Add GPT-6 Astra evaluation artifacts (high / xhigh / max) for 99 CPU tasks
Adds 3 runs of GPT-6 Astra on the 99 selected CPU tasks, one directory per reasoning effort: gpt-6-astra-high/, gpt-6-astra-xhigh/, gpt-6-astra-max/. Every recorded attempt is retained (405 attempts). GPU task tb3/exam-pdf-eval was not run.
Scored results (one attempt per task; the later attempt replaces the earlier one):
gpt-6-astra-high/: 48 pass, 48 fail, 1 invalid, 2 infrastructure_error; valid pass/fail 96/99; 122 attempts recorded. Format-tolerant variant (separately labeled): 50 pass, 48 fail, 1 invalid, 0 infrastructure_error.gpt-6-astra-xhigh/: 44 pass, 54 fail, 1 invalid, 0 infrastructure_error; valid pass/fail 98/99; 115 attempts recorded.gpt-6-astra-max/: 34 pass, 62 fail, 3 invalid, 0 infrastructure_error; valid pass/fail 96/99; 168 attempts recorded.
Each directory contains manifest.json, summary.json, scores.json (scripts/score.py), attempts.jsonl (every attempt with its round and whether it is the scored one), notes.txt, artifact-index.jsonl (size and SHA-256 of every archive member) and task-artifacts.tar.gz (per attempt: results, run config, clocks, episodes, coding-agent / simulator / judge call logs, judge and checker reports, patches and logs). Extract the archive inside the run directory; manifest artifact_path values name archive members. Directories with a format-tolerant variant also carry summary-format-tolerant.json and scores-format-tolerant.json: format-only breaks (one action wrapped in prose or repeated) tolerated, principle breaks still rejected; not the standard score.
Setup: upstream main 5933dd2, harness 0.5.0rc9. Coding agent, user simulator and judge on Amazon Bedrock (us-west-2), through a local SigV4 forwarder exposing Bedrock's OpenAI-compatible endpoint unchanged. Simulator and judge: gpt-6-luna at reasoning_effort medium (README's default is an Azure deployment). Coding agent: reasoning_effort high / xhigh / max, max_completion_tokens 128000.
Policies and deviations (details in each notes.txt): a blocked calibration gate was re-sampled with a fresh cache (up to 8 attempts); every invalid or infrastructure_error task and every fail with total_timeout or api_interrupted was rerun once; targeted fix-ups for infrastructure interruptions and judge protocol errors; drained batches continue in "-cont" batches (same attempt). Raw error classifications are preserved; 99 recorded tasks are not 99 valid completions.
Redaction and exclusions: no credentials were present; local machine paths were replaced with "~". Copied runtime and project sources, images, Docker build contexts, workspace snapshots and other binaries were excluded.
Validation: 99 unique task IDs per directory, 405 attempt records, status totals, 29,372 archive member hashes, presence of every scored attempt in its archive, and sensitive-pattern scans passed locally.