[jon314159]

PR6045 leaves combinations of carrier sources open, and the earlier endpoint screen tested individual donor availability. This package supplies reproducible, scoped negative results on the unchanged PR6030 word so those searches need not be repeated.

Results

• 4,033 leaking unit-injection classes: an exact global dependency filter excludes 4,009. Of the remaining 24, 22 have no first same-chart combination carrier. Dynamic affine closure excludes both survivors, including opportunities enabled by earlier corrections.
• 764 incoming-only MOVE intervals: the exact net incoming contribution fails the combined rational spans of both full legal endpoint windows. This includes split replacements and cancellation among incoming ADDs: 693 coordinate exclusions and 71 exact elimination exclusions.
• Five private rank-four query outputs: none can be reconstructed from other live destination-chart values at its immediate output seam, even granting equivalent frame presentations.

For the two dynamic cases, the final correction space and protected-output projection both have rank five. Independently implemented dense suffix, determinant and affine checks confirm that clearing protected outputs erases the whole new-helper signal.

Scope and reproducibility

All changes are in research/p104-fixed-word-negative-screens/. The selected result remains conditional κ = 0.003125804. These exclusions concern the pinned fixed word, specified cuts/windows, and unchanged lifetimes. They do not exclude corrections that temporarily change the original scalar map, changed injections or continuations, new MOVEs/storage, receiver scaling in the endpoint model, or other architectures.

The PR6040/6045 dependencies are retained byte-for-byte with source hashes and attribution. The unchanged PR6030 word is the sole external live input. Replay requires Python 3.12 and its standard library; it needs no pending predecessor commit or network fetch.

Validation

Passed on CPython 3.12.3 under Linux:

• Registered normal and optimized isolated replay; both produce semantic SHA-256 e9de43aef960895096b75c5e057ee0119276cdbdbbbc81fe698be60dae61f683.
• 27 module controls, independently implemented terminal certificates, adversarial boundary controls, and five explicit dual separating certificates.
• Clean exported-package replay and rejection of changed word bytes, changed frozen vendor bytes, and a forged semantic receipt.
• Metadata/source pins and exact committed-byte preflight for commit b3f5aabddbae2d65f80bb8356e6993e0a65cd245.

python3.12 -B scripts/research_package.py --package research/p104-fixed-word-negative-screens
python3.12 -B scripts/check_research_commit.py --package research/p104-fixed-word-negative-screens

The optional full-Git-history audit was not run in the shallow checkout; the self-contained export passed. No full native/compiler/caller/outer replay or new κ is claimed.

Predecessors: #6045, #6040, #5994, #5718, #6030 and #5827; related chronological repair theory: #6019. Exact credit and source notices are retained in NOTICE.md and VENDOR.json.

Prepared for Jonathan Capalbo (jon314159) with substantial OpenAI Codex assistance, including separate implementation and review agents. This is AI-assisted research and review, not independent human approval or formal verification.

[jon314159]

AI-assisted review completed for commit b3f5aab. No blocking findings within the declared fixed-word scopes.

Reviewed the complete baseline-kernel argument, dynamic affine enclosure (including newly enabled carriers), both-window net-contribution condition, private-query donor selection, source authentication, predecessor attribution, and claim limits. A separate AI review implementation independently checked the terminal affine certificates and explicit dual separating certificates.

Validation passed:

• Python 3.12.3 registered normal and optimized isolated replay, with the same semantic certificate.
• 27 module controls plus independent certificate and adversarial boundary checks.
• Clean exported-package replay; changed word bytes, changed vendor bytes, and a forged receipt all rejected.
• Metadata/source pins and exact committed-byte preflight (18 package files plus the declared word input).

The results establish only the stated finite scalar exclusions. No new κ, general optimality, native/compiler/caller/outer admission, independent human approval, or formal verification is claimed. The optional full-source-history audit was not run in the shallow checkout. Hosted CI is separate from these completed local checks.

Reviewed using OpenAI Codex, with separate implementation and review agents, at the author's explicit request.