The Burstiness Engine
Reproduce the study
Scoring code, nine-family results, the human baseline and available raw generations.
Download code and dataNeurIPS 2026 · LP4FM Poster
Burstiness Was Measured Wrong, and Prompting Cannot Aim It
Liceo Scientifico Augusto Righi · Metaventions AI
A scale-free measure of sentence rhythm, a normalization audit, and a nine-family test of whether prompting can target human dispersion.
Code and data
Materials for the study.
Manuscript source
The submitted TeX, bibliography, style and figures, ready to import into Overleaf.
Download sourceExplore the numbers
The published estimates, confidence intervals and companion diagnostics in a machine-readable form.
Download benchmarkThe submitted public dataset mirror is unchanged: seven archived legs reproduce the published scores exactly, and the Qwen archive is incomplete. Recovered original Opus outputs and the new October DeepSeek runs are archived separately with private access. October 18 data update · Reproduction limits.
Camera-ready data · October 18, 2026
Recovered originals and separately identified new runs.
Raw data for the camera-ready handoff is archived. Access to the linked datasets requires collaborator permission; the public download contains provenance metadata, not these private generation files.
| Dataset | Rows | Provenance | Access |
|---|---|---|---|
| Original Opus 4.8 | 3,000 | Recovered original run; self-refinement drafts and critiques are archived alongside it. | Private archive ↗ |
| New DeepSeek paraphrases | 600 | October 18 generations: three variants × 200 frozen briefs. | Private archive ↗ |
| New DeepSeek baseline | 3,000 | October 18 generations: five modes × 200 briefs × three iterations; all 4,200 selected stages are preserved. | Private archive ↗ |
The new DeepSeek runs use deepseek-v4-pro, documented as DeepSeek-V4-Pro-0813 (August 13, 2026). This is a newer snapshot than the July runs. The original July paraphrase outputs were not recovered; the October outputs are new generations.
One DeepSeek baseline response remains included with an off-topic quality flag: self-refine, scenario 93, iteration 1. Its original output is preserved without rerun, replacement or exclusion.
Rescoring is pending
Vittoria Lanzo is handling the camera-ready rescoring. The PDF, benchmark tables and scored JSON on this site remain the October 1 publication. This data handoff does not publish revised findings.
Public provenance metadata (JSON) · Data handoff and reproduction notes
The paper
Open in a separate tab ↗The PDF retains the anonymous submission author block. Vittoria Lanzo and Dico Angelo are identified in the public OpenReview record.
Citation
Reference the study.
Vittoria Lanzo and Dico Angelo. 2026. Burstiness Was Measured Wrong, and Prompting Cannot Aim It. NeurIPS 2026 Workshop LP4FM, Poster. OpenReview.
BibTeX
@misc{lanzo_angelo_burstiness_2026,
title = {Burstiness Was Measured Wrong, and Prompting Cannot Aim It},
author = {Lanzo, Vittoria and Angelo, Dico},
year = {2026},
howpublished = {NeurIPS 2026 Workshop LP4FM, Poster},
url = {https://openreview.net/forum?id=1E20ig92Zi}
}Paper: CC BY 4.0 · Code: Apache-2.0 · Wikinews baseline: CC BY 2.5.
Publication record
Source and verification.
Artifact provenance and integrity
The PDF is the unchanged October 1 OpenReview manuscript. It was accepted as a poster on September 29. The submitted source, recovered from the paper’s anonymous artifact repository, builds to 20 pages with identical extracted text after whitespace normalization.
All 261 checked entries from Table 1, Appendix B and the multiplicity census match the scored artifacts. The source and PDF retain submission formatting; these materials do not imply a later camera-ready revision.
PDF SHA-256:
ee5301d4067b76d143aef274d5cf8339bb60c4aa1fbf0f36a63575cea083c8adSource hashes and validation · Publication metadata · Earlier versions