Research: metered QQ16 Gaussian axis on six bit tapes

[Clear1nsight]

Change and mathematical scope

Adds only the eight-file research/qq16-gaussian-axis-tapes/ package. Original fork PR: Clear1nsight#7. This is a signed-bit implementation of one exactly divisible, nonuniform Gaussian tensor-axis line primitive on two supplied data tapes and four owned scratch tapes. It includes movement, scratch zeroing and head returns, with exact READ=6NW, WRITE=6NW, MOVE=20NW and total 32NW metered cell actions.

For N=2^n, n>=1, signed little-endian two's-complement W>=2, and axis<n, the caller supplies separately based owned windows and must guarantee every numerator is even and in signed W-bit range. The program does not check these preconditions non-destructively. Four discardable scratch tapes may start with arbitrary bits. The outputs implement the pinned native formula, scratch ends zero, all heads return, and exterior cells are untouched. The carry proof covers every admitted n,W; finite fixtures are separate evidence.

Decision: close this bounded local primitive and stop. This is not a complete F0/B1 producer, consumer, dirty-helper return, admitted recursive supplier or kappa improvement. Input layout/encoding, addressing/control, adjacent monomials, cover maps, scalar instructions, restoration, F17 consumption and recursive/tape transfers remain chargeable. The conditional axis-only 354432lambdaN*W subtotal from #3303 is not a complete-producer bill.

Task and overlap

Scope: QQ16 physical tape implementation, refined to one caller-preconditioned axis component; TASK.json records task/dependencies. Base: 20ecf28; priority revision 6. Scientific head: 2be0f34. The browser comparison has eight commits and eight files, confined to this directory. Current AGENTS.md, autoresearch.md and CONTRIBUTING.md were read before submission.

Ali Romman's native QQ16 formula is retained, including the output convention discussed in #3652/#3671. #3303/#3270 give all-depth/factor-count evidence; #3241/#3377 address distinct access boundaries; #3394 prices owner-cover permutations. The caller/consumer missing in #3339 remains missing. This contributes the local coordinate primitive, rather than replacing those results or claiming a solved physical interface. Exact dependency hashes are preserved in SOURCES.json. No exhaustive novelty claim is made by this publication transfer.

Validation

python3 -I -B research/qq16-gaussian-axis-tapes/verify.py
python3 -I -O -B research/qq16-gaussian-axis-tapes/test_verify.py
python3 -B scripts/research_package.py --package research/qq16-gaussian-axis-tapes

The original investigation recorded local CPython 3.13.5 runs with -I -Werror -B and -I -O -Werror -B: 8192 signed pairs; 56 routed cases (dimensions 2 through 128); 72 signed boundary cases (widths through 545); 18 comparisons to the actual pinned 512-coordinate native line (nine axes, both signs); four consecutive axis/inverse operations sharing scratch; twelve rejecting behavioural controls. Every cell operation is metered, sources are authenticated, and the complete RESULT.json is compared. Published code/receipt Git blob hashes were matched against the tested bytes.

The registered Python 3.12 metadata/package replay is supplied but not claimed as executed in that local record. The fork merge shows two passing checks; that is not represented as full hosted scientific replay. No new execution was performed during this browser transfer. New upstream CI remains separate. Full F0/B1/native/selected cold replay, Lean verification, complete dirty-helper return, complete producer/consumer admission and human review are not claimed.

Credit

Credit Ali Romman's operator/native lineage; Nick Petalas (#3303/#3270); ToikY (#3652); #3671 for output convention; Sagar Thakrar (#3241/#3377); Dugongue (#3394); and the retained predecessor/source notices. New work by Clear1nsight with substantial OpenAI ChatGPT assistance. Apache-2.0. No independent human or formal-kernel acceptance is claimed.

Open next questions

See research/qq16-gaussian-axis-tapes/README.md and TASK.json. First bind one literal F0/B1 native event to actual separately based owner/frame/tape intervals and its signed-grid/scratch entrance. Then price one adjacent event pair without free relocation/addressing before expanding to the full 11076-axis layer and consumer. Finally test a named frame-relative borrowed-helper interface on arbitrary nonconstant data (#3293). Caller ownership, entrance protocol and uncharged layout/consumer transitions are blockers; no supplier expansion is claimed.

Workflow feedback

The original upstream creation attempt returned HTTP 403 through the integration. This browser-created PR publishes the focused contribution separately and retains fork #7 as its historical review record.

[Clear1nsight]

FINDING — the first-caller follow-up is now PR3932, scientific head 12c324c576821f78070627c5eeb684d03704e20e; fork PR10 is updated.

The actual first F0 child is event 9894 / L546 / frame 0→4. A reachable nonconstant impulse obstructs unconditional unchanged-grid substitution. One extra fractional bit repairs this rank-one child’s arithmetic; PR3303’s existing finer format already passes that local test conditionally. Source CLEARs identify an abstract C0/C1 scratch candidate, while separate physical tapes/origins and the complete connection, adjacent ADD, return and consumer costs remain unbound. No physical caller, supplier or new kappa is admitted.

Normal/optimized Python 3.12, 1024 native columns, 20 dense tape cases, 15 test groups, registered replay, read-only Git-free export and publication-byte/source-history checks pass. Fork hosted replay passed; upstream CI is separate on PR3932. The updated handoff releases the remaining physical task. Source attribution and substantial OpenAI assistance are retained.