Summary

Hosts the baseline rollout traces consumed by the can-ai-agents-improve-ai-agents starter — the companion code for the Can AI Agents Improve AI Agents? TensorZero blog post.

The starter ships in tensorzero/tensorzero, but the inference traces (98 MiB) are too large to live there — pre-commit blocks at 1 MiB and that repo doesn't use Git LFS. The starter's baseline_data/fetch.sh will download these jsonls from this repo's LFS-backed media URLs.

Tracked through Git LFS — *.jsonl filter=lfs ... added to .gitattributes (next to the existing *.csv LFS rule).

Files

Provenance

Real baseline rollout of yc_bench_tutorial_v0::yc_bench_act against the initial variant on openai::gpt-5.4-mini: 80 unique train tasks + 20 unique test tasks (Codex YC Bench seed 0, 2026-04-23). Matches the artifacts the autopilot-evals harness dumps to <run_dir>/claude_code/baseline_data/ before invoking the optimizer agent.

Raw URLs (post-merge)

https://media.githubusercontent.com/media/anndvision/data/main/can-ai-agents-improve-ai-agents/baseline_data/inferences.jsonl
https://media.githubusercontent.com/media/anndvision/data/main/can-ai-agents-improve-ai-agents/baseline_data/feedback.jsonl

(LFS-backed files use media.githubusercontent.com/media/... to fetch actual content rather than the LFS pointer that raw.githubusercontent.com returns.)

Test plan

• Both URLs resolve post-merge and serve the correct bytes (SHA-256 matches the table above).
• Starter's bash baseline_data/fetch.sh downloads + verifies both files.