What a failed model teaches

Rejected candidates narrow specific claims without closing the broader recognition program.

98.44% → 0 / 23

What changed the product

Official-shard top-1 accuracy did not transfer to critical continuous-ASL terms.

Decision trace

Can a benchmark score travel into continuous academic signing?

ASL-Citizen checkpoint · official shard + locked continuous holdout

Reproduced

The checkpoint was real.

The public model repeated its reported behavior on its own official distribution.

A model can load correctly, repeat deterministically, and still fail the product claim. Those facts are worth preserving.

One checkpoint reproduced 98.44% top-1 on its own official shard and then recovered none of 23 predeclared critical tokens on continuous-ASL holdout clips. The model was real; the transfer assumption was wrong.

Sign First locks development and diagnostic evidence before evaluation, records the exact model and preprocessing path, and separates own-distribution performance from continuous-ASL utility.

A near-miss may identify useful representation signal. It cannot be tuned on the opened diagnostic and then presented as independent evidence.