Assurance and Evidence
This page tells a skeptical reader how to check IntentSutra instead of trusting it. It is written from the standpoint of an independent auditor.
What was independently produced
| Artifact | Where | What it shows |
|---|---|---|
| Raw TAP log | test-results/test-run.tap | every test name and result: 116 passed, 0 failed, 2 skipped |
| Fuzz results | test-results/fuzz-results.json | 5398 generated inputs; 398 audit-log edits all detected |
| CLI transcripts | samples/cli-*.txt | a real lifecycle, including refused actions and exit codes |
| Sample audit, honest | samples/audit/audit-report.html | verdict PASS, 11 checks |
| Sample audit, forged | samples/audit-forged/audit-report.html | the same workspace after a silent edit: verdict FAIL |
| Digests | test-results/SHA256SUMS.txt | SHA-256 of each proof file |
The forged sample is the important one: an audit that never fails proves nothing. Here the audit demonstrably fails when state is edited without a matching audit event.
Verify it yourself in five minutes
git clone https://github.com/shahvatsalofficial/intentsutra-ai && cd intentsutra-ai
npm test # unit, integration, security, CLI tests
npm run proof # regenerates every proof file from scratch
node packages/cli/index.mjs verify --report docs/proof/audit-sample/audit-report.json
Then open docs/proof/audit-sample-forged/audit-report.html and read finding AUD-08.
What a reviewer should challenge
- Does the audit trust the registry? No. It recomputes content hashes, evidence hashes and the attested state from raw data, then cross-checks counts against the audit log.
- Could
auditbe gamed by deleting the log? An empty log with a non-empty registry fails AUD-08. Truncating the newest events is not detectable locally; keep the folder in Git. - Is the seal a signature? No. It detects modification of a report; anyone can regenerate a new report with a new seal.
- Are imported test results trustworthy? Not by themselves. They are labelled so in every report, and AUD-06 fails if a label is upgraded.
Not covered (honest list)
Native execution inside any IDE · signatures and key management · multi-process concurrency · legal or regulatory conformance · accessibility conformance testing with assistive technology · behavior on macOS and Linux outside hosted CI · performance at large scale.
Severity policy
critical and high failures produce verdict FAIL and exit code 2 with --strict. medium and low give PASS-WITH-FINDINGS. Informational checks never fail.