audit-agent-run-evidence – skill Agent & Tự động hóa cho agent AI

Audit Agent Run Evidence

Overview

Turn an end-to-end success statement into independently decidable claims. Reconstruct what happened from available records, grade each claim against the strongest witness, and keep missing evidence distinct from failure.

This is a read-only audit. Do not rerun tools, approve actions, resume workers, deploy artifacts, or modify evidence unless the user separately authorizes those actions.

When to Use

• Auditing a completed or interrupted agent run from traces and artifacts.
• Checking whether an agent's end-to-end success claim is actually supported.
• Reviewing MCP, gateway, sandbox, checkpoint, retry, memory, approval, or deployment evidence.
• Separating autonomous success from human-assisted or merely requested outcomes.

Do not use this skill to design instrumentation for a future run or to perform the missing actions. It evaluates evidence that already exists.

Establish the Contract

Record these inputs before judging the run:

• declared goal and terminal success criteria;
• run, workflow, task, and parent identifiers;
• immutable code, configuration, model, prompt, tool-schema, and artifact revisions when available;
• actors and trust boundaries: orchestrator, worker, sandbox, MCP server, gateway, human approver, CI, and deployment platform;
• retry, deadline, token, cost, concurrency, and human-escalation budgets;
• supplied evidence inventory and known collection gaps.

Do not silently strengthen the original success criteria. Do not weaken them to match the evidence that happens to exist.

Build a Claim Ledger

Split the overall claim into atomic predicates. Give every row a stable claim ID.

| Field | Required content |

|---|---|

| claim_id | Stable identifier |

| predicate | One falsifiable statement |

| required_witness | Source that can independently prove it |

| evidence_refs | Exact event, log, artifact, or record IDs |

| counterevidence_refs | Conflicting records |

| coverage | Required instances versus observed instances |

| verdict | proven, partially_proven, contradicted, or not_proven |

| gap | Missing field, actor, interval, or verification |

Typical predicates include:

• every required step reached its terminal postcondition;
• sandbox isolation held for every executing worker;
• each required MCP/tool call has a correlated response;
• retries respected idempotency and did not duplicate committed effects;
• a checkpoint was durably written, verified, and actually used for resume;
• parallel branches satisfied the declared join policy;
• memory reads cite a versioned source rather than an untracked summary;
• retry, deadline, token, cost, and escalation budgets were respected;
• approval was granted by an authorized human for the exact artifact and target;
• the platform deployed that same artifact and passed the declared health checks.

Normalize Evidence

Preserve original records and create a normalized event view with:

{
  "run_id": "run-123",
  "event_id": "evt-42",
  "sequence": 42,
  "observed_at": "RFC3339 timestamp",
  "actor": {"type": "worker", "id": "worker-2"},
  "operation": "mcp.search",
  "state_before": "researching",
  "state_after": "researching",
  "attempt": 2,
  "request_id": "req-9",
  "idempotency_key": "task-7:search:2",
  "input_digest": "sha256:...",
  "output_digest": "sha256:...",
  "checkpoint_seq": 3,
  "parent_event_id": "evt-41",
  "status": "succeeded",
  "evidence_ref": "tool-log:991"
}

Use null or unknown for absent values. Never synthesize IDs, timestamps, digests, costs, approvals, or outcomes.

Verify bundle hashes or signatures when supplied. Check duplicate IDs, broken parent links, non-monotonic per-source sequences, impossible state transitions, unaccounted clock skew, and unexplained trace gaps. Treat an integrity failure as counterevidence for claims that depend on the affected records.

Rank Witnesses

Prefer the witness closest to the effect:

| Claim | Strong witness | Insufficient alone |

|---|---|---|

| Code changed | Commit/tree and diff | Agent narration |

| Test passed | Complete test result bound to revision | Command invocation |

| MCP effect occurred | Server or provider audit record | Client request |

| Checkpoint resumed | Durable checkpoint plus verified load event | Checkpoint file exists |

| Human approved | Authorization-system decision bound to artifact and target | Approval requested |

| Deployment succeeded | Platform record plus required health checks | Deployment started |

| Memory grounded a decision | Versioned memory read and citation | Final answer resembles memory |

An orchestrator and its child worker are not independent witnesses when they repeat the same unverified result. A cryptographic digest proves byte identity, not semantic correctness.

Reconstruct the Run

1. Order events by causal links and per-source sequence; use timestamps only as supporting evidence.

2. Build the state-transition path and mark every gap or illegal transition.

3. Link each retry chain by logical operation, request ID, and idempotency key.

4. Link checkpoints to the state they contain and the resume event that consumes them.

5. Preserve every parallel branch outcome; apply the declared all_required, quorum, first_success, or other join rule.

6. Track remaining budgets at each transition. A late success after budget exhaustion is a budget violation.

7. Bind approvals and deployment records to exact artifact digests and targets.

Do not infer successful completion from a final state label when required intermediate predicates are missing.

Assign Verdicts

• proven: authentic evidence covers every instance of the predicate and no reliable counterevidence remains.
• partially_proven: some required instances or fields are proven and the uncovered portion is named.
• contradicted: reliable evidence conflicts with the predicate.
• not_proven: evidence is absent, circular, unverifiable, or only self-reported.

Use not_proven, not contradicted, for missing logs. Use contradicted when the trace shows a failed health check, duplicate effect, unauthorized approver, corrupt checkpoint, skipped required branch, or exhausted budget.

The end-to-end verdict cannot be stronger than its weakest required predicate. Optional diagnostics may remain unproven without failing the run if they were never part of the declared contract.

Report

Return sections in this order:

1. Scope and evidence inventory — run identity, declared criteria, records inspected, integrity checks.

2. Claim ledger — one row per predicate with verdict and exact references.

3. Reconstructed timeline — only state-changing, fault, retry, checkpoint, join, approval, and deployment events.

4. Gaps and counterevidence — identify the affected claims and whether collection can still recover the evidence.

5. Overall verdict — one sentence plus the blocking claim IDs.

Example conclusion:

> partially_proven: repository steps C1-C18 and checkpoint recovery C22 are proven, but deployment success is not proven because C31 has only a client-side start event and no platform health result.

Common Mistakes

• Treating a successful process exit as proof of the business postcondition.
• Counting retries as separate successful logical operations.
• Accepting a child agent's summary as independent corroboration.
• Calling a checkpoint recoverable without observing a verified reload.
• Calling an approval request an approval grant.
• Reporting percentages without listing the denominator and missing instances.
• Recommending instrumentation as though it were evidence from the completed run.

Limitations

• An audit cannot recover facts that no trusted source recorded.
• Provider logs may establish external effects without proving the agent's internal reasoning.
• Redaction may be necessary for secrets and personal data; record the redaction scope and preserve stable references.
• If evidence collection would mutate external state or expose sensitive data, stop and request authorization.