# Starrydata thermoelectric ZT frontier audit

This audit tests whether Starrydata can support a monitored experimental thermoelectric-performance series.

## Proposed series

- **Primary:** standing bulk-material ZT frontier at 300 K by paper publication date.
- **Secondary:** standing bulk-material ZT frontier at 700 K.
- **Robust companion:** annual 99th-percentile ZT when at least 20 eligible samples are available.

## Fixed eligibility rule

A curve is eligible when:

- it belongs to the `ThermoelectricMaterials` project;
- `prop_x` is `Temperature` and `unit_x` is `K`;
- `prop_y` is `ZT` and `unit_y` is dimensionless (`-`);
- the joined sample has `DataType = Experiment` and `Form = Bulk`;
- the paper has a valid Starrydata issued date;
- the target temperature lies inside the measured range;
- linear interpolation uses a bracketing gap no larger than 50 K;
- no extrapolation is used.

Values are deduplicated to one value per `(SID, sample_id)` and target temperature. If several eligible curves exist for the same sample, the largest value is retained and its curve identifier remains visible.

## Current audit

The pinned release contains 9,486 thermoelectric papers, 55,198 samples, and 155,885 curves. The strict rule produces 4,672 experimental bulk samples, including 2,295 values at 300 K and 3,036 at 700 K.

The provisional frontiers contain 11 steps at 300 K and 13 steps at 700 K. They remain **under audit**. Every frontier step requires paper-level checking, and the upstream repository does not currently expose an explicit data-license file.

The audit therefore publishes the transformation and derived records without promoting the series to monitoring.

## Files

- `source-manifest.json` — release, source hashes, row counts, and reuse note.
- `sample-values.csv` — all eligible sample-level 300 K and 700 K values.
- `summary.json` — frontiers, annual p99 values, filter counts, outliers, and gate status.
- `audit.py` — deterministic rebuild from the pinned release.
- `check.py` — independent consistency checks.

## Attribution

Source data: Starrydata Project, `ThermoelectricMaterials` release files. Cite Starrydata when using the data. The panel architecture is also informed by METR's “LLMs' Contribution to Discoveries” and the public `tecunningham/ai-discovery-data` repository.