NATO-1000 Operations and Protocol Manual
Runtime lifecycle
An operator first validates the selected JSON configuration, confirms that network and tools are disabled unless explicitly required, creates a run identifier, performs a health check, executes the request, reviews structured logs, and verifies the audit chain. A release is not considered operational merely because a process exits successfully; the operator must inspect the recorded status, backend, parameters, and error path.
Error classes and response
| Class | Meaning | Required response |
|---|---|---|
| Configuration error | Invalid model ID, parameter, or permission | Correct configuration; do not start runtime |
| Input error | Empty, unsupported, oversized, or malformed input | Reject and quarantine when configured |
| Backend error | Adapter unavailable, timeout, or malformed result | Fail closed, record traceback locally, escalate |
| Audit error | Hash-chain mismatch or unreadable ledger | Stop release; preserve evidence; investigate |
| Permission error | Network/tool access not allowed | Stop and require explicit operator change |
Debugging procedure
Reproduce with the deterministic backend, capture the run identifier, inspect logs/runtime.log, inspect the matching JSONL audit events, and confirm that secrets are absent from both. Use python3 -m compileall -q . for syntax checks and pytest -q for contract checks. Do not debug by enabling unrestricted network or executing model-generated shell commands.
Audit and retention
The audit ledger is append-only by convention and hash-linked by implementation. Verify it before and after incident response. Preserve the original ledger, runtime log, configuration hash, source revision, and quarantined input. Rotation and remote retention should be implemented by the deployment owner with access controls and legal retention rules; this repository does not silently upload logs.
Production adapter checklist
A production adapter must have a documented upstream model and license, deterministic configuration mapping, tokenizer and context limits, timeout and retry policy, resource requirements, redacted telemetry, offline and online tests, prompt injection tests, model-output validation, rollback instructions, and a human approval gate for external side effects. The adapter must not claim to be trained or evaluated until weights and benchmark evidence are actually supplied.