the roadmap is a queue, not a promise: a thing ships when its receipts exist.

gaige is released and operating: the package installs from pypi today and the committed receipts on this site are its output. this page is the queue behind the released tool.

shipped, dated

newest first. every date is the day the receipts went public.

drift observatory2026-08-04 · live

live at gaige.dev/observatory/, rebuilt nightly from the committed exports: fingerprint, interval, and refusal state beside every number shown, so the site holds itself to the standard it sells.

gaige admit2026-07-30 · 0.0.4

the custody flow in one command. point it at the baseline you already trust and a pile of new material, and the receipt says how far the new material sits from what you accepted, with intervals, the strata where it differs, and floors that refuse when the evidence is thin. it never says admit or reject; that part stays yours.

gaige verify2026-07-30 · 0.0.3

not a scorer but a provenance evidence sweep: c2pa credentials with typed absence, the dwtdct watermark behind a per-image carrier self-test, keyed schemes as NEEDS_KEYS. statuses, never a score.

gaige export + the receipts gallery2026-07-29 · 0.0.2

every measurement leaves as a committed artifact, and the receipts gallery renders the committed exports this site is built from.

public release2026-07-27 · 0.0.1

pypi package, public repository, and this site, all in one day; the first committed receipts shipped with them.

the shape of more

gaige grows the way it measures. an entry below graduates when it can show the run that proves it, the test that refuses dishonestly, the documented edge, and not before. dates on shipped entries are receipts. the queue carries no dates on purpose: a date on the future is a claim, and this project only makes claims it has measured. the queue is ordered by what the work needs, not by what would sell.

the queue takes outside input. bring improvement recommendations, measured use cases, and sightings of gaige cited in the wild to the Discussions page: threads get triaged, and what survives lands below with a receipt attached.

queued

these land in the order their receipts arrive.

Mondrian conformalper-subgroup guarantees

group-adaptive operating points carrying the same finite-sample honesty: needs ≥99 human samples per bucket at α=.01, and the data decides whether you get one.

batched scoringbenchmark scale

full-benchmark runs without the single-text loop; batch size and padding enter the fingerprint, because they change the numbers.

synthid-text receiptswatermark lane

watermark receipts for synthid-text (detection is per-configuration, and the thresholds are delegated to deployers: calibration homework), plus revocation and trust-list state in verify receipts, and c2pa-in-text.

degradation receiptsmeasured, not invented

published robustness panels scored under our own instruments: how much detection degrades under paraphrase and rewrite attacks, with intervals, on attack-stratified benchmark slices.

bigger instrumentsfull re-baseline

larger reference models and wider corpora go through a full re-baseline when they land, because a changed instrument that kept its old numbers would be lying.

what money never changes

the free floor. anything that surfaces an error, an interval, or a caveat stays free, forever; that is the point of the tool, not a tier.
the license. AGPL-3.0, with commercial licensing for those who need it; the code stays open either way.
the refusals. no verdicts, no universal thresholds, no number without its instrument, at any price.
the order of proof. credibility features are never paid.

adjacent: support · the package · the receipts gallery