Evidence-based rating and iteration

A design score is useful only when each point maps to observable evidence and a named approver. The producing agent may run a first critique, but it cannot certify its own work. Combine deterministic checks, a fresh review pass, and human product or design judgment; record defects and dispositions instead of repeatedly prompting until the model declares 10/10.

Four dimensions

Intent fidelity

Does the artifact satisfy the accepted outcome and non-goals? Cite each criterion.

State completeness

Are required states, breakpoints, content, keyboard paths, and failures present?

System coherence

Does it use canonical tokens, components, interaction patterns, and terminology?

Evidence quality

Can another reviewer reproduce the state, route, viewport, check, and finding?

Score each dimension from 0 to 10 and use the minimum as the handoff score. A 10 is not “looks excellent”; it means every declared criterion has reproducible proof and no open blocker. For normal work, set an explicit ship threshold such as 9/10 with no high-severity finding and required human acceptance.

Review prompt

Review ARTIFACT against ACCEPTANCE_CRITERIA as an independent reviewer.
For each criterion return pass, fail, or not-tested with evidence.
Score intent fidelity, state completeness, system coherence, and evidence
quality from 0-10. The overall score is the minimum.
List blockers, non-blocking gaps, and the owner of every manual decision.
Do not edit the artifact during review.

After fixes, rerun the same checklist in a fresh context. Preserve previous findings and mark them fixed, accepted, duplicate, or still open.

Stop conditions

Stop and hand control to a human when:

a product, brand, legal, privacy, or accessibility decision is missing;
two iterations do not improve the same failing criterion;
a tool cannot reproduce the artifact or collect required evidence;
resolving the gap changes accepted intent or architecture;
the threshold passes and remaining gaps have explicit dispositions.

Failure modes

The score rises but the artifact does not. Require evidence per criterion and compare the actual version, not the agent’s summary.

Everything receives 10/10. Tighten criteria, add negative states, and sample ratings against human review.

Review edits the artifact in place. Split read-only review from a separately authorized fix task.

Taste dominates functional quality. Block on intent, states, accessibility, and system constraints before polish.

Verify the rating loop

The rubric points to accepted criteria and artifact versions.
Every failing score cites reproducible evidence.
A fresh review validates fixes.
Human decisions have owners and dispositions.
The loop has a threshold and explicit stop conditions.
The accepted score and evidence travel to the next SDLC stage.

Return to the pipeline

Apply the rubric at each handoff in the four-artifact pipeline.