[amaybaum]

This is the guiding draft PR for NP-AVG-1 (#929), as §A.43 requires. It is planning only: no preregistration is frozen and no enumeration has run.

The base is #912's branch at 2af1d5d5, which contains v8 at fc452ea8, so this diff shows only the planning files.

Owner decision (2026-10-26)

The decision: proceed with NP-AVG-1 Stage 0 first, and keep PE-02 (#934) open as the next experimental candidate.

What Stage 0 is: a geometry-only investigation of what the masking structure preserves or removes. It is not a prediction experiment.

Approved scope:

• Fix the configuration space and masking rules before enumeration, including all symmetries and equivalence relations.
• Count survivors exactly, with reproducible integer arithmetic.
• Classify survivors by structural properties, including distinguishability under the permitted observer interface.
• Identify degeneracies and counterexamples rather than excluding inconvenient configurations.
• Compare against known combinatorial results.
• Retain the complete evidence: code, controls, counts and explicit limits of coverage.

Acceptance: counts that can be reproduced independently, and a precise characterization of what the masking geometry preserves or removes. An empty, small or unfavourable survivor set is a valid outcome.

Out of scope: Stage 0 claims no unpredictability, computational hardness or quantum behaviour on the strength of survivor counts.

Action	Status
Draft the Stage 0 preregistration	Approved, in progress
Run cheap design controls that bear on no result	Approved
Independent plan review	Required
Freeze Stage 0	After review, and any owner decisions it raises
Launch the full Stage 0 enumeration	Not authorized
Launch PE-02	Not authorized

Contents at 7cc21833

Under composition-access-20261024/native_predicate_avg_v9/planning/:

File	What it is
PLAN-r1.md	The first plan draft.
review/REVIEW.md	Its independent review: ACCEPT WITH CHANGES.
PLAN-r2.md, CHANGES-r2.md	The revision applying all 13 required changes, and the change map.
review/RECHECK-r2.md	The re-check of r2: BLOCK for freeze, on nonredundancy. It is the motivation for Stage 0.
review/toy/	Toy checks of the exact reductions L1–L3.
NOTES.md	The drafter's sources, and its interpretations of #929.

The Stage 0 draft will be added here once written, then reviewed independently.

---

Generated by Claude Code

[amaybaum]

Checkpoint at 1f027140: Stage 0 draft preregistration (WIP, unreviewed, not frozen).

The draft is in planning/stage0/STAGE0-PLAN.md, with STAGE0-NOTES.md beside it. It is a geometry-only preregistration of the masking structure, and it fixes:

• The configuration space: records of at most two of v8's stream items.
• The equivalences: identity merge, parity, unordered pairs, and the C4v symmetry group (order 8).
• The masking rules: asymmetric cone radii.
  • L3′ is a rule the drafter added. It also eliminates the far-slice axis extremes of each cell.
  • L3 is reported beside it for comparison.
• The survivor definitions and structural classification.
• The collapse propositions, with proof obligations PO1–PO9, and a MISMATCH rule for any disagreement between proof and computation.
• The decision labels: VOID, MISMATCH, INCOMPLETE, COLLAPSE_ALL, RESIDUE_UNLINKED, RESIDUE. No label is read as unpredictability, hardness or quantum behaviour.
• Controls C0–C14, and rule mutants.

Design controls run (owner-approved). These touch the configuration combinatorics and toy geometries only; nothing is computed on v8's class. They are in stage0/design/ with their outputs.

• Symmetry group orders.
• Orbit counts by direct canonicalization and by Burnside: 875 singles and 815,175 pairs.
• Cone radii and bijectivity of the axis tips on toy instances.
• A planted-tie detector.
• A timing estimate: about 6–12 CPU-hours per convention and rule before the orbit reduction.

Next:

• Independent plan review, which is in progress. It checks the new L3′ rule first.
• Owner decisions on open items S1–S10. S6 is a small exact law check inside a geometry-only stage, so it needs approval.

Not authorized: the enumeration.

---

Generated by Claude Code

[amaybaum]

Stage 0 independent review of the draft at 1f027140: ACCEPT WITH CHANGES.

The review found the masking rule L3′ sound against v8's rule. It checked this exactly on a 1D analogue, exhaustively over 2^21 worlds, and a deliberately wrong far-tip rule was caught. The required changes R1–R9 concern the controls:

• As written, the controls could not catch over-elimination, which biases toward "everything collapses".
• One mutant is equivalent to the rule and can never fail.
• The symmetry control would void a correct run.
• The design toys used the wrong parity.
• The "observer interface" class was nearly vacuous.

Owner decisions (2026-10-26):

Item	Decision	Condition
S6, exact law check	Approved	Additional assurance beside the three-case proof.
Search size	Global closure bound now	Larger searches deferred. The bound decides whether one is justified.
Symmetry reduction	Approved	With the corrected control, and a check that it preserves the relevant counts.
Literature	All five named works	Checked first-hand before any novelty claim.

All nine corrections are to be applied.

Safeguards required by the owner, using independent reference calculations where practical:

• Soundness: every elimination is justified by the frozen rule, and an incorrect one fails a control.
• Completeness: every eligible configuration is considered, and every required elimination is applied.
• Non-vacuity: planted survivors and non-survivors must separate genuine collapse from over-elimination.

The observer-interface classification keeps the reviewer's three-way distinction.

Sequence:

1. Apply R1–R9 and regenerate the design evidence.
2. Run S6, the corrected mutations, the planted controls and the small exact reference checks, on toy and small-window instances only.
3. Independent review.
4. Freeze, with exact code and configuration hashes.
5. Separate authorization to enumerate v8's actual configurations.

---

Generated by Claude Code

[amaybaum]

Owner direction on Stage 0 r2 (2026-10-26). r2 at baf408d9 goes to independent review unchanged. The replacement of the symmetric-radius countercontrol is accepted, provided the reviewer confirms the parity argument.

If the review is a clean accept, with no required changes:

1. Freeze the exact plan, code, configuration, controls and hashes.
2. Preserve the review and all design evidence.
3. Prepare a pre-execution certificate covering:
   • the precise v8 inputs;
   • the search bounds;
   • the resource limits;
   • the expected outputs;
   • the failure conditions.
4. Present the certificate for separate execution authorization.

Fail-closed requirement. The production run must fail closed if its configuration differs from the frozen certificate, or if any soundness, completeness or non-vacuity control fails.

Not authorized: the full v8 enumeration.

---

Generated by Claude Code

[amaybaum]

Stage 0 r2 independent review: ACCEPT WITH CHANGES. This is not a clean acceptance, so nothing is frozen.

What holds (nothing was run on v8's class):

• The rule, the masking engine, the three safeguards and the toy evidence reproduce byte for byte, including under -O.
• The reference code shares no code with production.
• Every non-equivalent mutant is caught by a named control.
• The parity argument is confirmed. For an even-class quantity, a cell on the far slice lies at a distance with the parity of d−1, so radius d and radius d−1 select the same cells (0 differences in 400). The symmetric-radius control could therefore never fail, and replacing it was correct.
• S6 is correct for the 21-bit carrier structure.
• The v8 guard refuses to run without its flag.

Required changes:

Item	Change
R1	The plan says it freezes "every file in stage0-code/", but the enumeration program its decision rule depends on does not exist yet. The reviewer recommends a two-layer freeze: freeze the plan, engine, references, controls, mutants and S6 now; review and freeze the enumeration program separately before any launch. Owner decision.
R2	Run the C1, C2 and C3 depth ranges stated in the plan; the code runs them at shallower depths.
R3	Check count preservation over every reported count.
R4	Small fixes.
R5	Retain all eight run outputs, deterministically.
R6	Correct the §A.41 rows.

R2–R6 are being applied in r3.

Fail-closed (information). The current code does not verify its own frozen hashes, and no production runner is gated on the controls. The reviewer lists eight requirements for the pre-execution certificate:

1. a manifest check at start-up;
2. pre-flight controls at the exact commit;
3. passing mutants for the same hashes;
4. in-run gates that lead to VOID or MISMATCH;
5. a digest- and coverage-checked merge;
6. v8 authorization bound to the exact commit;
7. -O determinism and a mandatory dry run;
8. labels produced only by the merge.

Not authorized: the full v8 enumeration.

---

Generated by Claude Code

[amaybaum]

CI on 5941bc58: "Lean kernel check" and "Mathlib bridge" failed in their Install elan step, before any build ran. The cause is that releases.lean-lang.org returned HTTP 504 while serving lean-4.35.0-linux.tar.zst. The same outage hit #938 and #939 at the same time.

This is not a failure of this PR. The commit adds only review records under verification/programmes/…, and no Lean code ran. I will re-run the failed jobs once when the run completes; the API refuses a re-run while it is still in progress.

---

Generated by Claude Code

[amaybaum]

Closing without merging (D-70: retire to #929; PR-AUDIT-1, #1088). The branch and its commits are kept.

Owner decision (amaybaum, Claude Code session chat, 2026-11-04):

Approved, with one administrative condition concerning O-REUSE. I accept the final independent confirmation and approve retiring both #935 and #912 without merging. Neither PR's scientific results are adopted by this decision.

D-70 — #935: Approve retirement to #929. Close without merging.

This is an administrative retirement, not a scientific adoption. #935 is a reviewed plan, not an executed v8 result. No native mechanism or OI–QM equivalence is established, and no scientific status changes.

• Accountable home: Issue NP-AVG-1: does the native acceptance predicate's negative depend on the pointwise background convention? #929, handoff comment. It carries the exact scope:
  • Stage 0 Layer 1, frozen at f33d8d63a14b3bbaa3c1eb1e52aeffcb55b0c966;
  • erratum E1, in effect at 53786e6e;
  • Layer 2 r5 at 6ceedd82 and its review at 32aa481a (ACCEPT WITH CHANGES), with the required and unverified items;
  • the feasibility finding: C7 set 2 is about 4× the ceiling, measured on an analogue;
  • every open owner decision and branch-only ruling, with the three meanings of "O1" separated;
  • the authorization limits.
• Provenance kept:
  • branch research/np-avg-1-plan-20261026 at 32aa481a94c008bb6c48eba93a4637669604508b, the only branch holding f33d8d63;
  • base branch draft/oi-composition-access-20261024 at 2af1d5d5, which holds v8 at fc452ea8;
  • branch research/oi-pseudoentropy-20261025, holding 47a1f19b, cited at NOTES.md:40–41 and PLAN-r2.md:578;
  • every plan, review, probe and design-evidence file under …/native_predicate_avg_v9/planning/, unchanged;
  • this PR's discussion.
• Why nothing is landed: nothing here is a result. Nothing was computed on v8's configuration class, and no claim on main cites these records.
• Resumption: a successor draft branches from 32aa481a, or from main only after Composition and access: finite-depth composition contract, logical acceptance circuit and complete-record laws (research record) #912 lands by ordinary merge. The frozen plan requires a base containing fc452ea8 (STAGE0-PLAN-r5.md:37–39). No rebase.
• Standing restrictions:
  • do not delete this branch;
  • never merge this PR into its base;
  • no v8 run, count, calibration dispatch, enumeration or PE-05 (PE-05: separate active from passive observers and charged from uncharged lookup tables (OI-PE-1 follow-up) #934) is authorized.

---

Generated by Claude Code