Every Model's Actual Answers — Prompt · Correct Answer · Model Answer

📖 How to review: click any model name or the green view answers → button. On each model page, every test card shows three things: the prompt (top), the correct answer (gold panel, left) and the model's actual answer (white panel, right — click to expand). Each card opens with a one-line plain-English answer, then every field in rows. Category sections collapse: click a category to open or hide its tests. Models still completing the suite are shown in yellow-striped rows with their current partial score and test count.

⚖️ How grading works (explicitly): every response is graded by deterministic Python code (scorers.py) against frozen ground-truth files (truth_new.json / truth_canonical.json) that were computed programmatically from the same CSV data the models saw. NO LLM grades any response — no judge model, no AI evaluation, ever. Three check layers: (1) numeric presence within tolerance bands (e.g. ±5% for ratios, ±1–5% for money), (2) regex/phrase presence for required findings, (3) an entity whitelist that flags any deal/company alias not in the source data (fabrication detection). Every check shown as ✓/✗ chips is one of these deterministic checks. The identical code and identical truth files grade all 119 models — byte-for-byte the same standard. The full methodology is in the test specification.