gaige
package v0.0.4 · released

the pages that go deeper than a card

detector guides with committed numbers, the threshold arithmetic worked end to end, drift practice, and dated regulatory reading. every entry opens a page whose numbers carry their own receipts.

start anywhere: each page stands alone, carries its own receipt line, and links its neighbors. the shelf grows when a question keeps arriving or a case earns a writeup; the date on each row says when a page last earned an update. new to the bench? the ground-up track below reads in order.

the ground-up track

read in order and the whole argument assembles: what goes wrong when nobody measures, and how a score becomes a decision someone can defend. the track fills in as new pages land; every step that exists is listed here, dated, and none of it assumes you arrive knowing the vocabulary.

1 · corpusthe pile every number stands on · 2026-08-07

what a corpus is, how to build one you can vouch for, and what happens to a detector’s numbers when the pile changes. starts from zero; every number carries its receipt.

2 · why gaige, a case studymeasurement debt, priced · 2026-08-04

a reported $1.8M overrun on a menial AI task, read as a measurement story: five joints where it failed, and the boring counters that catch this failure class. citations carried; no invoice monitoring claimed.

3 · thresholds, workedoperating points · 2026-08-07

empirical vs conformal operating points and the base-rate arithmetic that decides how many flags are wrong, worked from committed receipts. builds on the case study: this is the arithmetic that was missing there.

4 · false positivesthe arithmetic of a wrong flag · 2026-08-07

false-positive rates against base rates: why most flags can land on humans even when the rate sounds small, worked from cited studies and committed receipts. six questions any flag should answer.

5 · receipt anatomythe reference receipt, read line by line · 2026-08-04

every field of a committed receipt explained from zero: corpus block, instrument block, result, two threshold families, the refusal line, and the reproduce command a stranger can run.

6 · refusalsthe honest no, collected · 2026-08-17

every refusal gaige ships, with the measured line that triggers each: conformal floor, memory floor, quantization gate, admission floors, the absence rule. never priced.

detector guides

one page per scorer the bench ships, each holding the published figures and an independent measurement in the same frame.

fast-detectgpt in practicedetector guide · 2026-08-04

the shipped scorer with committed numbers: what the published defaults do, what calibration buys, and where quantization moves the answer. its operating points are worked in full on thresholds.

binoculars in practicedetector guide · 2026-08-04

the authors’ figures beside an independent measurement, and why a global threshold does not transfer. both numbers belong in the same sentence. reads beside fast-detectgpt: same bench, different instrument shape.

running the bench

the practice pages: what the instrument commits to, how the measuring corpus stays yours, what watches for drift, and the dated regulatory reading.

the instrumentai metrology in practice · 2026-08-12

what the word in every footer commits gaige to: calibration before trust, an instrument fingerprint on every number, a drift cadence, and refusal floors where the data cannot support the ask. the concept page for the whole bench.

custodythe corpus stays yours · 2026-08-04

chain of custody for the material you measure on: pin it, hash it, vet additions with admit, and prove at any date which bytes every number came from.

drift practiceseries monitoring · 2026-08-17

instrument drift monitoring: registered series, replicate variance floors, per-interval conformal alarm bounds.

drift observatorythe committed receipts, rendered live · 2026-08-04

every committed receipt on one static page, rebuilt nightly as a pure function of the export directory: aggregates with the slices they hide one toggle away, refusals rendered as refusals, zero javascript. when the exports are inconsistent it refuses to render at all.

what changed on august 2regulatory dates, maintained · verified 2026-08-03

the eu ai act article 50 marking obligations and california ab 853, read against the primary sources: a maintained date table, and what detectable and reliable look like as evidence rather than adjectives. dated, linked, and not legal advice.

adjacent: the package · receipts · faq · about