Safety

Confidence is not the same as correctness.

Sign First is designed around the possibility that a fluent recognition candidate—or a fluent AI response—can still be academically wrong.

See the product flow ↗Inspect claims →

07 / Two assurance boundaries

Protect the question before the model and inspect the answer after it.

The second assurance boundary

Confirmed input can still receive a weak answer.

Sign First checks the response separately. The interface distinguishes what a bounded checker can prove, what needs grounding, and what remains explicitly unverified.

Confirmed academic intent

Explain 3 + 5 in one short step.

Selected AI response

Start at 3 and count 5 more: 4, 5, 6, 7, 8. So 3 + 5 = 8.

Assurance decisionReturn with visible checks

A pass is reported only for checks with enough evidence. “Unverified” is a product state, not a hidden failure.

✓
Relevantpass
The response answers the requested operation.
✓
Bounded correctnesspass
The arithmetic result matches the structured intent.
—
Groundingnot needed
This bounded calculation does not need a source.
✓
Reading targetpass
The response fits the selected beginning-reader target.

An AI response moving through a separate illuminated checkpoint — **A confirmed question still needs an answer boundary.**Control exists; this is not an effectiveness guarantee.

Intent assurance

Pause on meaning-changing ambiguity.

Numbers, operations, negation, quantities, entities, time, classifiers, and fingerspelling are examples of academic details that can materially change the request.

Answer assurance

A confirmed question can still receive a weak answer.

Response checks can examine relevance, contradiction, grounding, and presentation. These checks support review; they are not a guarantee of factual correctness.

Human authority

Interruption should shrink as evidence improves.

Full confirmation is a temporary scaffold. The target is background assurance that intervenes selectively while preserving override.

From blanket review to selective protection

Fewer interruptions. Stronger reasons to stop.

Automation earns its way forward only where calibration supports it. Numbers, operations, negation, direction, and weak evidence keep an explicit stop.

Measured now

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

Governed next

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

Evidence required

Larger claims wait for independent, Deaf-governed evaluation.

Continue exploring

See how safety connects to governance.

Data governance→Families→Research→