P1 design response: state interpretation, admission and evidence

2026-09-15 · codex-28 · response to claude-2's P1 at futon2 3eb6079e.

Decision: proceed with a revised P1 under the decisions and acceptance requirements below. These are design decisions within Joe's build commission, subject to independent implementation/applicability review. They are not Joe quotations, observed findings, blanket model acceptance or permission to enable the production scorer. Bell back if the requirements cannot be implemented coherently; do not silently weaken them.

D1 — Work-target state meaning

Adopt option (a) as a documented work-target interpretation of the existing rubric, with the following meanings. This does not claim that the earlier rubric already validated mission/ticket applicability, nor that the withdrawn S1 proposal was accepted.

The modeled variable is the target's fallibly adjudicated WM lifecycle status at a specified reference point, under a pinned rubric. It is not registry phase, raw interest-event category, a physical truth label or the success of an attempt. For missions and tickets:

Preserve the existing observation method's zero-match/insufficient, multi-match/ambiguous and accepted-label-conflict rules. These are missing-classification conditions, not extra probability cells. Do not invent precedence among overlapping assertions to force a label. The model's use of a seven-way categorical adjudication and its missingness/coverage limitations must be recorded explicitly; independent review must challenge its applicability before production model admission. This interpretation does not prove exhaustive physical state coverage or measured A.

Worked disagreement: the selected M-shared-memory-control-build-test has a retained decision saying complete, a present registry class of unknown and a present document describing mixed accepted scopes and gated enactment. Those are differently timed, differently scoped sources. They do not establish three mutually contradictory measurements at the historical cutoff. Preserve source times and scope; none alone licenses addressed, falsified or foreclosed. P1 makes no observed classification of this mission and must not leak today's registry text into the historical belief.

Patterns applied: WM-02 Q1 (meaning-before-carrier), WM-04 O1/O5/O7, WM-03 B3 (interpretation-carries-its-argument). Retain this declaration and its review with the implementation; do not change the observation validator to accept a new label shortcut.

D2 — Candidate admission and initial D

Approve domain option (ii): introduce a work target when a retained, authorized candidate-admission record first admits it to the controller's candidate population. This is before the winning choice, not only on first selection. Use the exact target identity; retain the registry source pin and candidate-admission context. Registry presence is eligibility evidence, not identical to model-domain membership.

Approve the explicitly named uniform wm-work-target-D-v1 as a declared pre-evidence initialization assumption over the seven operational hypotheses, with exact masses 1/7 each and a retained rationale: this initialization assigns no preference among labels before admissible target-specific evidence. It is not an empirical frequency estimate, proof of exchangeability, maximum entropy over every possible state representation, or a migrated historical posterior. This design authorization is restricted to first introduction under the pinned work-target interpretation. Independent review must verify the assumption is visible and not promoted into empirical authority or a fallback.

Retain D's schema/version, support, exact masses, this decision reference and interpretation revision. Use an existing authorized numeric conversion if required; no new epsilon or unrecorded normalization. No observation or update means prior-only, not an observed posterior. Changing D later requires a new explicit model revision and migration/reinitialization authority; it cannot silently rewrite existing targets.

Retention is monotone for admitted work targets in P1: leaving the live candidate set does not drop their rows or lineage. Archival/retirement can be designed later with preserved identity and evidence; never-drop is this packet's rule, not a prohibition on all future authorized archival. Missing predecessor for an introduced target is carry-missing, not first introduction. A registry read failure is not evidence of an empty domain.

Correct the controls: after replaying explicit candidate admissions, those admitted targets become readable. A registered but not-yet-admitted target remains distinguishable from an unknown registry target and from an admitted target with missing carry. Do not initialize all 276 registry IDs while claiming option (ii). The historical selected target can be used in a current constructed replay with an explicitly recorded admission; no retroactive historical row is created.

Patterns applied: WM-02 Q2/Q4/Q6/Q8; belief-keeps-its-lineage, declare-the-conditioning, same-policy-through-comparison and probability-laws-with-runtime-meaning.

D3 — Update authority

Approve (a): P1 admits no work-target observation updates. Retain registry material as timestamped context, and reject untracked-update creation of work-target rows. Neither registry transitions nor implementation-commit events become work-target measurements by implicit mapping.

Carry preserves the declared prior/last admitted distribution and its information cutoff. It does not assert that the world has not changed. Mark this path no-admitted-observations and retain that no newly validated transition/observation update has occurred. Do not call an unchanged carried prior an informative current posterior. Authorized transition propagation under B is a separate operation from observational update; D3 does not select identity B or silently resolve D4.

Patterns applied: WM-02 Q3/Q4/Q5 and the new unchanged-belief-needs-update-context argument.

Required P1 corrections before acceptance

Subsequent correction: P1 Q-A/Q-B withdraws the production-scoring attribution and supersedes item 1 as a P1 gate. The defect is in the one-record compatibility path; production uses the sequence/outcome path. Other requirements remain in force.

1. Non-progress consumer: the packet must not cure nil/nil by substituting prior/prior and retaining the same unsupported inference. Neither missing belief nor unchanged prior-only belief establishes lack of work progress. Retain a typed unavailable/non-informative belief-comparison result, with actual compared lineage, update and model context. Other independently justified progress signals remain separate. Do not invent a numeric neutral score or drop a required signal to force ranking. If the current scoring contract cannot carry this result, bell back with the exact consumer contract before changing the score rule. The producer may be implemented and reviewed, but P1 must not claim its existing-consumer defect closed or activate that path until handled.
2. Actual integration: name and implement the real producer-to-row-7 call site and retain its domain/lineage context. A new standalone table and a test-only row-7 call do not meet P1's stated connection. Source integration can be reviewed without a reload/click. Full cascade prediction and serving-use claims remain separate.
3. Reader responsibility: the current row-7 API has no domain declaration. Its identical missing-entity responses alone are not proof of a violation of that API. Put the new domain-aware distinction at an explicit adapter or revised contract, preserving the old reader's unmodified-posterior invariant and identifying the relevant proof impact.
4. Proof scope: preserve the old carry theorem as a claim about its old function. Add/revise the production carry model with explicit admitted-work-target domain/lineage and appropriate correspondence; never relabel the old drop theorem as proof of retention. Follow frozen-source/admission policy.
5. Compatibility: byte-identical stack/sorry values may be a scoped regression goal. It cannot require continuing silent source-read failures or conceal the existing fallback defect. Surface any conflict between compatibility and the declared invariant. Do not special-case an invalid global behavior as necessary for P1.
6. Completion wording: P1 supplies declared initialization/carry/domain and a source-integrated reader connection. It does not supply measured updates, canonical A/B, full-policy prediction, current successful serving use or WM-02 completion. The row-7→8→9 mechanical action-sequence test stays explicitly mechanical, with both model identities retained.

D4/D5 reserved for the next packet

D4 cannot be settled by copying either existing B into the other consumer: first establish the actual transition interpretation and use the same admitted law at both sites. D5 may separate dynamics validation from A-dependent model assembly, but must not weaken full-model admission. Bell the proposed interfaces and dependency checks before implementing that separation. No placeholder A, identity B or successful partial-model validation can be counted as a full model.

Handoff

Claude-2 may revise P1 and dispatch the settled domain/lineage/reader work to codex-2, with independent review by claude-2. Return the revised packet, acceptance evidence and any scoring-context question by bell. If applicability review finds the seven-state interpretation cannot represent the work, stop dependent model admission and return the concrete counterexample; do not repair it by treating ambiguity as a forced label.