Test Report
Machine-generated figures from the recorded run. Regenerate with npm run proof.
Environment
| Field | Value |
|---|---|
| Tool | LocalSutra AI 0.2.1 |
| Node.js | v24.15.0 |
| Platform | win32/x64 |
| Recorded | 2026-11-17T17:19:18.302Z |
Result
| Metric | Value |
|---|---|
| Tests | 126 |
| Passed | 126 |
| Failed | 0 |
| Skipped | 0 |
| Fuzz cases | 9500 |
| Unexpected fuzz crashes | 0 |
Raw logs: test-run.tap, fuzz-results.json, environment.json, SHA256SUMS.txt.
Test pyramid
| Level | Examples |
|---|---|
| Unit | numeric bounds, statistics, canonical JSON, seal, redaction, URL policy |
| Boundary / negative | malformed JSON, oversized bodies, remote endpoints, hostile nvidia-smi output, invalid descriptors |
| Integration | real loopback servers: healthy, redirecting, hanging, flooding, junk, 503 |
| Filesystem | installer idempotence, overwrite refusal, symlink and junction refusal, dry-run |
| Server | token, method, bind address, tampered report |
| End-to-end | CLI commands via child process, audit file outputs, verify tamper detection |
| Fuzz | seeded hostile model descriptors, GPU text, runtime payloads, seal round-trips |
Sample transcripts (real output)
- doctor, fit, probe with nothing listening, probe with remote endpoint blocked, benchmark, privacy, install-skills dry run.
- Full audit report (HTML), Markdown, JSON.
Bugs found and fixed during this release (honest log)
- Installer rejected valid Windows project paths that used 8.3 short names (false symlink error). Fixed with segment-wise
lstatchecks. - Symlink tests failed with
EPERMon Windows. Fixed by using junctions. canonical()handledundefinedinconsistently and could break a seal after a JSON round-trip. Fixed and tested.- Markdown tables were not escaping
|(caught by the audit's own check E05). Fixed. - A GPU name containing markup was accepted by the first parser draft (caught by check D02). Parser now allow-lists characters.
- Report server sent a permissive CORS header and answered POST with 404. Now no CORS header and 405.
What this report does not show
Results on other operating systems (see the CI matrix), live inference on a real Ollama/llama.cpp model (not run), GPU detection on real NVIDIA hardware (parser tested with fixtures; no GPU on the recording host), native IDE skill activation, independent review.