[
  {
    "date": "2026-10-09",
    "version": "1.0",
    "change": "v1.0 published from the rescore at llm-eval-suite 32c27a23 (runs at 42532423), with the AIEval DoW Bench Score (equal-weight average of the three suites). Phi-4-mini rows (pending re-judge) and Nemotron runtime stats (pending rerun) are provisional."
  },
  {
    "date": "2026-10-10",
    "version": "1.0",
    "change": "Final stage-1 results. Phi-4-mini-instruct and Phi-4-mini-reasoning re-judged by Granite 4.2 3B under the same-family rule (all 88 judge-graded items each, 0 over budget, 0 network entries; llm-eval-suite c52abb57): both now have full suite scores and a ranked AIEval DoW Bench Score. NVIDIA Nemotron 3 Nano 4B rerun after an end-token fix (run 7fef66483e47, llm-eval-suite 32c27a23): judged scores identical; run time, peak VRAM, and token-cap stats replaced. No rows are provisional. Judge agreement remains pending."
  },
  {
    "date": "2026-10-12",
    "version": "1.1",
    "change": "v1.1, Oct 12, 2026: added a Public benchmarks section (six offline public benchmarks, indicative and not ranked) and accessibility fixes."
  },
  {
    "date": "2026-10-12",
    "version": "1.1.1",
    "change": "v1.1.1, Oct 12, 2026: reasoning-model coding scores regraded with the corrected scorer (\u00b12 points on MBPP+ and one HumanEval+ score)."
  }
]