0066: Keep common plan review and revision generic
Status: accepted (2026-08-09) · Scope: product tier (core/governor planning)
Context
Section titled “Context”ADR 0064 removed general program-shape and delivery-strategy advice from the first planner request but provisionally kept two independent-review criteria and one matching revision instruction. The review criteria asked reviewers to reject implicit consequential choices and unjustified delayed integration. The revision instruction asked planners to expose program shape and prefer observable end-to-end slices.
A held-out comparison evaluated these surfaces separately. It used twelve new review candidates, two repeats per variant, six fixed revision cases, and two repeats per variant. All 72 reviewer and revision calls succeeded.
The review criteria improved relevant seeded-defect recall from 50.0% to 91.7% with 10.4% more visible tokens and 10.4% more model duration. They also rejected four of twelve valid controls, including both local reversible controls. That 33.3% false-positive rate exceeded the fixed 20% ceiling and failed the zero-local-false-positive rule.
The revision instruction produced valid plans in every run and used a median of three work items versus 3.5 without the instruction. Two blinded primary judges agreed on 91.7% of twelve pairs. After one blind adjudication, the instruction was preferred in eleven pairs and introduced no material requirements outside the mission or finding. Its median model duration was 42.7% higher, above the fixed 25% ceiling.
Alternatives considered
Section titled “Alternatives considered”- Keep both surfaces because they improve semantic quality. Rejected because common policy must meet its reliability and cost gates, not just its quality gates.
- Keep review criteria but exempt local work. Rejected because that is new wording that was not tested. A post-result rewrite would need a new study.
- Keep revision guidance despite the duration miss. Rejected because the ceiling was fixed before calls began. The strong preference result does not authorize moving it after the fact.
- Turn the criteria into deterministic validation. Rejected because Harnery cannot infer semantic system layers, maintainability, or meaningful observability from plan text and dependency edges.
- Forbid hosts from supplying these criteria. Rejected because a domain reviewer may know contracts and risks that the common runtime cannot know.
Decision
Section titled “Decision”Remove the two program-shape and delayed-integration criteria from the common governor review prompt. Remove the matching delivery-strategy instruction from the common revision prompt.
Keep common review focused on candidate completeness, scope, frozen goal context, blocking findings, human-attention boundaries, and the proposal-root contract. Keep revision focused on resolving the actual reviewer findings within frozen goal and template constraints. Reviewer specialist profiles may add domain-specific criteria without changing Harnery’s common contract.
Result and gotchas
Section titled “Result and gotchas”- No common governor prompt now carries the evaluated program-shape or observable-slice wording.
- The immutable proposal, review receipt, revision rounds, reviewer authority, explicit approval, and inferred-root contracts do not change.
- Removing universal wording does not claim that program shape or integrated feedback is unimportant. It means Harnery lacks a common formulation that is reliable and efficient across both consequential and local work.
- This decision closes the provisional prompt adoption. It does not authorize another workflow stage, design schema, or deterministic semantic checker.