The approach

Understand the decision.
Inspect the reasoning.

Useful feedback should explain what mattered, what could have changed the decision, and where the explanation stops.

A small case. A complete reasoning loop.

Start with a short presentation. Choose which findings to request, then make a decision when you think you have enough. An optional probability estimate and reflection let you record your thinking without interrupting the consultation.

The review compares your decision with the alternatives available at that information point. It does not assume you knew a finding you never requested. Try a contrasting examination afterwards to explore how the same question can lead to a different stopping point.

The browser does not invent the verdict.

The case is authored as a finite model: possible causes, observations, decisions, probabilities, resources and utility values. The engine enumerates the admitted decision trees. A separate checker verifies the calculations, and the feedback layer assesses each management choice.

This website displays an exported record for each of the 18 reachable information states. There are 54 checked decision records. The page and data are bound to the same artifact version; incompatible data prevent the case from starting.

Four conclusions, each with a boundary

What the mathematical feedback means
ConclusionMeaning within the declared model
Supported hereYour decision reaches the best expected value in every admitted model.
Another route supportedAn alternative has greater expected value in every admitted model.
Reasonable alternativeYour decision ties the best result under the chosen minimax-regret objective.
Assumption dependentNo uniformly better alternative is established, but your choice is not optimal under that objective.

These are mathematical statements. The numbers and trade-offs are authored judgements; exact arithmetic does not make them clinically correct.

Download the data and an independent numerical verifier →

Show uncertainty without false precision.

The demonstration uses two possible causes and two starting-probability assumptions. Its range varies the starting probability only. The likelihoods and utilities are fixed, so it does not represent the full uncertainty of a real consultation or a diagnostic confidence interval.

When is another question worth asking?

The primary objective is minimax regret: minimise the largest expected-value shortfall relative to the best policy in each admitted model. If several routes tie, the displayed route favours the lowest worst-case expected work, then a stable ordering.

The distinction matters. A route can be optimal under the primary objective and still contain an unnecessary question. The secondary rule removes that avoidable work from the example route without quietly changing what “optimal” means.

The current resource units are illustrative. They are not seconds, money or a validated measure of clinical efficiency. Several first actions can be equally supported; the review does not pretend there is always one perfect next question.

A checked calculation is a starting point.

Clinical validation has not been established.

This is a synthetic educational prototype. Do not use its probabilities, utilities or preferred prescriptions for patient care or to assess a clinician’s competence.

The sore-throat case is deliberately incomplete

Urgent illness and other causes are excluded in the scenario. It models only a streptococcal and a viral cause, two grouped information actions and three antibiotic strategies. Its grouped findings do not implement a complete FeverPAIN score. It cannot train the initial exclusion of serious illness.

Real management also depends on illness severity, complications, patient preferences, contraindications and the complete assessment. The clinical context in every review links to NICE guidance, including safety-netting for every management option. It is separate from the mathematical comparison.

What has not been demonstrated

This prototype does not establish better clinical outcomes, transfer to real consultations, validated diagnostic probabilities, reliable calibration, SCA performance or CPD accreditation. Repeating an exposed case is practice, not a fresh assessment.

The numerical clinical-release profile in the core remains unavailable. Individual human review is necessary for future cases, and it cannot substitute for the evidence, model validation and release controls that are still required.

Your thinking stays with you.

The drill saves one current practice record and its unfinished drafts in this browser’s local storage. It can restore an interrupted consultation or a completed review. If a saved record belongs to a different case version or cannot be restored, it is kept until you choose to replace it. If storage is unavailable, you can still practise for the current page session.

Use “Export practice” to download your choices at any stage, or “Clear saved practice” to remove saved practice, including records from the earlier demo. You can print a completed review as a reflection aid. If another tab changes your saved practice, you can load that record or keep working here without saving. Records and reflections are not sent to an application backend. Do not enter patient information.

This is editable local practice. It is not a tamper-proof commitment, a signed assessment, or an authenticated record of when you made a prediction. An artifact hash detects content changes; it does not authenticate the publisher.

The site contains no analytics, third-party embeds or tracking cookies. Our hosting provider serves the website and may process ordinary request information. Browser storage can be read by anyone with access to your browser profile.

Built to earn the next claim.

The ambition is broader: clinically reviewed UK primary-care cases, natural consultation input, adaptive practice and meaningful progress over time. Those capabilities need evidence and careful testing, not stronger marketing language.

The current public experience is intentionally smaller: one working case family, two authored variations, transparent assumptions and reproducible numerical feedback. There is no booked follow-up, supervisor portal, live clinical dispute process or paid plan in this release.

Try the case