Verification Manual
Insight LM's central claim is an Assistant that shows its work. This Verification Manual applies that standard to the product itself: a growing library of verification entries, each documenting a capability, the method used to verify it, concrete acceptance gates, and the result — including failures.
The model is the verification manual of a classical engineering tool: a numbered, versioned library of problems the tool is demonstrated to solve correctly, re-run as the tool evolves.
Entry format
Every entry carries:
| Field | Meaning |
|---|---|
| VM-ID | Stable identifier (VM-NNN) |
| Objective | The user-level capability under verification |
| Method | How it was verified — always through the real application as a user would drive it |
| Acceptance gates | Concrete PASS/FAIL checks defined before the run |
| Result | PASS / FAIL per gate, with evidence |
| Authority class | A: independently reproducible from the entry alone · B: reproducible with the described fixture setup · C: demonstrated, fixture not yet packaged for reproduction |
| Build | The application version the run executed against |
Principles
- Real paths only. Verification drives the actual application; mocked or backend-only checks never count as acceptance.
- Gates before runs. Acceptance criteria are written down before the run, not fitted to its outcome.
- Failures are published. A failed run is a finding, and findings are part of the record.