Benchmarks

77.7% on Stanford’s LaborBench.

One run of Stonewall’s harness, not the signed-in app, on Claude Opus 5.5 across 1,647 statutory questions: above published Lexis+ AI and Westlaw AI, below the study’s STARA.

See the full result

Statutory research

LaborBench

Accuracy and F1 · %

Afane et al. · CSLAW 2026 · Table 1

Method and sources

Published study of 1,647 statutory questions, using the same uncorrected DOL labels. F1 is one score that falls when the system misses provisions that exist or asserts ones that do not. These are study results, not a current-product retest; Westlaw received compressed context.

Stonewall’s result is measured, not estimated: one complete local Shelby harness run on 28 Oct 2026 with Claude Opus 5.5 at high effort, all 1,647 questions graded against the same DOL labels with the same metrics. 95% bootstrap intervals: accuracy 75.7–79.8%, F1 74.5–79.1%. This measures the harness, not the deployed application. Stonewall answered without a statute database. LaborBench has been public since May 2025, so the model may have seen it in training. The study rows are the study’s published results, not a same-day retest. STARA remains ahead.

Agent work

HLAB-Vals-1.1

Task pass rate · %

Vals AI · 26 Oct 2026

Method and sources

Harvey created this benchmark; these are model scores, not a score for the Harvey product. Vals version 1.1 results are not comparable with the archived v1 board or the independent LAB-AA evaluation.

Stonewall’s measured score and projection are not yet published. An estimate will be labeled separately from a measured result.

Harvey’s own product evaluation uses a different method and cannot share this ranking.

Pattern graph

What travels together.

Stonewall tags each document in a litigation record with the patterns it shows. These ten labels are stand-in names for ten of those patterns, the real names stay private, and the counts are real. Pick a label to see what travels with it.

Pick a label

Ten demonstration labels

Labels run clockwise from the top of the ring. Pick one to see which others share its documents.

    The names are stand-ins; the document counts are real. Arcs show how many documents carry each label. Ribbons show the documents two labels share, on the same scale; overlapping ends mean a document carries three or more labels. Pairs sharing fewer than two are not drawn.

    Three readings. Each is computed from the same counts the graph draws.

    1. Most partners

      —

      The label that shares two or more documents with the most other labels.

    2. Least overlap

      —

      The label with the lowest share of documents in common with its closest partner.

    3. Three labels at once

      —

      Labels whose shared counts show that some documents carry three or more labels.

    See Stonewall work with sources