An uncertainty score is not an instruction. It may describe a model signal, but it does not decide what an interface permits a person or system to do. Product teams need a separate interaction policy: a rule that translates the quality of available evidence into an appropriate human-facing action. The four bands below—proceed, ask, defer and stop—make that translation explicit. They apply the public Viotus principle that judgement, authorship and responsibility remain human while AI expands what a focused team can understand, design and build.
1. Separate evidence from permission
Begin with two columns. The evidence column records what supports an output: completeness of the input, agreement among relevant signals, and any unresolved contradiction. The permission column records the interaction allowed at that evidence state. Keeping them separate prevents a numerical confidence display from quietly becoming authority.
Describe evidence in terms a reviewer can inspect. Avoid pretending that one model score measures truth. A band may use qualitative criteria, or illustrative thresholds when the team labels them hypothetical and validates them for the specific task. The essential point is that permission comes from product policy and human responsibility, not from the model alone.
2. Proceed when evidence supports the bounded task
Use proceed for a bounded, reversible action whose required evidence is present and consistent. The interface still explains the basis, preserves a clear route to review, and keeps the action inside its declared scope. Proceed does not mean certainty; it means the remaining uncertainty is proportionate to the limited decision.
For an illustrative example, a writing assistant may offer a wording correction when the original sentence and the proposed edit remain visible. This example describes a design pattern, not a Viotus capability. If the action changes a consequential record, the same evidence may belong in a stricter band because impact changes the permission.
3. Ask when one answer resolves the gap
Use ask when the system can name a missing fact that the person can reasonably supply. Ask one decision-relevant question rather than displaying a vague warning or demanding all possible context. The prompt should state why the answer matters and how it changes the available action.
Do not turn ask into data collection by default. Request only the minimum information needed for the present decision. If an example involves personal, financial or business data, redact where appropriate, restrict access, and delete the added context after the defined task or retention period. The interface must not imply that more sensitive data automatically creates better evidence.
4. Defer when review needs different evidence
Use defer when the current interaction cannot repair the uncertainty, yet another bounded review path can. The interface names what is missing, identifies the responsible human role or evidence source, and holds the action without presenting a guessed completion. Deferral is a positive routing decision, not a softer version of proceed.
A useful defer state distinguishes incomplete evidence from conflicting evidence. Incomplete evidence calls for a specified source; conflict calls for comparison and judgement. Both cases keep authorship and final decisions with people, matching the published Viotus position on human direction without claiming any particular Viotus interface or unreleased function.
5. Stop when uncertainty meets unacceptable consequence
Use stop when evidence is absent, contradictory beyond repair in the flow, outside the system’s declared purpose, or paired with an irreversible or high-consequence action. Remove the action control or leave it disabled, explain the policy reason in plain language, and provide only a safe route out. A decorative warning beside an active control is not a stop.
Stop criteria should be testable. Teams can review sample cases at the boundaries between ask and defer, or defer and stop, then record why the policy selects each response. Any numeric boundary in such a test is illustrative unless supported by task-specific evidence. The test evaluates policy consistency rather than claiming a universal uncertainty threshold.
6. Test the four-band table as an interaction contract
Build rows from realistic decision states and fill four fields: observable evidence, permitted action, explanation shown to the person, and recovery path. Then challenge adjacent rows with the same output but different consequence. If permission never changes when impact changes, the table probably treats confidence as authority instead of governing interaction.
The completed table gives teams a reviewable contract: proceed for supported bounded action, ask for a resolvable gap, defer for another evidence path, and stop for unacceptable uncertainty or consequence. This method extends the public Viotus principle rather than inventing a product feature. Human judgement defines the bands, reviews exceptions and remains accountable for the final decision.