shared-source-multi-claim
claim-duration
claim-durationevidence-duration
claim-durationnone
MATERIAL_CLAIM_MISSING_EVIDENCE

Process-affected cases are sourced from candidate findings and missing material-claim evidence links, not from preserved decision fields.

CaseProcess statusDecision fieldsChanged fieldsProcess findingsClaim-evidence links
shared-source-multi-claim
failed invariant
preserved
evidence refs, claim-evidence links
1 material claim link regression
MATERIAL_CLAIM_MISSING_EVIDENCE
Baselineclaim-durationevidence-durationCandidate- claim-duration none
Agent Assure Evidence Diff · generated locally · evidence packet available2 / 5

Decision preserved, process failed

Visible Output Stayed Stable

The business decision did not change. The governance evidence around that decision did.

Decision-Field Comparison

Decision fields (recommendation, outcome)
preserved
Case coverage
same case IDs
Baseline evaluation state
pass
Candidate evaluation state
fail
CaseBaseline recommendationCandidate recommendationBaseline outcomeCandidate outcomeDecision fields
shared-source-multi-claimapproveapproveapproveapprovepreserved

Process Invariant Findings

CaseControlTargetReason codeStateMessage
shared-source-multi-claimmaterial_claims_have_evidenceclaim:claim-durationMATERIAL_CLAIM_MISSING_EVIDENCEfailfixture-declared material claim has no evidence link
Agent Assure Evidence Diff · generated locally · evidence packet available3 / 5

Technical appendix

Before and After Traceability Detail

Engineers can inspect claim IDs, evidence refs and sources, linked evidence, policy states, provider/model fields, tool names, operational counters, and measured usage summaries used in the run.

Full Baseline Process Evidence

CaseClaimsEvidence refsLinked claimsPolicy statesHuman reviewProvider/modelToolsOperational metricsMeasured usage
shared-source-multi-claim
claim-duration
evidence-duration source=guideline-duration
claim-duration
none
not required
not recorded
benefit-policy-lookup
not recorded
not observed

Full Candidate Process Evidence

CaseClaimsEvidence refsLinked claimsPolicy statesHuman reviewProvider/modelToolsOperational metricsMeasured usage
shared-source-multi-claim
claim-duration
none
none
none
not required
not recorded
benefit-policy-lookup
not recorded
not observed
Agent Assure Evidence Diff · generated locally · evidence packet available4 / 5

Evidence packet appendix

Release Evidence for Review

Fixture equivalence, artifact paths, and SHA-256 digests make the blocking decision reproducible without exposing local machine paths.

Comparison Classification

Classification
new_failure
Baseline state
pass
Candidate state
fail
Provenance changes
none
Verdict findings
1 item
  • MATERIAL_CLAIM_MISSING_EVIDENCE

Fixture-Equivalence State

State
pass
Review note
Baseline and candidate fixture material remained equivalent for comparison.

CI Gate Result

Result
blocked
Evidence packet
packet-duration
Packet limitations
1 item
  • Local deterministic fixture evidence for human review.

Claim Boundary

This report is not a compliance attestation.

This artifact does not certify safety.

Display note: external URLs and sensitive-looking values may be redacted in this HTML; source artifacts remain unchanged.

Artifact Paths

ArtifactPath
baseline run setbaseline.runset.json
candidate run setcandidate.runset.json
evidence packetevidence-packet.json

Artifact Digests

RolePathSHA-256
evaluation-summarypacket artifact digestaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa
compiled-suitecompiled.jsonaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa
candidate-runsetcandidate.jsonaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa