what this instrument refuses to say

one sentence appears in all three committed receipts, verbatim: a tighter guarantee than your data supports is not a guarantee. it is not an error message. it is the product working: where the data can carry a commitment, the receipt states one, and where it cannot, the receipt says so in words a committee can read.

a finite-sample false-positive guarantee at level alpha needs a minimum number of human calibration samples, and the requirement is arithmetic, not policy. all three committed receipts hit it at alpha=.005, which needs at least 199: the reference corpus brought 100, the binoculars run the same 100, the raid slice 120. three different piles, one identical refusal line, quoted on the receipt-anatomy page. the alternative, printing a threshold anyway and letting the reader assume the guarantee, is how confident nonsense gets made.

the front page’s gaige plan card is a table of refusals with the measured numbers attached. on the bench that generated it, with a resident scoring daemon holding most of the gpu: falcon-7b in 4-bit, no, needs 8.0 GB free and the machine had 7.9. falcon-7b at fp16, no, needs 13.7. the binoculars pair, no, needs 9.0. what fits ran, and carried its measured pace. the card keeps no quality column, deliberately: separation lives in receipts, never in a hardware table, and a plan that promised quality would be a different kind of document.

every load is verified before a single score is emitted: the receipt records how many modules actually loaded at the requested precision, what stayed resident, and whether any fallback occurred. a run that silently fell back to a different precision would be a different instrument wearing the same name, so gaige refuses the silent version; the fallback field in the fingerprint is there to prove a no happened, not to soften a yes.

gaige admit vets candidate material against a trusted baseline: divergence with intervals, stratified where it differs, refusal floors when a candidate slice is too small to support the conclusion you want. and its strongest refusal is permanent: the receipt never says admit or reject. the measurement is gaige’s; the decision stays yours. the custody page walks the five-step practice this verb belongs to.

gaige verify sweeps a file for provenance evidence, c2pa credentials and watermarks, as typed statuses rather than scores. its refusal discipline runs one level deeper: absence only counts when a per-image self-test proves the carrier could have held the mark. an absence that was never testable is not evidence of anything, and the tool declines to pretend otherwise.

none of this is a paywall in costume. the standing promise on the roadmap holds here: anything that surfaces an error, an interval, or a caveat is free forever. a refusal is exactly that class of information. it will never be rate-limited, upsold, or softened for a paying tier, because a bench you can pay to stop saying no is not a measurement instrument.

this is the last step of the ground-up track: what you measure on (corpus), what unmeasured costs (why gaige), how a score becomes a decision (thresholds), who wrong decisions land on (false positives), what an honest record looks like (receipt-anatomy), and, here, the noes that keep all of it honest. the guides shelf holds the whole set, dated.