Constrain the question
Give the planning process a defined objective, relevant evidence and protected inputs. Permission to propose does not grant permission to execute.
Express the objective as a bounded question with supported parameters. Record what the planner may change, which inputs are protected and how many trials or requests are permitted. Giving a model more context does not give it more authority over an environment.
- Set trial and request bounds.
- Keep acceptance assertions outside model control.
- Record the provider and proposal provenance.
Challenge the alternative
Review proposed values and assumptions before running bounded trials. Preserve the original and edited versions so a reviewer can understand the change.
Inspect the proposal for unsupported fields, uncited assumptions and changes outside the declared objective. Keep the original response and any reviewed edits distinct. The evidence should explain which candidate was actually evaluated, not just which explanation sounded convincing.
Proposal, independent verification and human review
Constrain the question, relevant evidence and request budget.
- Inspect unsupported or uncited claims.
- Keep baseline and targets fixed.
- Evaluate training and independent held-out evidence.
Keep abstention visible
A malformed response, provider outage or failed verification is not a recommendation. Withhold alternatives that do not satisfy the declared gate.
Abstention is a useful result when the provider is unavailable, the response is malformed or the candidate fails verification. Do not substitute a new successful response for a consumed failed outcome and then report only the success. Keep the full decision history visible.
Should an AI-proposed scheduling change be considered further?
Limit the planner to supported parameters and a declared trial budget. Freeze the baseline and assertions, then retain the candidate values, trial results and independent verification outcome.
A passing numerical result goes to domain review. A failed or unsupported proposal remains withheld with its evidence, rather than being presented as an accepted recommendation.
- Retain failed and withheld outcomes.
- Require independent review where the operation demands it.
- Separate simulation evidence from real-world authorization.
Keep the test outside the proposal
A candidate should not be allowed to redefine the assertion that judges it. Freeze the baseline, independent expected outcomes and permitted parameter bounds before qualification. Separate development trials from protected evaluation.
A plausible explanation can help a reviewer understand intent, but it is not a substitute for executable checks. Evaluate the actual candidate values and retained results. If evidence is incomplete, report the gap rather than upgrading the proposal into a recommendation.
Keep outcome states meaningful
| State | Meaning | Review response |
|---|---|---|
| Malformed or unsupported | The proposal cannot be evaluated within the contract. | Retain the response and withhold execution. |
| Verification failed | The candidate did not meet the declared gate. | Inspect the failure without rewriting its history. |
| Checks passed | The candidate met specified numerical conditions. | Review usefulness and any separate action authority. |
- Protect acceptance assertions from candidate control.
- Pin model and evidence versions.
- Keep development and qualification outcomes distinct.
Review usefulness independently
Numerical success and practical usefulness answer different questions. A domain reviewer should judge whether the change addresses the intended problem and whether its tradeoffs are acceptable under the stated assumptions.
External action requires its own authority. The surrounding platform uses independently approved identities and supported action profiles; the AI cannot approve its own plan or turn a simulation result into customer production permission.
Keep the evidence and its conditions together. A later success does not erase an earlier failure or broaden the original scope.
- Inspect the actual outcome and its assumptions.
- Retain the reviewer decision separately from numerical checks.
- Bind any later action to the required independent approval.
Prepare the proposal review
Bring the objective, original proposal, candidate values, fixed baseline and complete outcome set to the reviewer. Include malformed, failed and withheld cases where they affect the conclusion.
The checklist records review preparation, not approval. A bounded internal experiment can inform the next decision while leaving customer integration, domain acceptance and production action authority unresolved.
Review preparation
0 / 6 prepared- Name the reviewer and permitted decision.
- Retain provenance and failures.
- Record remaining qualification and action boundaries.