Engineering notebook

Dataset provenance and reproducibility

Keep inputs, labels, recipes and evaluation evidence connected.

01

The dataset bundle

A reproducible release includes the dataset, ground truth, generation manifest and checksums. Record both archive integrity and the checksum-index digest: they answer different questions about the bytes being distributed.

The archive hash identifies the packaged object. The checksum index supports verification of the files inside it. Neither alone proves annotation correctness or physical realism.

02

The evaluation bundle

Each accepted sample links an immutable input ID to a provider attempt, raw response, normalized text, scored result and durable evidence record. Retain the model configuration and metric-policy version alongside the run.

03

Version without rewriting history

The public benchmark identifies its source leaderboard and frozen-truth digests. New v1.0 dataset artifacts do not retroactively replace the v0.1 evaluation inputs.

Build with evidence

See what breaks
before production does.

Run StressBench against your document model and get a condition-level robustness readout.