# Submitted dataset mirror

`hf-datasets.zip` is the unchanged dataset archive recovered on October 4, 2026 from the [anonymous repository linked by the accepted paper](https://anonymous.4open.science/r/burstiness-was-measured-wrong-lp4fm).

- SHA-256: `4e1e7aced0203b9c20e4d52f04cc1111be9a9de8ffab58a42d551250bc83f47e`
- Size: 17,979,599 bytes; 11 files, 55,890,631 bytes uncompressed.
- Contents: frozen 200-article human baseline, earlier 100-pair ablation corpus, eight raw generation legs and the original README.

The human baseline is attributed to Wikinews contributors under [CC BY 2.5](https://creativecommons.org/licenses/by/2.5/). Each baseline record preserves an article title, date and `source_url` for article/contributor attribution. The human text was extracted and normalized for research. Generated records are machine-generated outputs, identified by model/provider; Kimi K2.6 outputs are attributed to Moonshot AI and are not human-authored. See [`/assets/THIRD_PARTY_LICENSES.md`](/assets/THIRD_PARTY_LICENSES.md).

## Availability and verification

Seven archived raw legs reproduce the committed scored JSON byte-for-byte using the paper's normalization and archived scorer: DeepSeek V4 Pro, Gemma 3 27B, Gemini 3.1 Pro, GLM-5.2, Grok 4.3, Grok 4.5 and Kimi K2.6.

The Qwen3.6-27B mirror contains **2,987 rows**, fewer than the 2,996 usable generations described by the published scoring run. Recomputing it changes some few-shot, plan and consistency statistics; for example, few-shot median B_n is 0.118 from this mirror versus 0.119 in the published scored snapshot. The archive is preserved as submitted, and the published scored data is unchanged.

Claude Opus 4.8 raw outputs are absent from this submitted archive. The original README's broad end-to-end reproduction statement must therefore be read with these two limitations. All nine published scored snapshots remain available, and all 261 checked printed entries match the accepted PDF.

## Separate October 18 handoff

The original 3,000 Opus generations have since been recovered and uploaded to a private collaborator dataset. Separate new DeepSeek runs contain 600 paraphrase outputs and a 3,000-output baseline from the later August 13 model snapshot. These private raw files are excluded from this unchanged archive and the public package. Collaborator access is required for the linked Hugging Face datasets.

See the [handoff notes](/assets/camera-ready-data-2026-10-18.md) and [verified inventory](/assets/camera-ready-data-2026-10-18.json) for exact revisions, hashes, token-budget disclosures and the retained off-topic baseline flag. Sentence-segmentation correction and rescoring are pending; the accepted PDF and scored snapshots remain unchanged.

See [`/assets/source-provenance.json`](/assets/source-provenance.json) for original file hashes and [`/assets/PACKAGE-CONTENTS.md`](/assets/PACKAGE-CONTENTS.md) for offline reproduction commands.