For Deaf and hard-of-hearing students who sign · ages 5+
Signed intent, confirmed before AI answers.
Every major AI tool assumes text. A student who thinks in ASL must translate their question into written English before they can even ask it. Sign First reverses that: the student signs, sees what the AI understood, fixes what changed, and sends only confirmed meaning.
Signer controls meaningRaw video not retainedStudent outcomes not yet validated
01 / The product in motion
Access moves. Authority stays.
Two concept films explain the product boundaries: the protected bridge and focused repair. Films that depicted generated signing are set aside until Deaf reviewers approve real footage.
01 / Protected bridge
Recognition crosses a boundary without gaining authority.
A replaceable model can propose meaning. The signer controls what becomes an AI prompt.
Concept films explain the intended product model. Signing that appears in generated footage is illustrative and has not been verified as ASL by Deaf reviewers; none of these films are footage of the working recognizer, an ASL-language sample, or evidence of student outcomes.
02 / The missing layer
Recognition is a proposal. Education needs assurance.
A fluent answer can still begin with the wrong question. Sign First lets the learner see, fix, and approve meaning before an AI acts on it.
01Sign
02Recognize
03Detect risk
04Repair
05Confirm
06Route
07Check
08Learn
02 / The complete control path
Eight moves. One protected meaning.
Scroll the system from expression to answer. Every scene is a concept visualization paired with its evidence boundary.
CONCEPT / 01Expression enters
01handspresent
02face + bodypresent
03timesampled
candidate only
CONCEPT / 01 · Expression enters
01 · Expression enters
A recognizer proposes—not decides.
01Sign02Recognize
The camera or another approved input can produce a candidate. Sign First keeps that candidate separate from confirmed meaning, so recognition providers can change without changing who has authority.
Concept visualization · no recognition-accuracy claim
meaning-changing difference
3×5or3+5
hold before answer
CONCEPT / 02 · Risk becomes visible
02 · Risk becomes visible
The smallest important uncertainty gets attention.
03Detect risk
Numbers, operations, negation, and other meaning-changing differences can trigger a pause. The 3 × 5 versus 3 + 5 scene is an illustrative academic-risk pair—not a documented error from Google or another provider.
Illustrative critical pair · not a measured model result
learner repair
multiplyadd
✓approved bythe signer
CONCEPT / 03 · The learner repairs
03 · The learner repairs
Choice replaces a hidden correction.
04Repair05Confirm
The repair surface is designed to become age- and reading-aware while preserving the same rule: the system can suggest, but the learner controls what the question means. Child effectiveness has not yet been validated.
Interface concept · child outcomes not yet validated
confirmed intent packet
3 + 5confirmed
Gemini
Claude
OpenAI
✓ answer checked
CONCEPT / 04 · Confirmed meaning moves
04 · Confirmed meaning moves
AI answers the question the learner approved.
06Route07Check
Only confirmed text crosses the answer boundary. A separate answer-assurance check can then look for drift between that intent and the response while the signer retains a visible override.
Control-path concept · no effectiveness guarantee
governed evidenceoffline only
permissionpurpose-bound
lineageversioned
revocationrebuild required
synthetic proof today
CONCEPT / 05 · Authorized evidence returns
05 · Authorized evidence returns
Learning is a governed choice—not a hidden exhaust stream.
08Learn
Ordinary use stays private. Separately authorized, minimized repair evidence can enter an offline, versioned learning loop and must disappear from rebuilt assets after revocation. The verified loop uses synthetic text repairs; real-learner visual contribution remains disabled.
Synthetic governed-learning lifecycle · not student data or recognition improvement
02 / What compounds
Models change. Learner authority stays.
01
Signer authority
The model proposes. The learner owns the final meaning.
02
Selective assurance
The system steps in when meaning is at risk—not on every question.
03
Provider neutrality
Recognition and language models can change without replacing the trust layer.
04
Governed learning
Everyday use, personal support, and model contribution stay separate choices.
03 / Evidence, honestly
Infrastructure works. Recognition is still the hard part.
The moat is not pretending ASL recognition is solved. It is keeping the learner in control while better recognizers arrive.
Verified
Confirmed-text gate
No answer route accepts unconfirmed recognition.
Verified
No raw-media retention
The product runtime does not persist video or audio.
Demonstrated
Controlled technical journey
Camera-derived candidate to Gemini response on a locked public fixture.
Unvalidated
Reliable student outcome
Deaf-governed, signer-independent evidence is still required.
Checks run quietly with signer override always available.
TARGETTarget
Agentic flow
As natural as typing into an AI—without giving up authority.
TARGET
05 / Clear boundaries
What Sign First is—and is not.
Trust starts with saying the quiet parts clearly.
No. Sign First is an access and assurance layer for talking with AI. It does not replace interpreters in class, in conversation, or in community life.
Not yet. The control path works and can be shown. But reliable recognition and student outcomes are not proven. The site shows that boundary instead of hiding it.
No. The architecture is recognizer-neutral. Better recognition can plug into the same signer-controlled workflow when evidence supports it.
No. Private use is the default. Personalized learning and model contribution are separate choices, and the runtime keeps no raw video or audio.
