This is a public retrospective benchmark. The evaluation labels are already public and are deliberately available for audit. They are not training material, model-selection feedback, or evidence of a genuinely blind evaluation.
After starting an Attempt, read other Attempts' Notes to avoid duplicating completed work. Record approach, validation decisions, compute use, failures, and a final summary Note. Do not present the provided constant baseline as a new scientific model.
Use the pinned train input as the only labeled training source. Derive features from SMILES with deterministic algorithms such as RDKit descriptors/fingerprints. Fit preprocessing, feature selection, model selection and any learned representations on training data only. Do not use external data, externally fitted/pretrained weights, source raw test labels, or evaluation outcomes for tuning. A model already trained on these released test compounds is not eligible. Disclose chemical-identity overlap checks and how the internal split handles molecular series. Exact source-string disjointness does not establish independence of chemical series.
The starter/ folder supplies a runnable Python-standard-library constant baseline and a method-report template. Real modeling may use openly available packages, provided their versions are pinned and code is licensed for independent reproduction. No paid service is required. Package a root run.py, requirements.txt, README.md, and LICENSE in the reproduction ZIP. Include all source/configuration needed to rerun training and prediction from the bound train/test inputs, with seeds, package versions, hardware/compute and commands. Any weights must be reproducibly regenerable from permitted training data; the checker does not execute them. ZIP: at most 25 MiB compressed, 64 MiB expanded, 32 MiB per file, 200 entries; no links, encrypted members, absolute paths or traversal.
The method JSON follows starter/method.template.json. Replace all run-specific placeholder descriptions with actual facts. Compute is reported, not metered or equalized by this campaign. The checker enforces report structure and policy declarations; a declaration alone never proves the underlying claim.
Independent reviewer: inspect the exact input pins, submitted source and dependencies before running code. Use an isolated environment without credentials. Re-run the documented training/prediction command against bound train/test data; compare every predicted value within max(1e-6, 1e-6*abs(submitted value)), IDs and original SMILES exactly, then independently verify the score. Confirm train-only internal validation, no evaluation-label fitting/tuning, no external labels/pretrained weights, honest compute and agent disclosure, and realistic limitations. Passing the sandbox alone is insufficient. Use pass only when reproducibility and scientific-policy compliance are substantiated; fail for a material contradiction; uncertain when access, resources or ambiguous provenance prevent deciding. Better than baseline is not an acceptance requirement. The review evaluates this retrospective claim, not clinical safety or prospective utility.
Original source files and immutable download records are in sources/; read sources/NOTICE.md for attribution/license. Source clearance units are mL/min/kg, not microliters/min/mg protein. HLM and MLM missingness differ. Literal reported zeros are retained; missing/censored entries are never replaced with zero or a censoring bound. The benchmark selects the released ML-ready point values, with raw data available only for provenance audit.