Visual-first screens, large actions, optional read-aloud, and one decision at a time.
The access layer for ASL-first learning
AI starts with text. Sign First starts with the student.
Sign First gives ASL-first learners a signer-controlled path into AI, verifies the academic task and response, then turns separately authorized interactions into governed learning intelligence.
Working research platform · designed for learners ages 5+ · student outcomes not yet validated
FLOW From signing to an AI answer
01 The gap
The AI interface assumes text. Many students begin in ASL.
When access starts with a written-English prompt, an ASL-first learner must cross a language boundary before the AI can help. If meaning changes along the way, a fluent answer can still be the wrong answer.
ExpressionThe student has to translate the question before asking it.
MeaningNumbers, negation, direction, and relationships can change the academic task.
ParticipationPersonalized AI begins only after the student clears a text-first barrier.
02
One platform. Four layers of value.
The camera is the entrance. Structured intent and answer assurance protect the academic task. Connection creates immediate utility. Governed learning intelligence is the compounding asset.
- 01AccessBegin in ASL
The camera becomes a front door to AI for a learner who should not have to author the question in written English first.
- 02AssuranceProtect the meaning and answer
Sign First structures the academic intent, confirms only the details that can change the task, and checks the response separately.
- 03ConnectionUse the right AI
A provider-neutral handoff lets institutions connect confirmed intent to the model or learning tool they approve.
- 04IntelligenceImprove support
Separately authorized interaction signals can help tailor future support and reveal privacy-safe learning patterns.
03 How it works
Academic assurance before and after the language model.
Sign First never asks a student to trust a hidden translation or a fluent-looking answer. It makes proposed meaning visible, confirms the academic task, and checks the returned response separately.
- 01Sign naturally
A student begins with a short ASL question instead of composing an English prompt.
- 02See the candidate
Sign First proposes what it understood and makes uncertainty visible.
- 03Repair the meaning
The student can correct the details that would change the answer.
- 04Confirm the intent
Nothing moves forward until the student approves the meaning.
- 05Reach the chosen AI
Only confirmed intent is handed to an approved AI tutor or assistant.
- 06Check the answer
The response is reviewed for relevance, bounded correctness, grounding, and age-appropriate language.
04
The authority stays constant. The experience grows with the learner.
A beginning reader and a high-school student should not receive the same reading load, number of decisions, or explanation. Both should control what the system sends.
Short guided decisions with concrete language and visible progress.
Independent review with more detail and control over the learning response.
Independent detail without confusing reading level with intelligence.
Implemented, not child-validated: the learner band changes the interaction—not the signer's authority, the question's meaning, or the right to reject the system's candidate. Real child use remains gated.
05
The models will change. The trusted interaction layer compounds.
Sign First does not compete to build the world's largest language model. It owns the high-trust layer where student meaning becomes safe, useful AI interaction.
- 01Signer authority
The learner—not the recognition model—owns the final meaning.
- 02Academic intent and answer assurance
The system protects operations, quantities, units, negation, relationships, and the requested task—then checks the response separately.
- 03Provider neutrality
Recognition and language models can improve or change without replacing the trusted interaction layer.
- 04Governed learning intelligence
Consent, purpose, revocation, and deletion are part of the learning system rather than an afterthought.
- 05Education deployment
Age-aware interaction, school controls, and evidence boundaries turn an AI experiment into institutional infrastructure.
06 The learning flywheel
Access today can become better support tomorrow.
With separate authorization, confirmed interactions can create a longitudinal picture of concepts explored, support requested, and corrections made—without turning ordinary use into silent model training.
- 01Student confirms meaning
The trusted interaction creates better evidence than an unreviewed transcript.
- 02AI supports the task
The institution's approved model responds to the question the student actually authorized.
- 03Learner authorizes support
Useful context can inform a later interaction without making surveillance the price of access.
- 04Patterns improve the experience
Governed repair and learning signals can strengthen support, assurance, and institution insight.
The visual-data moat has its own gate: Sign First now enforces a synthetic/public research metadata, split, retention, and revocation contract. It does not collect participant video, and ordinary student use never opts anyone into visual-model training.
07
Built far enough to demonstrate. Honest enough to show the gap.
The strongest product story is not that recognition is solved. It is that Sign First remains safe and useful while recognition is imperfect—and can improve as better models become available.
Camera input, visible candidate, targeted repair, structured confirmation, AI handoff, and bounded answer checks run as one controlled experience.
Controlled research showed that general-purpose models can lose critical academic meaning—validating the need for assurance.
The product boundary remains useful when a recognizer is weak and is designed to benefit when stronger models arrive.
Private use, personalization, institutional insight, and future research contribution remain separate choices.
Current boundary: reliable open-domain ASL recognition and real-student outcomes have not been validated. Those claims require Deaf-led, governed evaluation.
08
The direct shortcut failed. The assurance layer preserved a safe path.
Recognition alone did not protect the academic question. Sign First adds a stop point where the learner can reject or repair meaning before a fluent answer makes the error harder to see.
| Path | Critical meaning retained | Human repair | What the evidence says |
|---|---|---|---|
| Typed English control | 91 of 91 | Not measured | Exact reference text; an oracle control, not observed student typing. |
| Gemini direct from video | 0 of 91 | None | Measured on 20 locked clips; it abstained on 0 despite losing every critical term. |
| Sign First before repair | 9 of 91 | Required | The current continuous-ASL candidate is not student-ready and averaged 16.95 word edits per item. |
| Sign First after correct repair | 91 of 91 conditional | Required | Protocol simulation only: the signer repairs to the locked reference before anything is routed. |
SupportedSigner-controlled assurance is necessary before routing.
RejectedDirect Gemini video is not a safe replacement for this workflow.
Still openWhether Sign First is easier, faster, or better for students than typing.
09 Path to scale
A focused wedge into a much larger learning platform.
Begin with a high-value access problem schools can see. Expand through integrations and governed intelligence without forcing institutions to replace the AI tools they already choose.
- EnterASL-first AI access
Give schools a controlled front door to the AI tools already entering classrooms.
- ConnectInstitution-approved models
Route confirmed intent through a consistent trust boundary instead of locking into one provider.
- CompoundAuthorized learner intelligence
Use longitudinal context to personalize support and produce privacy-safe institution insight.
- ExtendA reusable assurance layer
Bring the same signer-controlled pathway to more learning tools, programs, and partners.
A more equal starting point for AI
Let the student sign first. Let the AI answer second.
See how a signed question becomes confirmed intent, reaches an approved AI, and creates the foundation for governed learning intelligence.