aletr-0.1 — open-model agent traces on Agents' Last Exam

Full-fidelity LLM traces from open-weight models run through the ALE-Claw harness on ALE CLI tasks, all inference served locally (vLLM). Campaign: 9 models x 25 tasks x 3 seeds, plus a Gemma-4 thinking on/off extension (~1,370 graded units, ~36,700 LLM calls).

Layout

• parquet/ — the derived tables (join-ready):
  • units — one row per (experiment, task): status, score, termination, and session_ids, the join key to per-call run tags
  • llm_calls — one row per LLM exchange: timing, usage (incl. cached and reasoning tokens), parse/reasoning/logprob flags, blob hashes
  • tool_calls, steps, tokens, runs
• proxy_logs/<serving-key>/calls.jsonl — append-only per-call index; run_tag is <experiment-tag>/<12-hex session id>
• proxy_logs/<serving-key>/blobs/<aa>/<sha256>.json.gz — content-addressed full request/response bodies (logprobs, reasoning channels, tool calls). llm_calls.request_sha/response_sha reference these.

Provenance notes

• Proxy-injected fields are recorded per call in the index (injected): logprob requests, image stripping for text-only mains, Gemma enable_thinking, and seed perturbation on gpt-oss harmony-parse 500s.
• Some experiment names carry v2/_think suffixes; superseded earlier attempts remain in the logs for forensic completeness. Use units to select the experiments of record.