- Build a small, tested decision workflow, not a collection of agent names. Keep explicit state, distinct role prompts, typed decisions, bounded tool calls and an auditable trace.
- Preserve the difference between deterministic policy checks, replay fixtures and actual model inference. Never report a test double as model quality or copy another model's benchmark as our result.
- Use synthetic fixtures or explicitly licensed public datasets with immutable source revisions, attribution and reproducible selection. Do not import private contracts, financial accounts, employer data or personal records. Keep public benchmark tasks distinct from the fictional workflow fixtures.
- Legal examples classify text for a human review queue. Financial examples reconcile documents and flag discrepancies; they do not approve payments, lending, investments or legal outcomes.
- Prefer permissively licensed local decision models. Keep weights outside Git, pin tested model revision and package versions, record provenance and attribution. A compatible API is not evidence of equivalent accuracy or licensing.
- No paid API calls, external tracing, credential changes or infrastructure deployment without specific owner authorization. Disable telemetry and use local traces by default.
- Run offline unit/property tests, the fixture evaluation harness and any affected model adapter checks. Report actual runtime, case counts, failures and coverage; keep expected labels outside runtime workflow input.
- Freeze public task definitions, selection and prompts before model inference. Preserve failures in denominators, report each task separately, disclose known training overlap and unknown contamination, and do not silently truncate inputs or select cases after observing model output. Keep model weights outside Git and retain earlier reference runs unchanged.
- Preserve owner edits and true dates. Disclose AI assistance and leave owner study/review unclaimed until reported.