Working research, shown honestly. See what is verified and what is not →Working research · inspect evidence →
Sign First

Research

Progress includes what failed.

The research program tests public recognition candidates, locks evaluation sets before scoring, preserves near-misses, and refuses to convert fluent output into evidence of correct ASL meaning.

06 / Evidence discipline

A rejected model can still teach us where the next useful signal may live.

Research is a sequence of decisions

Open the program.
Keep the dead ends.

The model search matters, but the accumulated decision record is the differentiator: what transferred, what broke, what changed the product, and what cannot yet be claimed.

Not student-ready

The sentence models found language signal—and lost the academic task.

Across public ASL sentence diagnostics, the useful result was not a translator. It was reproducible proof that critical meaning can disappear behind fluent-looking text.
26.09%best literal critical-token recall on the first locked Uni-Sign diagnostic
147.71%mean word error on that six-clip diagnostic
0%critical-token recall from the earlier Qwen research bridge
What entered the product
The candidate must remain visibly provisional. Repair cannot be a hidden cleanup step after the model has already decided what the signer meant.
Next evidence gate
A future recognizer must pass signer-independent, academically critical, Deaf-adjudicated evidence before confirmation can become selective.
Source program
Uni-Sign and Qwen locked public-ASL diagnostics
Circular technical loop with checkpoints and a versioned center
Learning is a governed technical loop, not a silent by-product of use.Synthetic loop concept; no learner-effectiveness claim.
01

Recognition

Independent evidence before integration.

Continuous-ASL translation, isolated-sign specialists, fingerspelling, pose cues, and compound candidates are evaluated against fixed evidence and claim-specific gates.

02

Assurance

Calibration matters more than raw confidence.

A recognizer confidence score is not treated as student safety. The product needs held-out evidence showing when to allow flow, when to ask, and when to abstain.

03

Learning

Train, evaluate, promote—or reject.

The governed learning loop has been demonstrated with synthetic text repair assets. It is offline, versioned, revocation-aware, and not evidence that visual ASL recognition improves.

From model search to a disciplined portfolio

Every failure narrows the next question.

Rejected candidates stay useful when their task, split, preprocessing, result, and limitation are preserved. The program advances by changing the claim—not grading on hope.
01

Measured now

Every claim on this page traces to a locked, dated artifact.
02

Governed next

New evidence enters only through consent, provenance, and revocation gates.
03

Evidence required

Larger claims wait for independent, Deaf-governed evaluation.

Continue exploring

Separate verified controls from open research.

Evidence→Data governance→Insights→