Paper numerical reproduction — pprv1-20260910

Current manuscript: paper-r1.pdf and source. The explicit erratum corrects the Appendix H portability wording and one verb agreement. Original archive and paper assets are retained unchanged.

Public access verified: the archive was downloaded without GitHub credentials, matched its SHA-256 and passed all 98 strict checks in a fresh environment. The corrected manuscript assets were also anonymously downloaded and hash-verified. See publication receipt, anonymous replay and hosted replay.

Numerical reproduction package for Painter Naming and the Distributional Gap Between Generated Images and Original Paintings.

The archive contains the revised English paper, all eight figures, unchanged scientific computations, locked dependencies, compact measured features, prompt inventories, source metadata and explicit replay coverage. The temporal successor includes 264 new outputs measured in three pipelines. Both FLUX naming directions recur; the two additional named-palette effects remain unresolved.

Run the README command in an extracted archive to recompute the numerical results and figures. This release supports numerical replay from retained measurements. It excludes raw artwork, generated pixels, private transport responses, credentials and local Git history; it does not support public feature re-extraction. Source URLs and recorded license fields do not grant redistribution rights to absent artwork.

The new cohort was collected by the same maintainer. Computational challenge responses and hosted numerical replay do not establish human-perceptual validity, independent capture equivalence, or replication by independent investigators.

The manifest's prepared_not_published field records artifact-build status. The separate publication verification receipt records actual archive access, checksum and hosted/anonymous replay results.

Archive: pprv1-20260910.tar.gz (11,851,958 bytes).
SHA-256: 165ffbde2ba234e80a1da450f0e10453b127d79538317b6113ed8c145c6c6de4.

Sanitized source commit: 2592dfbe6667586e30d945119418bfefe24690ad.
The original copied inputs were verified against local source commit b2884c3cbc1ec281dc9d4383b3aceed90d6eff90; the archive does not include that local Git history.

The portable numerical check uses the recorded 1e-10 absolute/relative tolerance. A documented post-CI adapter correction applies that existing bound to recognized continuous Welch p-values: the diagnostic run found six copied values differing by at most 5.551115123125783e-17. Randomization p-values/counts, unknown p-valued fields, identities and scientific decisions remain exact; retained scientific outputs and hashes are unchanged. Strict local replay remains a distinct check.

Verification completed before publication:

• Fresh locked macOS replay: all 98 checks exact, including all eight PDF figures.
• Hosted Ubuntu run 34425107886: all 98 checks passed the documented portable contract; seven figure PDFs match byte-for-byte, and the challenge matrix renders from verified inputs with platform-dependent PDF bytes.

The full hosted receipt records continuous Welch differences up to 1.1102230246251565e-16 across the original and fresh palette cohorts; none changes an inferential decision. Anonymous archive-download verification passed and is recorded in the separate publication receipt.

Painter naming geometry: numerical addendum

This additive release contains the revised English manuscript and the complete numerical replay of the separately frozen painter-naming geometry and evaluation-centering analyses. It supplements the unchanged pprv1-20260910 release, which remains necessary for earlier paper results.

The archive includes 47 allowlisted files, nine figure PDFs, compact measurements, protocols, source, tests, hashes and dependency lock. It excludes raw artwork/generated images, provider responses, credentials, model weights, local Git history and Korean user drafts.

Both scientific analyses reproduce exactly in a fresh Python 3.13.11 environment; the 95 included scientific tests pass. Verify the extracted inventory before installing dependencies, then follow the two module commands in README.md. An initial local draft exposed a duplicate pytest filename collision; the final command uses --import-mode=importlib and was verified successfully.

Archive SHA256: 4a26b9c3f089d0a0c1dfaf7bd45a86f6658d6ee89ad5f6a47d6994a3ce5a9b30.
Local source provenance: 5485e36 (distinct from the sanitized public commit).

These are post-result computational diagnostics, not independent acquisition, human evaluation or institutionally independent peer review. The manuscript describes measurement, reference-panel and generalization limits. Numerical replay does not authenticate absent source pixels.

Prospective fixed-map numerical comparison — 10 September 2026

This additive numerical release covers the prospective FLUX/Cézanne fixed-map comparison on 12 new authored scenes: 240 images, 120 paired repetitions, and two fixed endpoints. Translation/scaling improves both reference energy (difference −0.390187) and repeat-corrected conditional-mean prediction (−5.046510); both approximate simultaneous intervals are below zero. The earlier opposing map ordering does not transfer to this panel.

The archive preserves the failed smaller-allocation planning record and the separate passing R10 qualification. It includes the fixed numerical source, 31-feature vectors, assignments, and terminal analysis. Its public check replays all 27 qualification cells and both observed endpoints without generation, feature extraction, or formal output writers. The observed Q interval half-width is 2.833, exceeding the proxy planning criterion 1.0; qualification did not guarantee achieved precision or actual service coverage.

Download pmv2r-20260910.tar.gz and its detached checksum; use the commands in the extracted README. The 74-file archive is the numerical replay root. The history-free release branch additionally contains a hosted-replay workflow outside that archive. Earlier numerical releases and paper assets remain unchanged.

Qualification floating values use the recorded 1e-10 absolute/relative tolerance; identities, counts, seeds, decisions, and reconstructed support hashes are exact. Observed analysis and report must match exactly. These restrictions can limit cross-platform replay. Local acquisition checks are represented by public attestations: images, raw responses, private metadata/credits, credentials, and repository history are excluded. This is maintainer-operated numerical reproducibility, not independent image acquisition, perceptual validation, or replication by separate investigators.

The final 39-page English manuscript is now available as paper.pdf, with a standalone 16-file paper-source.tar.gz and paper-assets-SHA256SUMS, from manuscript source commit 2969a31acb197ed5ae8fe728c2a0ae13f4c7c36e. Fresh Tectonic 0.17.0 compilation matches every page in text and rendered pixels; compiled PDF byte identity is not claimed. Anonymous downloads match all three published hashes and the geometry presentation replays exactly. These paper assets are separate from the unchanged numerical archive. No raw images, credentials, other local drafts or repository history are included.

The original strict Ubuntu attempt fails qualification comparison before observed replay. A separate completed diagnostic records 27 failing support hashes; all other qualification numerical differences are below 3.56e-15. Separately reached observed differences are below 2.67e-15 and the displayed scientific table is unchanged, while exact JSON/report comparisons remain false. The diagnostic does not turn the original strict attempt into a pass. The proposed fixed-support portability extension remains deferred.

Fixed-map precision qualification: reproducible numerical record

This additive numerical archive retains the complete failed fixed-map precision qualification: no allocation among 96, 144 or 192 outputs qualified, and no images were generated under that proposal. All 81 coverage criteria passed; the conditional-residual precision criterion failed at each allocation.

The 18-file archive contains unchanged scientific source, the fixed protocol, already-public historical numerical inputs, all 81 results, locked dependencies and a standalone replay adapter. Fresh local replay matches every numerical result exactly; all 73 included tests pass on the recorded macOS arm64/Python 3.13.11 environment. Exact reconstructed-support hashes can restrict portability across numerical libraries and platforms.

Verify the archive checksum and inventory before installing dependencies. The README gives the full commands and scope. Raw images, credentials, account metadata, provider bodies, private operational documents and local Git history are excluded. Replay does not validate actual service sampling, capture provenance or artistic similarity. Design, implementation and checks used maintainer-run LLM assistance, not separate human investigators. Earlier releases remain unchanged.

Painter-clause comparison: numerical addendum

This additive numerical archive preserves two separately analyzed, prospectively fixed painter-clause cohorts. The original 288-slot cohort returned 211 images and retains unavailable primary comparisons. The final, separate 96-image Cezanne/generic cohort completed all 48 pairs: primary energy change −1.195419353, Monte Carlo p=.00001 at the fixed .025 threshold, and observed named/generic trace ratio .552327.

The archive includes unchanged scientific source, protocols, allocation/precision records, locked dependencies, compact measurements, and both complete numerical reports. Its dedicated commands verify the inventory and recompute each cohort separately. Neither cohort is pooled with the other or with historical studies. Earlier numerical releases remain unchanged and are needed for earlier manuscript results.

Only allowlisted code and numerical records are included. The package excludes artwork, generated pixels, provider-response bodies, credentials, proxy snapshots, local Git history and Korean drafts. Replay does not establish perceptual validity, independent acquisition authentication or replication by separate investigators. Design, implementation and reviews were maintainer-run LLM work with involvement disclosed.

Archive SHA256: 6038da2daed74e6ed4b509464dc6f1a4dae386e644d265235de982db4bf23e40.

The final English manuscript is available as paper.pdf, with a self-contained paper-source.tar.gz and paper-assets-SHA256SUMS. The source contains the TeX, bibliography, nine figures, MIT notice and the readable Figure 3 renderer with its unchanged numerical input. A fresh Tectonic 0.17.0 build matches all 36 pages in text and 100-dpi rendering; PDF metadata may differ. All three assets were downloaded anonymously with matching hashes, and the presentation figure reproduces byte for byte using the locked geometry-release environment. The numerical archive remains unchanged and does not depend on these presentation assets.

Paper SHA256: f786c8a169dc0b5fb9f11df06b687837c6ab16cc48a6c9f492d613a1b2f4934c.
Source archive SHA256: 7e98f48ffe432b8d93a96d64a43fafe18e79c94139a473f3433dc9812a57e946.