Architecture diagrams¶
These diagrams are source-bound views of Crucible at commit 0a59896. Counts and weights come from a ground-truth.json extract, generated by go run ./tools/diagrams extract; every diagram states the source commit and review/extraction date.
The Mermaid and D2 files below are the auditable specifications: each states the source commit and the review or extraction date it was checked against.
| diagram | question answered | source |
|---|---|---|
| 1: System architecture | What crosses the generation, execution, assessment, and record boundaries? | generated Mermaid |
| 2: Strategy selection | How do category weight and category size combine into per-strategy probability? | generated Mermaid |
| 3: ggml-rpc exchange | What happens over time in one harness_rpc_graph iteration? | reviewed Mermaid sequence |
| 5: Crash lifecycle | Which evidence states and terminal outcomes can a crash enter? | reviewed Mermaid state diagram |
| 6: Evidence model | How are findings, artifacts, evidence, remedies, and disclosures related? | reviewed Mermaid ER diagram |
| 7: Formats and harnesses | Which formats are structure-aware, and which projects do the harness families drive? | generated D2 + SVG |
| 8: Build reproducibility | Where can the custom mutator silently disappear? | reviewed Mermaid flowchart |
| 9: Closed loop | Which feedback paths exist, and which remain absent? | generated Mermaid |
| 10: Operator journey | Where do manual work and human gates sit in the operating path? | reviewed D2 + SVG |
| 11: Correctness gates | Where does the tool refuse to proceed, and what is each refusal protecting? | generated Mermaid |
Generated diagrams are regenerated in CI and fail on diff. Reviewed diagrams carry a 90-day review-age warning.