[GamingPuzzled]

Conditional κ .00311313795 (tested .003113137955; exact raw crossing in (0.00311313797336, 0.00311313797337)), +1.219e-7 over #5751 (Ali Romman, .00311301600). Package: research/p104-block-reframe-r22/, stacked on #5751 head 8483f9a67 (#5713 → #5718 → #5751). This is the merged #5739 block move, ported.

Conditional on exactly #5751's assumptions and its P104 line, including free helper endpoints (#5561, audited in #5567). The word changes one helper start and one helper final.

Change

This is a large-neighbourhood step of the R22 exact reorder+re-frame MIP. All operations outside one ~1,000-op window are pinned. Inside the window, every chart within dim gap 4 of an op's bounds is allowed. The whole-word gap-4 model does not solve in 2h. The windowed model solves in about 100s.

• Block move: one 96-op block on registers 41, 42, 106, 400, 401, 403, 405, 410 and 741 leaves chart 477 (dim 8). 88 of its ops now run on a new chart span[265,194,164,75] (dim 4), and 8 run on chart 127.
• Helper endpoints: helper 741 now starts on the new chart, and helper 405 now ends there.
• Unchanged: no instruction, coefficient, order, owned lifetime, main endpoint or port changes.

Per local call: paid H1 +1, H4 +1, H5 −1 (66834 → 66852 paid calls). Helpers H1 −1, H4 +1, H5 +1, H8 −1. The gain comes from the helper credits. Surplus 760 and fields 12958 are unchanged.

Verification

python3 -W error -B research/p104-block-reframe-r22/verify.py passes and reproduces RESULT.json exactly. It checks:

• the source pin, and regeneration from EDITS.json (unique, disjoint event-run hunks, one appended chart, two helper endpoints);
• trace equivalence of the new order;
• Conditional κ 0.00311301600 (+2.77030e-6 over #5718): paid intermediate meetings #5751's own stack (its pinned local.py, four-level source chain): replay, transpose columns, 2927 native maps, physical caller, and Conditional κ 0.00311301600 (+2.77030e-6 over #5718): paid intermediate meetings #5751's pricing.py with only 4 literals changed;
• six controls, all rejected.

Search coverage (on the #5713 word, before porting; no further gain)

• Gap-4 windows of 1,000 ops: 32 windows, all optimal, gain 0.
• Gap-4 register neighbourhoods (every op on a group of 8 helpers, whole word): 67 groups, all optimal, gain 0.
• Gap 5 does not solve in the 12000–13100 op region, even on 200-op windows.

AI disclosure: Claude Opus 5.5 (Anthropic) via Claude Code, directed by Joel Pulikkan (GamingPuzzled). Apache-2.0.

🤖 Generated with Claude Code