an instrument is a promise: calibrate first, trust second.

a scale in a shipping bay carries a calibration sticker: who checked it, against what reference, when it is due again. the culture that puts stickers on scales is called metrology, and it exists because a measurement without a calibration story is just a confident number.

an AI-text detector is an instrument in exactly that sense. it takes material in and puts a number out, the number moves when the model, the quantization, or the corpus moves, and decisions get made on it. what detectors mostly do not have is the sticker. ai metrology, in plain words: calibration and receipts you can verify, applied to the tools that measure AI. gaige is the instrumentation layer that produces the sticker: it measures the measurers, on your material, and writes down everything a skeptic would ask for.

the word is a commitment, not a metaphor. numbers here ship with error bars, a fingerprint, and stated refusal conditions, or they do not ship.

a detector answers a question about a document: this text, human or machine. the gaige instrument answers a question about the detector: on material like yours, how often is that answer wrong, in which direction, at which threshold. the two get confused because they share a room, and the confusion is expensive in exactly one direction: a detector trusted without measurement.

a metrologist would put it this way: the reference bath does not tell you tomorrow’s weather. it tells you whether your thermometer can. gaige holds adapters for detectors (fast-detect-gpt and binoculars today) and runs them, but what comes out of a run is never a ruling on any text; it is the detector’s error rates on a stated corpus, conditions attached. the faq keeps the short version, and the case study reads a reported $1.8M overrun as exactly this confusion, priced.

an AUROC from march and an AUROC from today only compare if the same instrument produced both. that is the fingerprint’s whole job. each receipt writes down the model, the quantization as checked at load (not what the config claimed), which device actually ran, the library versions, the sha256 of the corpus, and the command that reruns everything. series are stricter still: once registered, a series will not take runs from a second instrument. swap a library version and gaige calls that a new instrument, not drift.

the committed receipts on the receipts page show the practice: fast-detect-gpt on hc3-mini at AUROC 0.9720 means that number, under that fingerprint, reproducible by that command, with the 95% interval (0.9458–0.9938) stated, because a point estimate without an interval is another confident number.

calibrating the instrument means running the detector you already use on material like yours, and deriving the operating point from what actually happened. gaige states thresholds two ways: the empirical threshold measured on your labeled set, and a split-conformal threshold with a finite-sample false-positive guarantee. the thresholds page works the arithmetic end to end, including the base-rate step most deployments skip, the one that decides how many of your flags are wrong.

aggregates get the same treatment: a falcon-7b run on a RAID slice showed one threshold catching 87.6% of greedy-decoded machine text while catching 39.7% of sampled text. the aggregate AUROC hid that split; the stratified receipt did not. calibration that stops at the headline number is not calibration.

a statistic on a receipt never travels alone. the headline AUROC above carries its bootstrap interval, and the practice continues all the way down the record: every threshold row states its own, so the reference receipt’s 91.0% catch rate at the 5% operating point is written as 85.0% to 96.0%, which is what one hundred texts actually earn you. the interval is not decoration. it is the difference between what got measured and what got hoped.

and where the data cannot support the ask at all, the receipt prints a refusal in the place a number would go:

alpha=0.005 needs >= 199 human calibration samples, got 100.
A tighter guarantee than your data supports is not a guarantee.

that line is receipt-verbatim. an instrument that only ever answers is not measuring, it is agreeing; the refusals page collects every no gaige knows how to say, with the measured trigger for each.

the last field of every receipt is a command. this is the one under the reference run:

$ gaige run --corpus hc3-mini --n 100 --seed 17 \
    --detector fast-detect-gpt --model tiiuae/falcon-7b \
    --quant 4bit --device cuda --max-tokens 1024

run it against the same corpus bytes (the sha256 in the fingerprint pins them) on the same instrument, and the numbers rebuild. that one line changes what kind of document a receipt is. without it, a receipt is testimony you decide to believe; with it, a receipt is a claim anyone can check.

the receipts page carries the demonstration: nineteen days after the reference run, the same command on the same instrument produced AUROC 0.9720 a second time, and both records are committed with their run ids.

reproduction does need the same instrument: the same weights, the same quantization, the versions the fingerprint recorded. that is not a loophole. it is what the fingerprint is for, and it means a rerun that lands somewhere else tells you which part moved rather than leaving you to shrug. receipt anatomy walks the whole record line by line, reproduce command included.

calibration has a shelf life. the model gets re-quantized, a library updates, the material you score changes shape under you. the fix is boring, and boring is the point: put the detector on a maintenance schedule the way a lab does with its equipment. register runs as a series, re-measure on a cadence, and let the alarm carry an interval so one noisy run cannot cry wolf. the mechanics live on the drift page.

gaige measures scoring processes. it never pronounces a document human or AI, never ships a universal threshold, and never claims a detector is accurate or bias-free; those claims belong to whoever makes them, with a receipt or without one. the terms page holds this as a condition of use, and the certificate states the full scope. that boundary is what makes the metrology honest: the instrument that checks instruments does not moonlight as a judge.