Retained native snapshots: no saving in the tested clean-word cache models

[aliromman]

FINDING / RELEASED — no paid saving in the tested clean-word cache models.

The scoped cache investigation is complete. It did not produce a cheaper native word or a kappa improvement. The broader search for a breakthrough remains unfinished. This reservation is released; it does not claim the center-free, coupled-helper or changed-contact investigations.

Source: canonical main 7c472563474e60185e6ca101a17622133d047e56, actual native payload SHA-256 cfcd36b40326d7a5815e1db5313444379b3fb756b27d5201c561574c085cc0dd. The prior checkpoint records source and overlap coverage. All source logical endpoint checks retained the 632 independent input columns in the full-row experiments. No selected file or predecessor package was changed.

Local discovery results:

• Repeated proportional-row snapshots at equal frames: 1,781 groups backward and 1,697 forward. Full path reconstruction leaves the paid profile unchanged.
• Same-frame rational spans of earlier observations: no replacement among 3,290 ordinary climbs in either orientation.
• All-term native anchor normalization: no candidate in either orientation. This tests the enumerated normalizations, not every Gaussian-free circuit.
• Monomial leading-input elimination: no candidate. Adding invertible two-term operator polynomials of order at most 12 still gives no candidate, after 1,461 backward and 1,477 forward uses of such inverses. Multi-term pivots remain outside this test.
• An optimistic fresh-data calculation with all helper inputs set to zero still pays positive Gaussian rank on every helper. Separating all arbitrary helper inputs into autonomous jobs adds fresh-component rank mass 2,993 backward or 2,680 forward. This rejects that simple separation budget; setting those inputs to zero is only a diagnostic projection, not a legal replacement of the original inputs.

These are exploratory finite screens. The code and receipts remain local. They are not an independently reviewed obstruction, a proof of optimality, or a complete native implementation. In particular, the frame-support diagnostic does not prove a universal invariant theorem, and none of these results excludes new paid queries, new scalar contact graphs, or a different producer/consumer contract. No cold supplier, compiler, precision, outer or full kappa replay was run because no changed word survived discovery.

Credit zxv's PR #5283 for the snapshot contract and Cao Yifan's PR #4610 native lineage. The next construction needs a changed mechanism and a fresh complete budget; another search of the same free-cache class is not justified by these results.

Ali Romman with substantial OpenAI Codex assistance.

[aliromman]

FINDING / CONTINUATION — retained snapshots in the original clean native word.

The source remains canonical main 7c472563474e60185e6ca101a17622133d047e56 and the actual word payload SHA-256 cfcd36b40326d7a5815e1db5313444379b3fb756b27d5201c561574c085cc0dd. A fresh fetch leaves this base unchanged. The 19:31–19:32 UTC overlap refresh covered 315 open PR bodies, 4,643 Discussion records, and 7,267 local README/TASK notes. This coverage does not include arbitrary diffs or unpublished work.

Local discovery has found no cheaper word. Equal-frame proportional-row caches give 1,781 repeated groups backward and 1,697 forward; reconstructing their full role paths leaves every paid profile unchanged. A chronological exact rational-span screen of the original observations also finds no replacement for any of the 3,290 ordinary climbs in either orientation. These checks retain all 632 independent inputs, all helper returns, and all eight center jobs. They check the complete original logical endpoints. They do not establish an impossibility theorem for other cache protocols.

The current experiment changes the reconstruction class. It permits a separate native monomial map for each retained observation, followed by scalar combinations. It also tests exact small polynomial inverses of native maps when the polynomial inverse is Gaussian-free. A positive result must still produce an actual cache word, exact physical phases, ownership and storage bounds, and the complete paid profile before a supplier or kappa claim is made. An input/frame support check leaves 941 backward and 717 forward climbs without a fresh-support witness; it does not close these cases. No general support-invariant theorem is claimed from this diagnostic.

This remains the original clean-word cache question from zxv's PR #5283, on Cao Yifan's PR #4610 lineage. It does not take the center-free caller, changed-contact synthesis, or coupled original-helper scopes. The claim remains active through 2026-10-23T21:12:00Z. These are local discovery results, not independent acceptance. The selected frontier is unchanged.

Ali Romman with substantial OpenAI Codex assistance.