Witnora maps framework-specific tool events into a small, stable capability ontology. The deterministic golden dataset pins public tool contracts from five external projects:
| Capability pack | External project | Calibration evidence |
|---|---|---|
| Browser | Browser Use | Checked-in real run plus pinned source contract |
| Coding | OpenHands | Pinned public tool contract |
| Data | LangChain Community SQLDatabaseToolkit | Pinned public tool contract |
| Messaging | Cloudflare Agents Email | Pinned public tool contract |
| Finance | Stripe Agent Toolkit and MCP | Pinned public tool contract |
Each adapter row includes the repository, exact git commit, source URL, observed tool name, expected Witnora capability, and rationale. The dataset is golden-v0.1.json and conforms to agentcert-semantic-golden-dataset.schema.json.
npm run semantic-calibration
The report is written to .agentcert/compatibility/semantic-adapter-matrix.json. Production Acceptance runs the same deterministic calibration and stores its copy beside the signed acceptance report.
The repository also keeps the reviewed v0.1 matrix snapshot so changes to adapter mappings and quality metrics are visible in code review.
exactMatchRate: expected known and unknown cases classified exactly as declared.falseUnknownRate: expected-known tool contracts that Witnora left unknown, divided by all expected-known cases.misclassified: known contracts mapped to the wrong capability.falseKnown: deliberately unknown controls incorrectly absorbed by a broad alias.The optional LLM suggestion provider is excluded. Human corrections remain project-scoped data and do not change the built-in golden score.
The pinned tool names and parameter-sensitive cases map deterministically to the expected ontology. In particular, OpenHands str_replace_editor is classified from its command argument, so view is a read and create or str_replace is a write.
Except for the existing Browser Use evidence, these rows do not execute complete third-party agent workflows. A semantic match does not prove framework runtime compatibility, authorization, gateway enforcement, outcome verification, or business correctness. Those require external smoke runs, controlled adapters, and workflow-specific assurance cases.