WattSutra AI — Independent self-audit
2026-11-16T17:56:27.058Z · WattSutra 0.2.1 · Node v24.15.0 · win32/x64
PASSverdict (scoped)
17/18checks verified
0failed
NOT_RUNenergy, carbon, cost gates
Evidence integrity, claim hygiene, comparison fairness and packaging. Energy, carbon and cost gates are NOT_RUN: no hardware meter was used.
Live sample
Baseline trial 1.185 ms vs candidate sieve 0.254 ms. Ratio 4.66x, 95% bootstrap interval [3.46, 5.34]. Latency gate PASS. Timings are from the machine that ran this audit, not an energy claim.
Findings
| ID | Check | Status | Severity | Evidence | Reproduce |
|---|---|---|---|---|---|
| A01 | Benchmark runs verify and keep quality parity | verified | info | baseline 9a1da4495e209034f9ee4c77481ef39150d65938e70d0bf8a36d6114c0d904b9 candidate 07d0c0ac64c6fa76987341a166ef263b1fbb18c3f6bf4aeb61cfd76a1819f2e3; prime count 2262, sum 21171191 | wattsutra bench --manifest fixtures/baseline.json |
| A02 | Edited reports are rejected (integrity hash) | verified | info | mutated wallMs.mean; verifyRun threw integrity mismatch | node --test tests/ |
| A03 | Internally inconsistent reports are rejected even with a recomputed hash | verified | info | mean edited, hash recomputed; validateRunShape recomputed stats from samples and refused | node --test tests/ |
| A04 | Measured runs contain no energy, carbon or cost values | verified | info | energyKWh/carbonGCO2e/cost are null; injected energyKWh=0.5 was refused | wattsutra bench |
| A05 | Incomparable workloads are refused | verified | info | n=20000 vs n=3000 refused | wattsutra compare |
| A06 | Quality mismatch yields a FAIL gate | verified | info | gate=FAIL | wattsutra compare |
| A07 | Environment mismatch downgrades the gate to REVIEW_REQUIRED | verified | info | mismatched: cpuModel | wattsutra compare |
| A08 | Too few samples never earns PASS | verified | info | n=2 per run, gate=REVIEW_REQUIRED | wattsutra compare |
| A09 | Estimator applies no default rates | verified | info | unknown: energyKWh, operationalGCO2e, cost | wattsutra estimate (empty assumptions) |
| A10 | Estimator rejects unknown fields and ambiguous energy input | verified | info | typo key and dual energy inputs refused | wattsutra estimate |
| A11 | Modeled and imported values are labelled | verified | info | kinds: modeled, imported-meter-observation (attributionUnverified=true) | wattsutra estimate |
| A12 | HTML report escapes hostile content | verified | info | payload escaped; CSP meta present: true | wattsutra report --format html |
| A13 | Hashing is deterministic regardless of key order | verified | info | canonical JSON, code-unit key order | node --test tests/ |
| A14 | Skills and agent manifest are consistent | verified | info | 20 skills, 20 manifest entries | wattsutra audit |
| A15 | No credentials in shipped files | verified | info | regex scan for GitHub, npm, AWS, OpenAI-style keys and private keys: no matches | wattsutra audit |
| A16 | Installer refuses to overwrite customised skills | verified | info | second install refused to clobber edited SKILL.md | wattsutra install |
| A17 | Hardware energy instrument | not-run | info | RAPL unavailable; tools: nvidia-smi=unavailable, powermetrics=unavailable, scaphandre=unavailable; measuredEnergyAvailable=false | wattsutra doctor |
| A18 | Statistical confidence | verified | low | latency gate PASS; n=12/12; ratio 4.66x; 95% bootstrap interval [3.46, 5.34]; CV baseline 42.6% | wattsutra compare |
Honest limits
- A hash detects edits; it does not prove a number is honest. A forger who rebuilds samples and hash is not detected without signatures.
- No hardware meter was used. Nothing here measures electricity or CO2e.
- This is a self-audit by the tool: reproducible evidence, not third-party certification. Re-run with
npx wattsutra audit.