0064: Review plan shape after the first planner request
Status: superseded by ADR 0066 (2026-08-09) · Scope: product tier (core/governor planning)
Context
Section titled “Context”ADR 0061 added two source-neutral instructions to governor planner requests. They asked for independently observable cross-layer slices and explicit contracts, module boundaries, types, and call paths. Independent reviewers and revision requests received matching checks.
A controlled comparison found a useful quality signal but missed its fixed planner-cost ceiling. The combined guidance won all three cross-layer pairs and moved the first credible integrated observation earlier, but median planner model duration increased 27.45% against a 25% ceiling.
A second experiment separated the planner concepts across 72 sequential calls. All proposals were valid and local plans stayed at one work item. The results were:
| Planner variant | Visible-token change | Model-duration change |
|---|---|---|
| Observable-slice instruction only | -3.35% | -28.31% |
| Program-shape instruction only | +17.58% | +108.73% |
| Shorter combined instruction | +8.29% | +63.08% |
The slice-only variant passed the cost screen and moved the cross-layer median first observation from item 3 to item 1. It did not pass blinded quality review. The two primary judges agreed on 55.6% of all pairs, below the fixed 70% floor, and disagreed on every cross-layer pair. The blinded adjudicator preferred the baseline in all three because the slice-only plans bundled persistence, surfaces, and proof into one item instead of exposing executable program seams and justified prerequisites.
Alternatives considered
Section titled “Alternatives considered”- Keep the original combined planner guidance. Rejected because it failed the fixed planner-duration gate. Changing the ceiling after observing the result would make the evaluation meaningless.
- Keep only the observable-slice instruction. Rejected because an earlier observation did not make the cross-layer plans better overall. The surviving candidate failed both the preference and judge-agreement gates.
- Use the shorter combined instruction. Rejected because word count was not the main cost. That variant was 63.08% slower even though visible-token growth stayed below 10%.
- Add a required design packet or another governor stage. Rejected for this change. The experiment evaluated prompt guidance, and a new durable artifact needs its own evidence and authority design.
- Remove the reviewer and revision checks too. Rejected for now. The prior reviewer comparison met its defect-recall, false-positive, local-ceremony, visible-token, and duration gates. Those surfaces require a separate study.
Decision
Section titled “Decision”Remove the two general delivery-strategy instructions from the first governor planner request. Keep the independent reviewer checks for implicit program-shape decisions and unjustified delayed integration. Keep the matching revision instruction so a rejected candidate can address those findings.
Planner profiles supplied by a host may still contain domain-specific planning instructions. Harnery’s common planner contract stays focused on the mission, durable state, dependency semantics, frozen template authority, limits, and the model-authored proposal schema.
Result and gotchas
Section titled “Result and gotchas”- A governor without configured plan review no longer receives a common observable-slice or program-shape quality rule. Harnery does not claim that graph validity proves semantic plan quality.
- A reviewed plan can still be sent back for implicit consequential choices or delayed integration. The correction happens after an inspectable candidate exists instead of making every first planner request reason through both strategies.
- The proposal schema, inferred-root contract, persisted plan records, and approval authority do not change.
- The reviewer and revision wording remains provisional until evaluated as a separate surface. This decision does not authorize another workflow stage or design schema.