paper

Paper

Reproducibility and release status for the September 8 benchmark surface.

Updated

Status

A formal paper is not published yet. The site is the public benchmark surface for the active release.

Current release includes

Public leaderboard, September 8 model roster, per-track tables, statement sample, score schema, and submission preview.

Planned artifacts

Paper draft, harness notes, submission format, reviewer protocol, and clearer release/version history.

Reproducibility note

Current release is operator-run. Submission review and reruns are batch-managed while the public flow is being formalized.

Draft outline

  • 1. Task definition and the six tracks
  • 2. Statement pack, gold labeling, and adjudication
  • 3. Harness, normalization, and the three-pass protocol
  • 4. Results: models vs. accounting agents, Jun 29 and Sep 8 rosters
  • 5. Error taxonomy from 1,140 reviewed misses
  • 6. Limitations: single statement shape, USD only, public rows scored

Timeline

Internal draftOct 2026
Harness noteswith v1.3 pack
Preprinttarget Nov 2026
Reviewer protocoldrafted, not published

Cite the release

Until the preprint is out, cite the site and pack version:

eval.tax, "AI Accounting Arena," pack v1.2, roster of Sep 8, 2026, harness 1.2.3. https://eval.tax/

@misc{evaltax2026,
  title  = {eval.tax: AI Accounting Arena},
  author = {{eval.tax}},
  year   = {2026},
  month  = sep,
  note   = {Pack v1.2, harness 1.2.3, roster of September 8, 2026},
  url    = {https://eval.tax/}
}

Data availability

statement.csvpublic, 512 rows
Gold labelsshared with submitters under review
Client profile + vendor historywith harness notes, v1.3
Per-pass raw outputsretained 180 days, on request

Paper status: in progress. Until then, methodology, dataset notes, and submission instructions are the primary public references for the September 8 refresh.