[thu-nmrc]

This changes two coupled frame itineraries in Ali Romman's complete P104 candidate #5545, raising its conditional κ from 0.00304117 to 0.00304153. Six ADDs and fifteen commuting SCALE chart annotations move; every scalar coefficient and its order remain identical.

The new local paid delta is +2H1 −3H2 +H3 −H4 +H5, repeated at all eighteen supplier invocations. Strict concavity proves a gain for every exponent in (0,1). All 735 arbitrary public inputs, all 527 productive helpers, all 743 endpoint rows and the eight owned lifetimes remain. The complete invoice still has 69,912 paid calls / 12,814 productive fields, rank surplus 760; no helper, query, restoration or ordinary work is omitted.

This is a different paid-path change from the existing single H1+H5→H6 contact and its narrow repetition ceiling. No new general optimizer or global optimum is claimed. PROOF.md gives the literal sites, all-positive-width typed implication, full price and limits; SEARCH.md describes discovery scope.

The exact raw crossing is in (.00304153116714,.00304153116715), tested at .0030415305. Cutoff 17 / floor 18, both inherited rational engines, completion/compiler/BIT stock and all 58 outer predicates pass. A separately written third rational log/exp engine independently proves the decisive signs and recounts the full invoice. The source native phase lowering and complete 265-axis caller are freshly replayed. The independent scalar/funding peer code, retained native compiler/interpreter, negative controls and full receipts are included.

The five inherited framework premises remain: native all-column implementation, owned erasure, uniform compiler, precision/tape costs and analytic recovery. The huge uniform word and rounded multiplier are not materialized. This is author-checked conditional research, without independent human/proof-kernel acceptance or selected promotion. The selected record is unchanged.

The package is isolated on canonical main 32890bc. Its authenticated 33-file official PR5545 source closure is retained as a dependency capsule; frozen predecessors and shared code are unchanged. No pending Git object or network is needed at runtime. Source attribution is preserved.

Reproduce with Python 3.12/stdlib:

python3 -B scripts/research_package.py --package research/p104-joint-contact
python3 -B scripts/check_research_commit.py --package research/p104-joint-contact --check-source-history

Expected markers: PASS_P104_JOINT_CONTACT_COMPLETE_KAPPA_00304153 and PASS_PUBLIC_P104_INDEPENDENT_FUNDING. VALIDATION.md records executed normal/optimized, source-history, committed-byte and Git-free cold replays. Passing CI is separate from scientific adoption.

Credit Ali Romman (#5544/#5545), Gareth Davidson (#2305), David Arias (#2152), Cinna Davis (#2219/controller), Jacob Alan Sussman, zxv, Cao Yifan and all common-chart, dual-ADD, geometry, native, COPY, precision and analytic predecessors in the retained proofs. Prepared for thu-nmrc with substantial OpenAI Codex assistance, including parallel subagents; Apache-2.0.

Open next question: independently review the actual six contact sites and complete conditional cost join, preserving the inherited premises. Further frontier improvement needs a different paid path, target contract or caller; refining this crossing alone is not another structural gain.

Hosted CI at exact head 79072da passed both jobs: run 40194668767, plan 3m55s and complete registered replay 8m18s, including unchanged committed-input checks. This does not claim maintainer scientific adoption.

A verification-only follow-up at head 1123635 adds an independent dense audit of all four newly introduced MOVE maps. All 2,048 columns and 1,048,576 complex entries are compared, including absolute phase, exact norm and measured rank. It reuses Chris R/Clear1nsight's credited, unchanged PR5548 arithmetic interpreter; no old native foundation is claimed closed. A wrapper binds the audit to this exact source and word and rejects two actual source corruptions before imports. Normal/optimized and Git-free read-only empty-PATH cold replays passed, with all 69 export files unchanged. WORD, RESULT, pricing and the earlier scientific core remain byte-identical. Two extra 30-second commands bring the six-command timeout total to 1,200 seconds.

Hosted CI for that exact current head also passed both jobs: run 40198198159, plan 3m28s and all six registered replay commands in 10m1s, including unchanged committed-input checks. It completed on 2026-10-25 at 13:47:39 UTC. This verifies reproducibility at the current head; scientific adoption remains unclaimed.

[Matarener]

An independent native-map inventory (all 3,188 operators, dense all-column), caller and funding audit of this PR is in #5572: no new κ, CI green. All layers pass and the raw crossing equals this PR's interval.

---

Generated by Claude Code