Verifiability

The practical distinction

A task is verifiable when a person or deterministic check can detect important errors before harm occurs. Verifiability depends on whether important errors can be detected before harm, by whom, using which evidence, and within the available time. Easy-to-read output is not necessarily easy to verify.

Worked example

A draft email is highly verifiable because the sender knows the facts and reads it before sending. Personalized medical advice may be poorly verifiable for the recipient even when the prose is clear.

Apply it

Lower autonomy or redesign when important errors cannot be checked.

1. Identify the highest-consequence error and the evidence needed to detect it.
2. Name the deterministic check or qualified reviewer that can perform that detection.
3. Lower autonomy or redesign if detection is unreliable, late, or unavailable.

Evidence to collect

• Use seeded material errors rather than asking reviewers whether an output seems plausible.
• Test review under realistic time, volume, and evidence-access conditions.
• Record missed-error and escalation rates for each consequential failure type.

Common mistake

Assuming that a human can verify an output merely because a human is available to look at it.

Scope limit

This guidance on verifiability helps define a task and its review evidence. It does not certify a model, source, reviewer, environment, legal position, or residual-risk level.