Scenario:
Candidate/reviewer:
Date:
Assumptions:
Timeline check
- Clarification complete by minute 5
- Requirements/non-goals by minute 8
- First architecture by minute 13
- Core flow/state by minute 25
- Deep dive by minute 34
- Evals/operations by minute 40
- Failures/tradeoffs and close by minute 45
Score
| Dimension | 0-4 | Evidence/feedback |
|---|---|---|
| Requirements and risk | ||
| User journey and authority | ||
| Architecture/topology | ||
| Tools/context/data | ||
| State/memory/durability | ||
| Security/identity/isolation | ||
| Evaluation/testing | ||
| Observability/scale/cost | ||
| Tradeoffs/evolution | ||
| Communication | ||
| Total / 40 |
Required checks
- Candidate states what the model may not do.
- At least one external effect has idempotency/reconciliation semantics.
- Untrusted content is separated from trusted instruction.
- Tenant/user identity reaches data and tools.
- Long-running or retry behavior is addressed when relevant.
- Evals include process/trajectory, not only final text.
- One SLO or quantitative capacity estimate is stated.
- Rollout/rollback is addressed.
- Two tradeoffs are tied to requirements.