Assurance and Evidence
This page lists what a skeptical reviewer can check, how to check it, and what remains unproven.
Verify it yourself in five commands
git clone https://github.com/shahvatsalofficial/localsutra-ai && cd localsutra-ai
npm test # 126 tests on the recorded run
npm run proof # regenerates docs/proof/*
npm run audit # writes examples/audit/audit-report.{html,md,json}
node packages/cli/index.mjs verify --report examples/audit/audit-report.json
Evidence index
| Claim | Evidence | How to re-check |
|---|---|---|
| Tests pass | test-run.tap: 126 passed, 0 failed, 0 skipped | npm test |
| Hostile inputs handled | fuzz-results.json: 9500 cases, 0 unexpected crashes | npm run proof |
| Safety gates work | audit report: 37 checks with expected vs observed | npm run audit |
| Files unchanged | SHA256SUMS.txt | sha256sum -c style comparison |
| Runs on 3 OS families | GitHub Actions matrix | Actions tab (not asserted here) |
| No dependencies | package.json | npm ls --all |
| No network beyond loopback | code: only fetch calls in core/runtime.mjs go through safeUrl | read the file; audit A01-A03 |
Independent-auditor view
Questions an auditor would ask, with honest answers:
| Question | Answer |
|---|---|
| Were the safety checks written by the same author as the code? | Yes. This is a self-assessment. A third-party review is planned. |
| Do checks use real I/O or only mocks? | Probe checks use real loopback HTTP servers that redirect, hang, flood and send junk. Inference uses a mocked fetch. |
| Was a real model ever run? | No. Live inference: not run. |
| Was a real GPU ever probed? | No GPU on the recording host. The parser was tested with fixture text. |
| Does the verdict mean "safe"? | No. It means the listed checks passed on that host. |
| What did the audit find during development? | Six defects, listed in the Test Report. The checks caught two of them. |
| Could the seal be forged? | Anyone can re-seal an edited file. Signatures are planned. |
Not proven
Behaviour with a real Ollama or llama.cpp install; non-NVIDIA GPU detection; native skill activation in each IDE; absence of egress from external runtimes; performance of any real model; third-party review.
Reproducibility notes
Timings vary by machine and load; the proof records them with sample counts and spread. The fuzz run is seeded for repeatability. All commands need only Node.js 22+.