Recorded run. The engine (v0.1.0, ruleset 2026.10.4) produced this report from the bundled synthetic sample, with AI synthesis by llama3.2:3b running locally through ollama. It is stored with the site so it opens without a running engine. Run it live.
Full contextSynthetic sample dataCoverage · cobertura✦ AI · llama3.2:3b
Sample: A pricing refactor that quietly breaks its callers
- Changed
- 3 files · +24 −5
- Symbols
- 5 symbols changed · 1 breaking
- Patch coverage
- 21.1% patch coverage
- Languages
- Python
- Analysed
13findings 9 deterministic 4 heuristic 0 ai-generated
- 2 high
- 9 medium
- 2 low
Review priorityrule-based label
High
2 high-severity deterministic or heuristic finding(s).
A triage label computed from severity and provenance, not a probability of failure.
Showing 13 of 13 findingsj/k to move