[aliromman]

This construction raises complete conditional kappa to 0.00310815849, a gain of 0.00000139516 (0.04491%) over PR #5660. Four temporary public-register bases remove internal contacts after every entry and exit transport is charged. This is a small gain; no large improvement or global optimum is claimed.

Proof, eight scalar block certificates, complete receipt, source pins, and validation record.

For each interval, insert R, replace each scalar block M by R M R^-1, and remove R at the exit. Here R = I + c E_ab. The exact block certificates prove the interval identity for arbitrary inputs. Both internal meeting charts are part of the literal paid word. All public boundary charts, productive targets, and owned instructions and transports remain fixed.

The four changes use pairs 438/733, 119/125, 125/128, and 128/139. Their total local paid delta is -7 H1 + 2 H2 + H3. Across 18 full calls, the complete bill has 66,402 paid calls, 12,958 productive targets, and rank surplus 760. All 743 independent inputs, 535 helper targets, 11 owned lifetimes, and peak two owned banks are retained. The complete ordinary-work bound is 184,647,552, below its 1,000,000,000 budget.

Two rational interval engines put the raw crossing in (0.00310815850980, 0.00310815850981). The tested saving is 0.003108158495. This fixed invoice has less than 2e-11 headroom above the claimed saving before its raw crossing. A larger gain needs a changed word or caller.

The package retains the complete authenticated PR5660 export at 3b0b4f37d5005f381ed9cfcccebfbe1ab35c71aa, including its nested predecessor export and notices. The branch starts from canonical main df0137efd35e11b4445734bf612bdd4c7ebc330c, priority revision 7. It changes only this new package. The selected frontier is unchanged.

python3.12 -B scripts/research_package.py --package research/p104-paid-interior-bases

Expected: PASS_P104_PAID_INTERIOR_BASES_KAPPA_00310815849 and PASS_REGISTERED_RESEARCH_PACKAGE. Python 3.12 and the standard library suffice. Normal and optimized registered replays pass. They check all construction steps, 565,504 transpose entries, 2,918 signed native maps, the complete physical caller, both moment engines, all compiler costs and all 58 outer conditions. Wrong basis signs, wrong scalar blocks, missing contacts or erasures, unpaid requests, extra targets, and inherited phase/caller/price corruptions are rejected.

Committed-byte and official export checks pass. A read-only Git-free optimized replay of the scientific commit passes in 89.70 seconds with empty PATH and unchanged hashes and modes. Twelve separate corruption tests on the two source exports reject before retained research imports. The publication commit changes only the validation record from that scientific commit, and its 16 package files pass the committed-byte check. Full-history ancestry mode was not run because the checkout is shallow. Hosted CI run 40271690035 failed in two unchanged shared Current CI navigation checks (review and source_manifest), before package replay. The hosted replay was skipped. This matches the reported main regression. Independent scientific review remains open.

The five inherited framework assumptions and helper endpoint contract remain. The huge uniform word and complete rounded multiplier are not materialized. These are author checks, not independent scientific acceptance.

Credit PR #5660 for the complete input and four-gate rewrite; @zxv's PR #5643 and PR #5650 for the A2 change and released composition; PR #5644 for the helper relations; and PR #5627 for the retained helper and full caller source. The full source notice retains the P104, scalar, native, controller, productive recursion, interval and compiler authors. The general basis identity is not claimed as new. New construction and verification by Ali Romman (@aliromman), with substantial OpenAI Codex assistance; no subagents. Apache-2.0.

Canonical coordination. The paginated 04:41-04:42 UTC checkpoint covered 361 open PRs and 4,986 Discussion records. Native-frame, retired-input and changed-caller scopes were checked and kept separate. The newer PR #5666 is related work, not a consumed input or additive claim; this value also exceeds its published 0.00310680117.