# Publication claim audit — 23 October 2026

The audited manuscript has useful engineering results, but its original current-results section and several related-work entries were not defensible as written. The isolated correction makes the contribution clearer without implying universal superiority or a new entropy-coding theory. It is a review package, not an independently validated publication or an integration into the peer-owned main tree.

## Review outputs and integration status

- Corrected manuscript, HTML, Typst PDF, LaTeX export and website: `target/publication-claim-review/` in the repository workspace.
- Source correction: [publication.patch](publication.patch). Apply only after checking [source_manifest.json](source_manifest.json) against the current files; the runtime owner may have changed them meanwhile.
- [Validation results](validation.json) record calculations, render consistency, table structure and export checks. [Evolution entry](EVOLUTION_ENTRY.md) is prepared for the integration owner.
- Main-tree runtime, version, manuscript and site remain under ANTIGRAVITY's ownership. The collaboration handoff identifies the correction package; no architect relay is required.

The runtime is unchanged by this audit. No new benchmark was run or silently substituted for a frozen experiment. No publication, release certification, upload or independent replication was performed.

## Rebuild the reviewed publications

After the integration owner checks and applies the patch, use the existing publication pipeline:

```sh
git apply --check docs/reviews/publication_claim_audit/publication.patch
# Apply only to the intended source snapshot after resolving any peer changes.
git apply docs/reviews/publication_claim_audit/publication.patch
python3 docs/paper/render_paper.py
python3 docs/paper/generate_pdf.py
python3 tools/render_site.py
python3 tools/build_site.py
```

Pandoc is required for HTML/MathML and PDF conversion. The PDF generator requires Typst (or its documented environment override); a TeX engine is separately required to validate the LaTeX submission. Generated review exports are retained in [corrected_publication.tar.gz](corrected_publication.tar.gz) with [SHA-256 fingerprints](corrected_manifest.json). Merge the prepared evolution entry into the current journal rather than replacing peer runtime entries.

## Findings and concrete corrections

| Severity | Original problem | Evidence / correction |
|---|---|---|
| Blocking | Current abstract named rc8 while section 4.1 reported 4.1.0 times and sizes | Replace the complete section with exact rc8 values from the final frozen `panel_history.json` entry, including every reference and fixed-preset win count |
| Blocking | Native small described as a modest CPU cost | rc8 uses 4,203.084 ms versus 13.014 ms for reused-workspace Zstd L3: about 323 times longer; make that cost explicit |
| Blocking | Related work implied PIZD introduced correlations beyond LogReducer and contrasted losslessness with CLP | LogReducer already describes correlation identification; CLP and μSlope are lossless. Rewrite comparisons as related mechanisms without unsupported priority claims |
| High | Incorrect μSlope, LogReducer and LogFold titles/authors/pages; unverifiable Bitshuffle bibliographic record | Replace with publisher/author records linked below; correct log-parser authors and distinguish Bitshuffle software from its cited paper |
| High | Current panel called binary-unit displays decimal MB and MB/s | Recompute current tables from integer bytes and milliseconds. 5,946,045 B is 5.946045 decimal MB, not 5.67 MB |
| High | Head-to-head counts could combine different PIZD configurations | Publish `wins_by_codec`, with one fixed preset in each row; separate metric wins are not joint dominance |
| High | 28–80% correlation savings, 2.46× matcher gain and −44.7% JSON probe gain lacked sufficiently pinned supporting artifacts at the claim | Remove these numerical assertions. This does not prove the experiments false; they require their exact inputs, versions, controls and raw measurements before reinstatement |
| High | JSON probe portrayed as a production-protected transform | Label `json_record_probe.rs` research-only; a prototype comparison is not a verified production integration |
| High | Size arbitration said to account for reconstruction cost | Distinguish complete serialized cost from separately measured reconstruction time; no multi-objective optimizer is established |
| High | Submission README called stale exports ready-to-upload and seamlessly compiled on arXiv | Regenerate exports from canonical Markdown, synchronize abstract, and explicitly retain unverified TeX/arXiv compilation status |
| High | Adaptive wording implied deterministic archives despite randomized equal-score keyword ordering in rc8 | Explicitly record observed archive variability, the two-byte baseline discrepancy and the pending isolated lexical-tie correction; distinguish reproducible source restoration from reproducible archive bytes |
| Medium | Website mixed historical cards and a newer source package | Explicitly distinguish frozen public scoreboard rc1, generated panel rc8 and isolated rc9 candidate; retain each experiment's identity |
| Medium | Lazy matching called optimal; SIMD parsing called zero-copy; scoped threading described as eliminating synchronization overhead | Replace with local-heuristic, architecture-dependent search and fully charged pipeline descriptions |
| Medium | HTTP/DNS/cloud examples presented empirical relations as universal laws | Explain learned mappings with exact exceptions; upstream dependency indices exclude cycles but do not alone implement a concurrent schedule |
| Medium | HTML renderer merged numbered references and displayed raw formulas | Use Pandoc's structured HTML and native MathML for manuscript body, without a remote rendering dependency; retain the existing page shell |

## What the corrected current panel actually supports

The generated 18-source development panel measures rc8, 5,946,045 original bytes, three-iteration arithmetic means and no warmups. It is not the historical 21-trial reviewer experiment. Source-verifying decode includes BLAKE3; independent SHA checks run outside timers. Reference streams contain an additional benchmark-specific digest envelope.

- Native small: 260,141 B versus Zstd L3's 560,319 B, or 53.57% fewer bytes. Encoding is about 323 times slower than reused-workspace Zstd L3.
- Hybrid fast: 455,429 B; 33.796 ms encode versus 13.014 ms for reused-workspace Zstd and 139.959 ms for Brotli Q6. The size and time objectives differ.
- Native small wins size on 17/18 inputs against the smallest measured reference, but wins neither encoding nor verified-decode speed on any input against the best reference for each metric.

These are observations on generated, familiar inputs and explicit presets. They do not establish that 94% of real deployments benefit, a current-release speedup over historical versions, or superiority over specialized compressors.

The separate isolated rc9 direct candidate reduces public30 Native thin/template bytes from 687,537 to 657,356 (4.39%). Its verified in-process decode is 5.778 times slower. These figures are paired candidate evidence, not the website's rc1 scoreboard. The stronger 656,103-byte research portfolio is not the direct runtime output. See [candidate evidence](../../benchmarks/adaptive_native_rc8/README.md).

## Value added that can be stated now

PIZD implements exact-layout structural projections, data-dependent typed and categorical models, exception-aware dependencies, and complete-cost choices among evaluated representations. The useful result is the measured size–CPU tradeoff of that integration, including cases where generic transforms and specialized field models lose. Finite exact-source restoration and old/new format checks support implementation capabilities on tested inputs; they are not proof of universal safety.

This is an engineering contribution. Schema separation, dictionaries, numeric deltas, field correlations, PPM and interleaved ANS are established. The project has not shown scientific priority for these mechanisms, a new coding theorem, or superiority over CLP, μSlope, LogReducer, LogPrism or LogFold in a matched experiment.

## Primary-source reference checks

- [LogReducer, FAST 2021](https://www.usenix.org/conference/fast21/presentation/wei): actual title, pages 249–262, parser-based approach and existing correlation work.
- [μSlope, OSDI 2024](https://www.usenix.org/conference/osdi24/presentation/wang-rui): actual title, authors, pages 529–544 and existing schema/table representation.
- [CLP, OSDI 2021](https://www.usenix.org/conference/osdi21/presentation/rodrigues): lossless compression and compressed search.
- [LogFold author manuscript](https://arxiv.org/abs/2603.20618): actual title and authors; structured-token and type-dependent encoding overlap with the problem addressed here.
- [LogPrism author manuscript](https://arxiv.org/abs/2601.17482): related structural/variable modeling, not a matched PIZD result.
- [Log-parser benchmark author manuscript](https://arxiv.org/abs/1811.03509): correct author list and publication context.
- [Bitshuffle official repository](https://github.com/kiyo-masui/bitshuffle) and [its cited paper](https://arxiv.org/abs/1503.00638): verifiable software/paper citation.
- [Zstandard RFC 8878](https://www.rfc-editor.org/rfc/rfc8878.html), [Brotli RFC 7932](https://www.rfc-editor.org/rfc/rfc7932.html), [ANS manuscript](https://arxiv.org/abs/1311.2540), [Gorilla paper](https://www.vldb.org/pvldb/vol8/p1816-teller.pdf) and [BLAKE3 authors' specification](https://github.com/BLAKE3-team/BLAKE3-specs): mechanism and attribution checks.

The review checks bibliographic identities and overlap, not formal equivalence of every implementation or an exhaustive prior-art search. Historical component claims retain their original version and evidence boundaries; this audit does not rerun their measurements.

## Remaining objections that wording cannot resolve

1. **Independent generalization:** current public excerpts and later windows are familiar source families. Freeze policies before acquiring new families, include drift and irregular layouts, and preserve failures.
2. **Specialized competitors:** compare actual implementations under a documented exact-restoration contract, full metadata accounting and matched resource budgets. Abstracts from other papers are not benchmark rows.
3. **Current causal attribution:** run fixed-backend ablations on current models, including metadata and rejected-candidate CPU. The older reviewer controls remain historical evidence.
4. **Host and effort dependence:** one host, unequal level effort and short excerpts do not settle deployment throughput, large-data density or cross-platform Pareto frontiers.
5. **Submission readiness:** Typst PDF rendering is checked locally. The regenerated LaTeX has not been compiled with an arXiv-supported engine in this environment, and arXiv acceptance is not established.

An honest release can explain these limits and its measured practical gains. Calling the paper unassailable would itself be an exaggeration.