Add certified BenchCheck evaluation logs: README.md

results/benchcheck_agentic_20260918/README.md

@@ -0,0 +1,35 @@
+# BenchCheck: scores and per-question answers
+
+Results from the complete 840-question VIDEO evaluation on 2026-09-18.
+
+| Model | Correct / total | Accuracy | Failed attempts |
+| --- | ---: | ---: | ---: |
+| GPT-6-Astra | 656/840 | 78.10% | 1 |
+| Claude Fable 5.1 | 583/840 | 69.40% | 38 |
+
+- [summary.csv](summary.csv) and [summary.json](summary.json): overall scores.
+- [category_scores.csv](category_scores.csv): capability, fine-category and source-benchmark scores.
+- [astra_answers.csv](astra_answers.csv) and [fable_answers.csv](fable_answers.csv): 840 answers each, with the published answer key, correctness and status.
+
+Each sample_id identifies this evaluation's item; item_id matches the source
+dataset. expected and prediction are answer letters. Empty predictions denote
+failed attempts. The CSVs preserve every failure as incorrect in the denominator.
+Fable's failures comprise 21 timeouts, eight OAuth errors, eight interruptions
+and one duplicate-terminal parsing error. Astra had one API error. Failed
+attempts were not retried or selectively recovered.
+
+Both models used maximum effort, local tools, model-directed video sampling,
+eight concurrent questions and 1800 seconds per CLI invocation. No BLIND run
+was performed. This differs from the benchmark's fixed-sampling protocol;
+the observed difference includes CLI/infrastructure reliability.
+
+Source dataset revision: 05e1acb6d6f6db9a6cf239795edfa21a1f75543a.
+The stored videos are already temporally subsampled and have no audio; original
+questions and labels were kept unchanged.
+
+This export contains only scores and per-question answers. It contains no
+frames, crops, videos, embedded image payloads, prompts, model process messages,
+tool commands/output, reasoning text, or raw execution logs. The CSV columns
+are explicitly selected from the audited results. MANIFEST.json records source
+checksums, FILES.jsonl records export checksums, and COMPLETE.json is published
+last after all remote files are verified.