submit

name	submit
description	Package a finished run into a TSEval Submission Report (result + agent trajectory + report) and contribute it via a GitHub PR to the TSEval leaderboard repo, via `tsf trace` / `tsf submit`. Use when the user wants to submit results to the TSEval leaderboard, capture an experiment trajectory, or build a leaderboard from submissions.

When to use / what to ask

The user wants to publish a result to the TSEval leaderboard (or audit-package it locally). Ask:

1. Dataset and model of the run to submit (must have run already — tsf submit reads work_dirs/<dataset>/<model>/records/<run_id>.json).
2. Which run — --latest (newest) or a specific --run-id.
3. Contribute to the board? tsf submit builds the bundle locally under work_dirs/_submissions/<submission_id>/. The TSEval leaderboard is GitHub-canonical: to publish, add that bundle to a clone of github.com/Diaugeia/TSEval under submissions/<track>/<dataset>/<model>/<submission_id>/ and open a PR (CI validates → aggregates → redeploys). There is no Hugging Face push; weights are never required. Confirm before opening the PR — it publishes the result.

Capture a trajectory first (recommended)

Start a trace session before running the experiment — it records every tsf command as audit evidence reviewers read. Agent-agnostic (CLI boundary).

uv run python tool/tsf.py trace start --label "<experiment-label>"
uv run python tool/tsf.py run configs/runs/<config>.toml   # commands are recorded
uv run python tool/tsf.py trace end                        # or: trace status

No session captured? tsf submit still works but marks the trajectory synthetic: true — a real one is preferred.

Package and submit

# Build the bundle — lands in work_dirs/_submissions/<submission_id>/
uv run python tool/tsf.py submit --dataset <DATASET> --model <MODEL> --latest

The bundle contains submission.json (schema-validated SubmissionReport), trajectory.jsonl, and report.md. To publish, add the bundle to a clone of github.com/Diaugeia/TSEval under submissions/<track>/<dataset>/<model>/<submission_id>/ and open a PR — CI validates against the TSF-Core schema, aggregates across seeds, and redeploys the board. (Full steps: SUBMITTING.md in that repo.)

Related commands

# Collate submissions into a ranked leaderboard.json (consumer side)
uv run python tool/tsf.py leaderboard-build --source work_dirs/_submissions --out leaderboard.json

# Regenerate / verify the JSON Schema contract (CI fails on stale schema)
uv run python tool/tsf.py schema-export [--check]

Reference

Full producer-side workflow and review process: docs/en/tseval-submit.md. The contract layer is src/tsf_core/ (pydantic only).