the reference receipt, read line by line

the record below was generated 2026-07-22 by gaige 0.0.1, measuring fast-detect-gpt on a pinned 200-text corpus. it has been committed, byte-stable, ever since: the receipts page renders it, this site’s front page quotes it, and the raw JSON is one click away. nothing on this page paraphrases; every quoted line is the record’s own.

five lines, and each one is a commitment. the name states the pile and the draw. the hash makes “same corpus” checkable forever. the counts state what each class contributed. the source and raw hash pin even the upstream bytes before filtering. and the note is the pile’s honest edge, recorded where nobody can miss it. the corpus page teaches this block from zero.

the load is verified, not assumed: the receipt records that 128 modules actually loaded in 4-bit and how much memory the model held resident, because a silent full-precision fallback would be a different instrument wearing the same name. the versions matter for the same reason. change any of these and the numbers can move; the fingerprint is how you know they have not.

AUROC summarizes how well the scorer separates the two classes across all possible thresholds: 1.0 is perfect separation, 0.5 is coin-flipping. the interval beside it is not decoration. it says that with 200 samples, the measurement itself is only known to within about five points, and the receipt prints that width instead of rounding it away.

the empirical lines say what happened on this pile, in-sample. the conformal lines carry a finite-sample guarantee on the false-positive rate (arXiv 2505.05084 is the method), bought at the price of sensitivity: compare the 1% empirical line’s 86% with the α=.01 conformal line’s 76%. neither family is “the” threshold. thresholds works the whole trade, including the base-rate arithmetic that decides how many flags are wrong.

one line of the record is a guarantee that does not exist, stated as such. the sentence it carries in full, quoted on the corpus page and collected with every other refusal on refusals, ends: a tighter guarantee than your data supports is not a guarantee. a receipt that only ever said yes would be an advertisement.

these lines are why the record outranks a screenshot. anyone with the command, the pinned corpus, and the fingerprinted environment can regenerate the analysis and compare byte-for-byte; the replay check does exactly that. trust here is not a request. it is a procedure.

the committed binoculars receipt has identical anatomy with one difference worth seeing: its instrument block records two models (an observer and a performer), each with its own verified quantization and residency, because a two-model score with a one-model fingerprint would be lying by omission. every receipt gaige emits, whatever the detector, carries this same skeleton: corpus, instrument, result, thresholds, refusals where honesty requires them, and the command that lets you check it all yourself.

this is step 5 of the ground-up track. the receipts page renders the committed records live; false positives takes the threshold lines into the arithmetic of who gets flagged; and the raw JSON behind this page is at data/receipts/20260722-163959-fast-detect-gpt.json, unchanged since the day it was generated.