Run outputs moved out of the task directories in the agent repository to
results/<label>/<task>/; the leaderboard's declarations (provider, model, effort)
come from the run summaries there. --results names the root, default
../agent/results.
25 days ago
The leaderboard from the agent package, shared with cataflow-bench-runs
sync.py renders the card's section with cataflow_agent.benchmark.leaderboard;
the local copy and its tests move there, so both dataset repositories draw the
same table from one renderer. The section is unchanged.
The r101 and r7 tasks, the index, the card with the r101.2 leaderboard
Thirteen generated tasks as the sandbox tools import them, each with exactly the
files its manifest names; their plans and recipes under truth/; tasks.jsonl with
counts, scenario tags and content hashes; the dataset card with the leaderboard of
benchmark r101 version 2, read from the runs' own declarations. sync.py publishes
tasks from the agent repository and rewrites the leaderboard; leaderboard.py
renders it. PDFs, spreadsheets and pictures go through LFS.
