Hinton-Test benchmark website

Interactive companion to Hinton-Test, a benchmark of the world model of the patient. The site exposes the verified manuscript results as searchable model, finding, organ-system, and task views.

Public website: mrgiovanni.github.io/Hinton-Test

Pushes to main automatically rebuild and publish the GitHub Pages site.

Rebuild the data

python3 scripts/build_site_data.py

The builder reads the canonical artifacts under docs/results/ and writes public/data/benchmark.json. It does not recompute scientific metrics.

Read cohort policy

The public Read leaderboard defaults to the original frozen cohort: 14,832 finding–examination cases for full-cohort endpoints, while GPT-5.5 and Claude Opus 4.8 retain their original 2,606-case audit. The later 16,532-case minimum-support expansion is preserved for future opt-in analyses in TMI-2026-Med-Reasoning/docs/results/read_breadth_min20_repair.

Visual and image provenance

• The paper plotting implementations under world_model_benchmark/figures/ are the authority for model families, model colors, metric direction, and finding order. fig_read_results/ is the reference for the assembled Read figure.
• The Read task includes a native web translation of figures/fig_read_findings: selectable per-finding score distributions, the sensitivity--specificity plane, and finding support versus performance. These panels read from benchmark.json; they do not scrape pixels from the paper figure.
• The dataset page uses the real de-identified three-examination trajectory FF000009 from the frozen no-action Predict cohort. Regenerate its slices with scripts/render_trajectory_slices.py; the follow-up examination is displayed only after an explicit reveal action.

Run locally

pnpm install
pnpm run dev

Pages

• / — benchmark story and headline findings
• /dataset — cohort and longitudinal structure
• /models — sortable task-specific leaderboard with clickable cross-task model profiles
• /tasks — interactive Assess, Read, Compare, Predict, Conclude, Advise, and integrated views
• /explorer — 179-finding search, body-system profile, and cross-model drill-down
• /submit — structured community model-submission workflow
• /about — methods and reproducibility resources