# Build a clean ExpansionRx nine-endpoint baseline

Implement and execute a small, reproducible RDKit-descriptor + LightGBM baseline on the **released** ExpansionRx split: `train` (5,326 rows) and `test` (2,282 rows). These labels are public; this is retrospective evaluation, not a blind competition entry. The reference `source_manifest` pins the releases and attribution, `data_card` defines endpoint units, and `tutorial` contains the upstream notebook and evaluation code.

Use only `train` labels for fitting, feature preprocessing, model selection and early stopping. Use no external labels, pretrained representations or leaderboard tuning. Freeze one recipe and record its seed before scoring `test`. Train one model per endpoint on available training measurements. Follow the tutorial's descriptor/LightGBM approach; make any change needed for reproducibility explicit. Add a training-median predictor as a comparator. Preserve the original split and every test identifier; report duplicate IDs, canonical-isomeric-SMILES overlap and invalid molecules, including a documented fallback prediction when necessary. Do not silently drop difficult cases or relabel a public test as hidden.

Use the nine named endpoints in `document_schema`. For modeled targets use LogD unchanged and log10(x+1) for the other endpoints (the tutorial's unit offsets are permissible if exactly inverted). Return predictions in the source units. For scoring use LogD unchanged and log10(max(x,0)+1) otherwise, as in the bound evaluator. Report observed-label counts, point MAE and RAE = sum(abs(pred-true))/sum(abs(true-mean(true))) for each endpoint, and the unweighted mean of the nine RAEs. Report null with a reason for undefined denominators; never coerce them to zero. Distinguish point estimates from the upstream bootstrap-mean implementation. Compute fixed-seed 1,000-resample intervals as additional evidence, reporting exactly what interval is used. Include the median baseline under the same evaluation. A good score is not a condition of acceptance.

Submit `document` JSON matching `document_schema`; `predictions` CSV with exactly all 2,282 test IDs, SMILES and all nine source endpoint columns; and a `reproducibility` ZIP containing runnable source, a pinned environment, a one-command rerun, run logs, split/data audit, and a concise report. Commands must use the bound input files without silently fetching changing data. Re-running evaluation on the submitted predictions must reproduce reported point metrics to 1e-8. Document library/platform-dependent training variation separately.

The machine check validates report structure only. Human review checks the actual run, units, complete predictions, train/test separation, source fidelity and reproducibility. This is a complete baseline task, not merely a proposed method. `raw` is provided only for documenting censoring and provenance; do not train on it because it includes held-out molecules. Inspect source_manifest and retain its CC-BY-4.0 attribution. After starting an Attempt, record working findings and end it with a summary.