The strong rivalMost users want an answer, not a statistics lesson

Progressive disclosure begins from a sound premise. Showing every caveat can transfer verification labour to the user, slow routine work, reduce comprehension and create indiscriminate distrust. A dense probability panel can make an expert product unusable, while repeated warnings teach people that banners are decorative. The interface should keep the default compact and let people inspect detail when it matters.

The rival fails when “simple” becomes precise-looking concealment. A badge can suppress the fact that a case sits outside the evaluated region. A generic disclaimer can appear on every output while the product still permits irreversible action. An optional details panel is insufficient if the user has no reason or time to open it. Progressive disclosure needs consequence-sensitive triggers that bring a specific missing fact, disagreement or scope failure into the main decision path.

The grammar approach has its own failure mode. Teams can display model internals that do not help the decision, add caveats to excuse correctable defects or assume statistical fluency. A calibrated aggregate can mislead an unrepresented subgroup. An animated distribution can become spectacle without a static range or accessible alternative. The test is behavioural: does the representation help this user choose an appropriate route?

Uncertainty-to-Action GrammarSeven layers, one named next action

The layers are not a disclosure checklist. They form a sequence for deciding what must enter the interface. If a missing jurisdiction changes the legal recommendation, the product asks for it instead of generating a broad answer with a footer. If two current sources disagree only about a financial assumption in the second sentence, the interface localises disagreement there rather than weakening the entire response.

Support and missingness need provenance. The product records which retrieval, database field, measurement or human input produced the evidence and when. Scope converts evaluation into a usable boundary. An overall calibration result cannot travel automatically to a new region, product or language. Change tests whether the boundary remains current. Consequence then sets the required friction and decides whether uncertainty can remain while action proceeds.

The route closes the loop. “Uncertain” without a next action transfers the problem to the user. A useful route may request a fact, compare scenarios, preserve work for a qualified owner, allow a reversible draft or stop a customer-facing action. Override should identify who can decide, what evidence they see and what receipt remains. A warning that does not alter authority, information or route is decoration.

Representation follows evidenceUse numbers only when they mean what people infer

A raw model score may rank classes or tokens without estimating the chance that a decision is correct. Self-reported confidence is generated language. Entropy summarises a distribution over model outputs. Ensemble disagreement compares selected models. None should be labelled “probability of truth” without empirical calibration against defined outcomes in a represented population and period.

Calibration asks whether forecasts and outcomes align across comparable cases. Accuracy and calibration are distinct: a classifier can be accurate overall while its probability estimates are systematically too high or too low. Proper scoring rules help evaluate probabilistic forecasts without rewarding strategic or overconfident reporting. Post-hoc methods such as temperature scaling worked well in the settings studied by Guo and colleagues, but that result does not establish a universal method or transport across shift.

When probability evidence is valid, a frequency can make it concrete: in the stated cohort and period, 18 of 100 comparable cases with this forecast range were wrong. The example is illustrative unless tied to actual evaluation. The interface names what “comparable” means and provides the calibration population, window and outcome definition. A range or natural-frequency translation may help some users, while others need categorical language or a scenario comparison. Test comprehension and action across roles and accessibility needs.

Intervals can represent a forecast range, measurement uncertainty or parameter uncertainty. The interface must say which. Coverage claims require evaluation under the method's conditions. A narrow interval is not automatically better when it omits model misspecification or shift. Static endpoints and quantiles should remain available even when animation helps people perceive a distribution.

Scenarios are useful when uncertainty comes from unresolved assumptions rather than a single probability. Compare the recommendation if demand holds, falls or recovers; identify which evidence would distinguish the paths. Categorical language is preferable when no valid number exists: “jurisdiction missing, do not issue advice” is more honest and actionable than an invented confidence percentage.

Consequence-sensitive frictionThe route changes before the colour does

Begin with the maximum credible effect, not the interface label. Identify the person, asset, service or obligation affected; when an error becomes detectable; what can be reversed; and whether time pressure makes delay harmful. A draft classification can remain visible for later correction. A denial communicated to a customer can trigger reliance, rights or downstream systems that a simple edit cannot undo.

Low-consequence, reversible work should remain light. Show a compact marker that distinguishes supported and unresolved claims, preserve source access and make correction easy. Avoid modal warnings that force acknowledgement without changing the task. Routine use teaches the meaning of the marker, so wording and placement should remain consistent through versions.

Comparison fits cases where evidence supports multiple credible paths and the decision depends on a visible assumption. Show ranges or scenarios side by side, identify the source of disagreement and let the user inspect the claim that changes. The product should not create artificial alternatives merely to look balanced. It should state why an option is credible and what observation would favour it.

Ask when a specific obtainable fact could change the route. A contract assistant should request jurisdiction, governing terms or date rather than generate a polished universal answer. The question preserves gathered work and explains why the fact matters. Repeated requests for information the system already holds are a quality defect, not uncertainty disclosure.

Abstain and escalate when evidence or scope cannot support the task but a qualified owner can. The product states what it can preserve, what it cannot infer and who receives the case. Block when a consequential action would cross a boundary that evidence, authority or recovery cannot support. A block may still allow research, drafting or evidence collection below the protected action.

Warning fatigue signals a broken route design. Count banners that change no information, authority or action. Replace recurring generic warnings with a stable status, localise the specific issue and reserve interruption for cases that need a different decision. If every output is “yellow”, the product has neither classified uncertainty nor designed a response.

Claim-level disclosurePut uncertainty where the decision can use it

Output-level warnings flatten different conditions. One response can contain a well-supported definition, a disputed financial assumption and an unsupported legal conclusion. A footer saying “AI can make mistakes” provides no route. The evidence map attaches support, missing facts, disagreement, scope and source age to the material claim that changes the decision.

Localisation should preserve reading flow. Use a compact visible state with meaningful text or icon and provide detail on demand. Colour can reinforce but never carry the distinction alone. Keyboard focus, screen-reader names, zoom, contrast and static alternatives are part of the uncertainty contract. A person should be able to locate unresolved claims without watching animation or distinguishing red from green.

Provenance is not a link count. Show whether a source supports, conflicts with or merely mentions the claim. Record retrieval time and version where change matters. When an inaccessible source prevents support checking, say so and route accordingly. Do not let the visual presence of citations imply proposition-level verification.

Alternative views should remain bounded. A credible alternative has relevant evidence or a decision-changing assumption; random generations do not constitute disagreement. Show why sources differ and whether reconciliation is possible. If the product cannot establish which alternative applies before a reversible draft, comparison may suffice. Before an irreversible action, unresolved disagreement can require escalation or block.

The interface should also reveal policy boundaries. A product may prohibit a recommendation because authority, fairness review or legal scope is unresolved even when a model produces a confident output. Label the condition as a policy or compliance limit, not a probability. Legal duties vary by role, system, provision, timing and jurisdiction and require qualified interpretation.

Calibration is lifecycle evidenceWithdraw the number when the population moves

A probability display needs a decision uncertainty receipt. Record the decision and affected party, material claim or action, uncertainty type and provenance, calibration population and date, missing information, scope exclusions, credible alternatives, consequence, reversibility, time pressure, route, override, owner and later outcome. The receipt is this article's synthesis and must not collapse into one uncertainty score.

Evaluation begins before deployment. Define the observable outcome and its time window, split data without leakage, measure discrimination and calibration on relevant cohorts, and use proper scoring appropriate to the forecast. A numerical display should have a traceable transform from model output to evaluated probability. If calibration is absent, the interface removes the number and uses support, missingness, scope or categorical action language.

Deployment completes the loop. Record what the person saw, which route they selected, whether they overrode and when the outcome became known. Reliance is not proved by a click. Compare acceptance, verification and error across forecast bands, roles and consequences. A display can be statistically calibrated and still create poor decisions because users misread it, lack time or face incentives to accept.

Shift interrupts inheritance. Ovadia and colleagues showed in their studied classification benchmarks that predictive uncertainty and calibration can degrade under dataset shift, and that common post-hoc calibration can fall short. Product teams should not generalise the ordering of methods beyond those settings. They should test current input and outcome distributions, especially new regions, products, languages, suppliers and workflow conditions.

When the case falls outside the calibration population, withdraw the probability rather than leave it with a small warning. Show the scope difference, retain useful evidence and choose another route. A new region may still permit comparison or a qualified review. It should not inherit last quarter's domestic frequency simply because the same model returns a score.

Revalidation follows material changes to the model, prompt, retrieval, feature pipeline, policy, interface or outcome process. A wording change can alter reliance without changing calibration. NIST's published AI RMF 1.0 and Appendix C emphasise explicit human roles, context and ongoing evaluation; the framework is currently under revision. These are risk-management foundations, not proof that a particular display works.

Ten boundary casesReveal uncertainty by changing the route

The examples show why one representation cannot win universally. Frequencies can aid comprehension when calibration is valid; ranges can show forecast spread; scenarios expose decision assumptions; categories name concrete limits. User testing should measure both understanding and action, including whether people notice a scope failure, inspect decisive evidence and choose the intended route.

Worked forecast panelAnimation can reveal variation without becoming evidence

Consider a monthly demand forecast used to plan reversible inventory. Hypothetical outcome plots animate draws from a forecast distribution. Studies by Hullman and colleagues and by Kale and colleagues found benefits for particular comparison and trend-judgement tasks. Those results do not make animation a universal solution. The visual is an interface technique whose comprehension, accessibility and decision effect require testing in the actual task.

The panel keeps a static view beside the animation: median, stated quantiles, units, evaluation period and frequency performance. The illustrative numbers in Figure 7 show structure, not a real product claim. A deployed panel would calculate them from a documented forecast and outcome cohort. The user can pause motion, read the range and compare assumptions without watching a moving path.

The panel also distinguishes aleatory variation represented by draws from unresolved structural uncertainty. If promotion timing can create materially different demand, show separate scenarios rather than folding the assumption invisibly into one distribution. If a supply commitment is reversible, comparison may be enough. If it creates an irreversible customer or financial obligation, route to qualified review or a protected limit.

Accessibility changes the representation, not the evidence. Reduced-motion preference removes the animation; a screen reader receives the static statistics, cohort, assumption and route; colour is repeated with labels and line descriptions; keyboard users can reach details and actions. The decision receipt records which view and version were available, not to monitor individuals but to understand whether the control worked.

Interactive instrumentUncertainty Interface Router

The Router turns decision context, evidence condition, calibration, disagreement, representation and authority into one of five action routes. It returns interface elements and a named next action. It never produces an uncertainty, trust or maturity score. If valid calibration is absent, numerical confidence is removed rather than replaced with model self-assessment.

Five interface routesPair every uncertainty state with an action

Answer with bounded evidence when support is direct, the case is inside the evaluated scope, no material fact is missing and consequence is low or controlled. Use a compact claim-level marker, source access, categorical limit or validated frequency and easy correction. Keep the answer within the product's authority and preserve the receipt.

Show a range and compare alternatives when credible scenarios or sources differ and the decision can use the distinction. Name assumptions, show static endpoints or quantiles, localise disagreement and identify what observation would favour a route. Do not compress structurally different scenarios into one precise number.

Ask for the missing information when a specific obtainable fact could change the answer. Explain why it matters, preserve work already done and prevent the generic response from becoming a substitute. The question should be answerable by the user or a named source, not a ritual request for assurance.

Abstain and escalate to a qualified owner when the case is outside scope, calibration is invalid, evidence is absent or a material conflict cannot be resolved by the interface. Remove numerical confidence, state the supported remainder, pass sources and missingness to an owner with the required qualification and record the transfer.

Block the consequential action when severe or irreversible effect would proceed despite unsupported evidence, unavailable material facts, invalid calibration or absent effective review. The product can keep lower-authority research and drafting available. A warning plus an enabled commit button is not a block.

The route needs an override boundary. A qualified decision owner may proceed under law, policy or professional judgement that the interface cannot encode. Show what was uncertain, what evidence the owner added, which authority applied and how outcome feedback will be reviewed. An override should not silently retrain the product's future default or erase the original evidence state.

The decision uncertainty receipt supports product learning. It joins the output shown, representation, population, source age, alternatives, route, action, override and later outcome. Review by task and consequence, not as an aggregate score. A low overall error rate can hide a shifted region; frequent overrides can indicate poor routing, a policy mismatch or missing expertise.

Uncertainty design succeeds when a person can tell what is supported, what is missing, where the evidence applies, which alternative matters and what to do next. The interface does not apologise for being uncertain. It converts uncertainty into bounded work, an answer, a comparison, a question, a qualified escalation or a protected stop.

Review the release as a control surface

Before launch, run matched cases through the full interaction. Include supported answers, plausible unsupported precision, missing fields, stale sources, represented and shifted populations, source disagreement, urgent decisions, severe effects and accessible alternatives. Test whether users reach the intended route without relying on colour, animation or prior statistical training.

Measure comprehension and behaviour separately. Ask users what the number or category means, which cohort it describes, what could change the conclusion and what action remains permitted. Then observe whether they inspect evidence, provide the missing fact, compare alternatives, escalate or stop. Preference and completion time matter, but neither proves that the interface calibrates reliance.

Monitor the route after release. Warning acknowledgement, click-through and open rates are weak evidence when users have learned to ignore them. Examine decision outcomes, route changes, overridden blocks, abandoned cases and cohorts with sparse feedback. A person may leave because non-reliance is rational. Do not redesign scepticism away when the evidence remains thin.

Retire a representation when its evidence expires. A supplier, region, outcome definition, calibration transform or workflow can change while the badge remains visually identical. Version the receipt, withdraw unsupported numbers, notify affected users when reliance could change and revalidate both statistical performance and human interpretation before restoration.

Decision receipt and operating proofKeep uncertainty attached to the outcome it governed

Build the receipt before designing the badge

The decision uncertainty receipt starts with a plain-language decision and affected party. “Customer risk” is too broad. “Whether to place this account into an enhanced-review queue” identifies an action, while “whether to deny the application” identifies a different consequence and authority. The same model output can support the first and be prohibited from the second. Recording the distinction prevents interface convenience from silently expanding the product contract.

List each material claim and its provenance. A claim may be supported by customer data, a current contract clause, an outcome model or a policy rule. Record missing facts and scope exclusions beside it. If a probability appears, store the transformation, calibration population, outcome definition, period, cohort results and date. If no probability appears, record the categorical evidence state and why a number would be invalid. The absence of a number is an evidence decision, not missing polish.

Consequence fields should describe the effect graph. Name who can be affected, which external systems receive the action, when reliance begins, what can be reversed and how long detection and recovery take. This makes friction proportional. A user can often correct a private draft after reading it. A customer may rely immediately on a denial notice, and connected systems may propagate it before a reviewer sees an alert.

The route records interface elements as well as the decision. Note which claims were marked, which details were available, whether motion was enabled, the question asked, the owner selected and the action blocked or permitted. Record an override with the added evidence and authority, not merely a button click. Later outcome linkage should preserve privacy and purpose limitation while still supporting evaluation of calibration and route quality.

Test the human interpretation, not just the rendering

Interface evaluation needs matched conditions. Compare a raw score, calibrated frequency, range, scenario and categorical route only where each is evidentially legitimate. Ask users to explain the representation in their own words, identify the population, state which fact could change the conclusion and choose the next action. A visually preferred design can fail if it increases acceptance of unsupported cases.

Include initial judgement where the workflow permits it. Human-AI studies indicate that whether people decide before seeing advice, how uncertainty is worded and who the user is can affect reliance. These results establish design variables, not universal prescriptions. Test domain expertise, numeracy, language, disability access, time pressure and organisational incentives in the intended setting. Avoid covert individual profiling; the product needs evidence about interaction conditions, not a speculative trust trait.

Warnings require negative controls. Repeat a generic banner across supported and unsupported cases, then compare a local missing-fact prompt or a protected action boundary. Observe whether users notice, understand and change behaviour. If the banner produces acknowledgement without better routing, remove it. An uncertainty display should earn its visual weight by changing information, authority or action.

Test accessible equivalence. With animation disabled, can the user recover static quantiles, frequency, assumptions and route? Without colour, are status and difference still labelled? At 200% zoom and 390 pixels, can controls be read and operated without horizontal page movement? With a keyboard or screen reader, do claim markers, details and route changes appear in meaningful order? Equivalent evidence matters more than identical presentation.

Operate the route under change and incident

Monitoring joins statistical and behavioural signals. Track calibration and proper scores by represented cohort, missingness rates, scope rejections, disagreement routes, overridden blocks, correction outcomes and time to qualified escalation. Separate low-consequence drafts from irreversible decisions. Aggregates can hide a small shifted population or a route whose errors are systematically hard to detect.

Material change withdraws inherited evidence. A new region, model, feature pipeline, retrieval source, supplier, policy, interface wording or outcome process may invalidate the population or the way users interpret the display. Remove the number or narrow the route first, then revalidate. Keeping an old percentage with a freshness icon preserves the false inference that the value still describes correctness.

When an incident reveals unsupported reliance, identify the receipts that used the affected source, calibration or interface version. Contain downstream action, notify relevant owners, correct the product state and reassess affected outcomes. Recovery tests the failure mechanism and user interpretation before restoring scope. A fresh badge, new colour or longer explanation cannot substitute for a repaired evidence loop.