short answers, receipts attached

no. the confusion is understandable because both live in the same room, but a detector reads an essay and leans human or machine, and gaige stands behind the detector with a clipboard. run it over a labeled corpus you chose and it writes down what the detector did there, which is usually messier than the marketing: the AUROC came out here, the interval is this wide, the threshold you would need for a 1% false-positive rate sits there, short texts misfired more than long ones did. Fast-DetectGPT and Binoculars ship as scorers it knows how to hold the clipboard for. what never comes out of it is a ruling on any document, and the trademark terms are written so nobody can borrow the name to fake one.

it can tell you what a calibrated scorer says about one document, with the error rates you measured riding along: that is gaige score, and it emits measurements only. what it will not do is convert that measurement into a verdict about a person or a paper. the report’s base-rate arithmetic is there to show how fast flags go wrong in mostly-human pools, before anyone acts on one.

because thresholds do not transfer. the RAID benchmark study (ACL 2024) measured open-source detectors at their default thresholds producing 47 to 100% false-positive rates, whilst commercial tools ship factory-calibrated at 1.7% or less. its own recommendation is to calibrate on in-domain data before use, and that sentence is this project’s thesis. the evidence section carries the receipts.

open one of the JSON files under data/ and the shape answers the question. the AUROC is in there, but so is the model that produced it, the quantization as it was verified at load time, the device, the library versions, the sha256 of the corpus, and one full command that rebuilds the run from scratch. that bundle is the receipt: a statistic that cannot be quoted apart from its instrument. the receipts page just renders those committed files in place, which is why nothing on it is typed in by hand.

start with what the ordinary number cannot do. measure a threshold on 100 texts and you learn what happened on those 100 texts, full stop. the conformal construction takes the same human calibration scores, sorts them, and picks an order statistic, and that mechanical little move is what earns a guarantee with actual reach: a finite-sample, marginal bound that holds beyond the sample, assuming exchangeability. the bill arrives as bluntness. the committed reference receipt shows α=.005 refused outright, because 100 human texts cannot support it and 199 could, and gaige would rather name the shortfall than invent the guarantee.

yes, by design. scoring runs on local models once the weights are staged; gaige analyze re-derives every number from stored scores with no model, no GPU, and no network, bit-identically; receipts are plain files that cross an air gap. pip install gaige phones nowhere, carries no telemetry, and disconnected operation is a documented mode, not an accident.

nothing, and the honest parts never will. the license is AGPL-3.0, and anything that surfaces an error, an interval, or a caveat stays free forever, on principle. commercial licensing exists for organizations that need different terms, and a support page exists for fuel. neither changes what any receipt says.

Fast-DetectGPT and Binoculars ship as scorers, each with its own guide: fast-detectgpt in practice and binoculars in practice. a watermark-class verifier is queued on the roadmap. the scorer interface is pluggable on purpose: anything that maps text to a score can be calibrated, compared, and watched.