gaige
ink
v0.0.1 · RELEASED

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

What support changes is pace. What it never changes is the bar, or what stays free.

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 are absent on purpose: a date is a claim about the future, 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.

SHIPS REGARDLESS

These land with or without support, slowly, in the order their receipts arrive.

gaige admitone-verb admission

Candidate material against your trusted baseline in one command: divergence with intervals, stratified where-it-differs, refusal floors. The custody flow, folded into a single receipt.

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.

Watermark verifierscorer #3

Watermark-class verification as another pluggable scorer beside Fast-DetectGPT and Binoculars. A prototype exists in the lab; it lands when its receipts do.

Degradation receiptsmeasured, not invented

Published humanization-attack panels scored under our own instruments: how much each attack actually degrades detection, with intervals, on attack-stratified benchmark slices.

Drift observatoryreceipts in public

A live page rendering registered series: fingerprint, interval, and alarm state beside every number shown, so the site holds itself to the standard it sells.

Bigger instrumentsas resources allow

When funding buys GPU time, it buys larger reference models and wider corpora, and each upgrade goes through a full re-baseline, because a changed instrument that kept its old numbers would be lying.

WHAT FUEL BUYS

Support converts to resources, and every resource has a name. Nothing here is a feature for sale: it is the same queue, moving at a pace the hardware no longer sets.

GPU HOURS

The watermark verifier and the batched runner both wait on measured runs. Compute is the slow half of every receipt on this site; fuel buys the hours that produce them.

CORPORA

Broader material beyond the current benchmark families (more domains, more registers, more length regimes), so calibration advice rests on more than one corpus lineage.

MORE INSTRUMENTS

More detector integrations, landing the way the first two did: the load proves its quantization before a single score comes out, the honest refusals stay wired in, and nothing ships ahead of its receipts.

TURNAROUND

Issues answered sooner, edges documented sooner, the queue above simply moving at the pace attention allows.

WITH SERIOUS RESOURCES

The audit is the original use case, not an add-on: gaige exists because somebody needed to check an instrument before trusting it with consequences. If your organization needs that check done rather than a tool to do it with, write to [email protected] with the question and the constraints. The engagement gets scoped on paper, and the deliverable is a set of receipts that rerun on your machines, so the conclusion never has to lean on our word.

WHAT MONEY NEVER CHANGES

FUEL

If gaige saved you from acting on an uncalibrated number, a coffee’s worth genuinely helps: it becomes the GPU hours and corpora named above. The link, and the plain accounting of what it buys, live on the SUPPORT page.