Repository navigation
Realise the P104 omitted completion at the actual entrance - #3609
Clear1nsight wants to merge 1 commit into
Conversation
Use existing dirty output pivots for a restored-source CA/2 accumulation, certify its actual native frames and phase, and retain the conditional consumer-budget stop. Preserve all predecessor sources and selected result.
Publication verification and reproducible source exportThe published head is Download p104-inplace-entrance-source.zip. The ZIP contains the 31 declared package/input files, including predecessor licences and notices, plus the export manifest. It is 242,074 bytes, with SHA-256 Extract the ZIP into a fresh directory and run from its root, with Python 3.12 and the standard library: python3.12 -I -B research/p104-inplace-entrance/verify.py
python3.12 -I -O -B research/p104-inplace-entrance/verify.pyBoth modes must print: The 124,335-byte Observed checks
Hosted run 39112397365, attempt 1, completed successfully at 2026-10-18 08:44:37 UTC. Both the planning and package replay jobs succeeded. The replay job used Python 3.12.15; its normal and optimised commands printed the receipt above, followed by The hosted checkout was the synthetic PR merge The package metadata records the historical state at its freeze. This comment supplies the subsequently completed publication gates without rewriting the frozen receipt. A green registered replay is executable evidence; independent human acceptance, proof-assistant verification, and a complete physical supplier bill remain unclaimed. The scientific decision remains NARROW: this is an exact local entrance construction with a reproducible partial bill. Its newly constructed profile fails the stated five-forward/four-dual consumer relaxation, so dependent supplier construction under that unchanged model is stopped. The three reopening routes in |
1 participant
Problem and result
PR3462 supplies the omitted P104 residual$R=CA/2$ , and PR3469 supplies endpoint adapters. A chronological implementation still had to acquire the Y readouts in a common frame, preserve dirty output contributions, return temporarily modified sources, and pay the intervening native maps.
This PR implements that completion at the actual original WORD entrance, where all Y fields have the empty frame. Eight required dirty X outputs serve as accumulators; the completion needs no additional Gaussian query bank. It delivers all 93 omitted banks in their required full frames and restores every Y before the unchanged 4,027-event PR3446 projection starts. The combined program has the original word's complete endpoint map.
The new omitted-section implementation uses 145 native calls:$52H_1+93H_9$ , replacing 473 calls while preserving rank mass 889. The resulting full word has 1,742 native calls and rank mass 3,542. Its local native tariff improves, but its newly substituted profile still fails the named five-forward/four-dual consumer's favourable cost screen.
Implementation and admitted inputs
The eight pivot rows of C are the identity. After moving all 52 omitted X inputs from their actual singleton frames to the empty frame, the scalar schedule performs
All preconditioning precedes the basis update, and the final scatter uses the updated basis. Thus arbitrary initial dirt is preserved. Each of 103 active Y fields is halved once, used for its A column through signed unit ADDs, and immediately doubled back. Y1 is unused by the prefix but remains required by the projection. The 52 X outputs and 41 omitted N/G banks then move to the full frame.
The local prefix accepts all 93 dirty fields in B9 and all 104 Y fields in 4B9, with temporarily writable Y access. A common sufficient full-word domain is 104 Y fields in 4B9, all 78 N fields in 2B9, and the other 176 public fields in B9. Inputs are arbitrary within that product domain; minimality is not claimed. The eight original private-C zero/erase lifetimes remain in the retained projection.
Only
research/p104-inplace-entrance/is added. It contains the explicit 1,238-operationPROGRAM.json, 53 distinct recipes inNATIVE.json, proofs, complementary verifiers, source pins and deterministic evidence. Frozen predecessor files and the selected frontier are preserved.Evidence and cost
The integrated scientific gate checks:
The new prefix costs 887 unit Gaussian ADD/SUBs = 908,288 real ADD/SUBs, plus 105,472 exact real halves and restoring doubles each. Across the complete changed word, scalar ADD/SUBs fall from 4,457,472 to 3,969,024, while halves and restoring doubles each rise by 7,168 to 2,757,632.
COSTS.md inventories the new monomial wrappers and sign operations, states the scalar width envelope, and leaves finite native temporary widths and routing costs open; the safe common bound is$103\max(H_X,H_Y)$ in the empty-frame scalar stage. Verification arrays are not claimed physical workspace.
Conditional stopping decision
The old fixed-profile theorem does not automatically cover this resynthesis. This package explicitly substitutes the constructed profile into PR2744's retained$m=73$ , five-forward/four-dual coverage model. Its new lower slope is 75,068, and the favourable moment still exceeds one at the retained selected benchmark, for every finite number of fully paid auxiliary banks.
Pause dependent supplier construction within that consumer model. A different complete consumer, coverage law or sufficiently improved network needs its own bill. Actual upstream preparation, owners, permissible source access, controller/routing, finite native widths and all-size recursive transfer remain open. No new supplier, selected saving, human review or proof-kernel acceptance is claimed.
Publication verification
Both registered normal and optimised Python 3.12 replays reproduce receipt
6d06db043316af94377653a8e3fbfa1dcd686e20760af943b08e7060896ccf8b.The local committed-byte/source-history gate passes for all 31 files against
the sole declared source-history ancestor, main
20ecf2891d80b5acf18d980b2164fb44ea7faa90.The deterministic export and transport check also pass: 31 files plus export
metadata.
The published head is
13f974dae4af66027afb01f3ab42df9e830c2e90, with complete treec198eb6e5e66e8c13741c8754fb52ede40e2fa6c, identical to the independently tested local tree. A fresh public-commit byte/source-history check passes. An independent optimised Git-free replay passes, and three actual CLI copies reject WORD byte drift, a changed native phase, and Booleantruesubstituted for an integer receipt field. Every copy preserves all bytes and modes apart from its intended mutation. The public export retains the same 31 scientific/input files. Hosted CI and the downloadable public source export are reported in the publication comment.Reproduce with Python 3.12 and the standard library:
Task and next questions
Task
p104-side-correction, priority revision 6, with changed scope from joint flags to actual entrance acquisition and dirty output accumulation. Scientific base is20ecf2891d80b5acf18d980b2164fb44ea7faa90. Canonical main remains unchanged at the final publication checkpoint. The final 08:35 UTC update scan includes the new, distinct PR3605 staged-release construction on PR3133; neither it nor PR3594's fixed retiming family supplies this entrance contract.The pinned task and open questions retain the bounded overlap coverage, exact source heads, and NARROW decision. Reopening requires a changed complete consumer/coverage law, a fully priced replacement network, or an independently specified physical caller. The first step is its explicit whole bill and access/return contract, before dependent implementation.
Coordination and credit
Canonical investigation: Discussion 2750, comment 19014226. A separate published scalar finding proves the exact$4\mathbb Z[i]$ rescaling ideal for PR3577's fixed kernel witness; that bounded result is separate from this entrance construction.
Credit Dumas and Grenet's general in-place accumulation method; Nick Petalas's PR3462/3469 and PR3416; Hiroki Kumazaki's PR3446; ywz65's WORD/native lineage; PR3382's module interface; and the retained compiler and phase-checker authors. Credit Jose Manuel first for the consumer-budget insight, then Vasek's PR2744 audit and Ross Richardson's forward/native lineage. PR3554/3577 supplied the caller questions. Clear1nsight, with substantial OpenAI assistance from a root agent and three concurrent specialists; predecessor notices and attribution remain in force.