Full results

By category

Where the job is hardest, across every model. Each row is the share of all 7 models’ answers in that category that fell into each of the four states. Percentages, because the categories differ in size; every row sums to 100.

• Confirmed — every judge agreed, or the single dissent rested on evidence that is not in the brief.
• Provisional — one judge dissented and its evidence checked out. Leans toward the model.
• Contested — the panel split evenly. Leans away from it.
• Failed — a majority rejected the answer, or the wrong tool ran.

Raw data

One file per model, containing every question, response, per-judge score and reason, token count, latency and cost.

The same four states as above, reported here as cumulative bounds rather than shares — the question is how well each model did, not how hard each category is. Confirmed is the floor; likely adds provisional; ceiling adds contested as well; failed is the other end. Breadth is the gap between confirmed and ceiling — how much the panel could not settle, which is a different question from how much a model got right. Confirmed, breadth and failed account for all 189 questions.