# Source and method references

The desk executes no model, fitting, calibration or external service. Formula references were read through official primary pages on 2026-11-04. No research text, figures or implementation code is bundled.

- [Scikit-learn Brier documentation](https://scikit-learn.org/stable/modules/generated/sklearn.metrics.brier_score_loss.html): mean sum over classes. This desk always retains the unhalved multiclass convention, including binary data; it does not infer alphabetic column order.
- [Scikit-learn log-loss documentation](https://scikit-learn.org/stable/modules/generated/sklearn.metrics.log_loss.html): natural-log loss. That helper clips to dtype epsilon. This desk deliberately does not clip: a zero gold probability makes mean NLL infinite, serialized null with an explicit flag.
- Guo, Pleiss, Sun and Weinberger (2017), [On Calibration of Modern Neural Networks](https://proceedings.mlr.press/v70/guo17a.html), ICML/PMLR70:1321–1330; [official arXiv HTML v2](https://arxiv.org/html/1706.04599v2). ECE is the sample-weighted absolute gap between bin accuracy and mean confidence. This desk fixes ten left-closed floor bins with1 in the last bin and lowest-taxonomy-index argmax ties. This is an explicitly specified descriptive statistic, not an exact reproduction of all paper bin choices or a per-case calibration guarantee. NLL is averaged over probability support rather than reported as total loss.

The known-result demonstration is from the accepted `work:tiny-hash-intent:v1` original package, manifest SHA256 `8879952142cf92415b7c052bdf1785954ef2f6bcc22c3952f6461326c76ba284`. [Reference map](reference/REFERENCE.json) preserves the exact subset bytes and renamed paths. `reference/original-MANIFEST.json`, `reference/original-license-scope.md` and `reference/original-NOTICE` are historical source records describing that original full package, not a manifest/license path inventory for this subset. The current release inventory is root MANIFEST.json.

BANKING77: PolyAI; Casanueva, Temcinas, Gerz, Henderson and Vulić (2020), [Efficient Intent Detection with Dual Sentence Encoders](https://arxiv.org/abs/2003.04807); [original repository](https://github.com/PolyAI-LDN/task-specific-datasets). Exact data commit `9d081458ff52e53cf7e848f414e6e9344e4e6696`, license commit `3bf93de788b9362c34d06cba1d271bc23dd1bcb6`, original license CC BY4.0. [Frozen source metadata](reference/banking77-source.json), [license](reference/LICENSE-CC-BY-4.0).

`make_demo.py` converts saved int8 logits through stable softmax at the original frozen temperature0.75. All3080 original test rows and ordered77 labels remain; argmax/confidence/gold probability/NLL/Brier are checked against the original CSV, maximum discrepancy below2e-15 except Brier below7e-16. No model or new predictions are produced. Earlier work already inspected this benchmark; seven normalized train/test overlaps remain, so this is neither blind data nor independent evidence of generalization. Hard-label metrics and probability metrics are recalculated from the same saved result.

Original synthetic fault fixture is authored by Renchao Wu, with intentionally meaningless text/labels and malformed lines, not real user data or adoption. Its ECE hand arithmetic typo was corrected before core implementation; original contract freeze plus explicit erratum retained.
