| # NATO-1000 Testing and Release Protocol |
|
|
| ## Required checks |
|
|
| Run `python3 -m compileall -q .`, then `pytest -q`, then execute one offline smoke request for every model ID in `MODEL_REGISTRY`. The release operator records the Python version, operating system, source revision, configuration digest, test output, and audit-chain result. |
|
|
| ## Test classes |
|
|
| | Test class | Coverage | Release gate | |
| |---|---|---| |
| | Unit | Config bounds, adapters, ingestion, redaction, quarantine, errors | Must pass | |
| | Integration | Runtime-to-ledger-to-log flow | Must pass | |
| | Security | Secret absence, passive ingestion, no tool/network defaults | Must pass | |
| | Failure injection | Empty input, oversized input, unknown model, tampered ledger | Must fail closed as designed | |
| | Burst/limit | Repeated offline calls and increasing input sizes | Record observations; no unbounded resource use | |
|
|
| ## Acceptance criteria |
|
|
| Every model adapter must return the same result schema, include its model ID, create start and finish audit events, and leave the chain verifiable. Invalid prompts must not produce a successful result. Ingestion errors must be visible in the ledger and quarantined when configured. No test may depend on a live credential or external network. |
|
|
| ## Known scope |
|
|
| The current suite validates runtime correctness and operational controls. It does not establish model quality, AGI capability, benchmark superiority, factuality, or safe real-world autonomy. Those claims require actual weights, representative evaluation data, independently reviewed metrics, and documented deployment conditions. |
|
|
