Earlier benchmarks

The questions changed. The original prompts, answers, and scoring records are still here. Scores from different versions are not comparable.

1.2 · Three everyday programs

Weighted bills, contact CSVs, and meeting times. The first GPT-6 Sol attempt failed at the provider and remains unscored.

Models wrote test traces to expose eight ledger faults. All 16 calibration attempts are preserved.

The ChatGPT subscription board, scored by GPT-5.6 Sol on a 0–10 scale.

The original runs using Sonnet 5 as judge. Kept separate from the subscription results.