Perform a source-grounded curve-quality review for exactly the 16 real Molecule Name values in selection. They span reported pEC50 uncertainty by a deterministic rank-based selection; this is a diagnostic sample, not a population-prevalence study. curves contains every matching curve row and its immutable zero-based row locator; original_curves contains the unchanged source parquet. counter_metadata supplies available PXR-null fit metadata, while selection includes primary-assay fitted estimates and uncertainty. Read data_notes, data_card and source_manifest before interpreting them.
For every selected compound, plot every concentration/signal array, with source-row labels and both normalized and raw-signal views. Inspect missing/nonfinite values, array lengths versus npoints, concentration coverage, direction/shape, response range, reversals/outliers, reported pEC50 versus observed concentration range, and reported estimate/CI/standard-error consistency. Report quantitative evidence and source locators for each flag. Flagging is triage, not proof of biological inactivity or measurement error; no defect is a valid outcome.
The released parquet has no assay/channel, replicate or plate/well identifier. Do not assign primary versus counter-assay or biological replicate identity solely from row order, brightness or expected behavior. Keep unknown identities explicit. If you offer a supported assignment inference, state assumptions, competing explanations, and what source metadata would settle it. The available counter-assay fits do not alone identify an unlabeled array. Do not manufacture per-well errors or derive confidence intervals without explaining their unavailable replicate structure. Optional refits must be marked sensitivity analyses with an explicit model and assumptions; they do not replace the source's published labels.
Deliver one record for each selected compound, preserving all source curve row locators. Distinguish observed facts from inferences, propose retain/review/additional-metadata actions, and discuss how any recommended exclusion or weighting would affect a later model without making unsupported performance claims. In particular, quantify a proposed rule's effect on these 16 cases, not on the whole training population. No fixed number of suspicious curves is expected.
Submit document JSON matching document_schema and a reproducibility ZIP with scripts, pinned plotting/data environment, exact run command/logs, all annotated plots, per-curve numeric summaries and a concise audit report. The arrays and selected metadata are real public assay data, not simulated fixtures. The machine check verifies only report structure; human review checks complete coverage, faithful plotting, source joins, justified flags, and restraint about unlabeled channels. Retain CC-BY-4.0 attribution. After starting an Attempt, record working findings and finish with a summary.
The training_metadata and selection_script inputs let you independently reproduce the selected IDs and curve rows. Save selection_script as Python or run it directly with Python; its two arguments are the reference directory and a separate output directory.