| name | setup |
| description | Set up everything needed to participate in the Viktor Challenge — environment, dataset, sanity checks, first baseline. |
/setup — get a team from zero to a running baseline
Do these steps in order. Announce each result in one line.
1. Environment
- Confirm Python 3.10+ (
python --version).
- Install the project dependencies with
python -m pip install -r requirements.txt.
- No GPU or API key is needed for the analysis-only workflow.
2. Dataset
- No dataset link yet?
python scripts/make_synthetic_sample.py writes a shape-identical
synthetic sample to ./export/. Build your pipeline on it.
- Real dataset — obtain the current private download link and published
checksum through the official challenge channel. Never store a signed URL in Git.
- Extract into
./export/: mkdir -p export && tar xzf trajectories_v1_01.jsonl.tar.gz -C export/
(scripts read any export/*.jsonl; the archive also contains the LICENSE — read it).
- Challenge use only, no redistribution.
Never commit
export/ or the archives to git (both are in .gitignore — keep them there).
3. Sanity checks
python scripts/load_trajectories.py export/ — prints request counts, per-model
distribution, reconstructed trajectory stats, and a schema check.- Open one line yourself and confirm: exactly
model, input, tools — no output,
no usage. Models are anonymized (families claude-opus/-sonnet/-fable, gpt-5.6-*,
across generations — ~9 ids); PII shows as named placeholders
(<PERSON_ROBERT>, <PROJECT_NAME>, <COMPANY_A>); images are redacted placeholders.
4. First baseline (the honest floor)
python scripts/baseline_router.py export/ — reconstructs trajectories, routes early
small-prompt calls to a configured sibling and prices cache-aware estimated tokens.
Writes results/routes.jsonl.python scripts/plot_frontier.py results/routes.jsonl — frontier CSV (+ PNG if matplotlib).
5. Where the points are
Beating the baseline is expected. The judged depth: a constructed outcome signal (earlier
outputs are recoverable from later inputs), an honest off-policy estimate with named failure
modes, and cache-aware pricing on clearly-labeled estimated tokens. See AGENTS.md → "The twist".
