FIELD GUIDE / HEALTHCARE AI

For product leads, CTOs & design teams

The citation looks right.
What would you verify?

Review six citation UX checks before you hand off a healthcare AI prototype. Take a concrete evidence request into your next team discussion.

Healthcare AI Citation UX Review Kit ·

A citation review follows the evidence
  1. 01 Specific claim
  2. 02 Supporting passage
  3. 03 Location & version
  4. 04 Reviewer decision

Opening a link is one step. Deciding whether it supports the claim is another.

Review one prototype. Use this worksheet for educational or knowledge products. It records your observations; it does not assess clinical accuracy, compliance or readiness to deploy.

01 / Inspect your prototype

What can your team show?

Mark Documented only when you can point to an artifact. Leave untested items as Not reviewed.

Your review · nothing reviewed yet
Six self-reported citation UX checks
01

Claim → supporting passage

Bring to the review
One annotated answer with each factual claim mapped to the passage intended to support it.
What to inspect & where it fails
Observe
Open a claim-level citation. Read the passage in context and have a reviewer decide whether it supports that exact claim.
Watch for
A real document can still be irrelevant to the claim. A link that opens is insufficient evidence.

Basis: S01, S02, S03. Editorial review request.

02

Location and edition

Bring to the review
A checked example showing document identity, edition, displayed locator and actual destination.
What to inspect & where it fails
Observe
Compare the displayed page or section with the destination. For PDFs, check printed numbering separately from viewer numbering.
Watch for
A plausible page number or section title can point to the wrong place. A mapping from one document cannot be reused blindly.

Basis: S01, S02. Editorial review request.

03

Samples stay visibly samples

Bring to the review
Side-by-side screens of a sample record and a checked record, including their labels and click behavior.
What to inspect & where it fails
Observe
Ask someone outside the design team to identify the sample. Keep its placeholder identity visible wherever the record appears.
Watch for
A realistic fabricated title or identifier can survive handoff as if someone verified it. Color alone is ambiguous.

Basis: S01. Editorial review request.

04

Missing evidence has a next step

Bring to the review
A small state table with a screen, trigger and next action for each missing-evidence condition.
What to inspect & where it fails
Observe
Walk through each state. Check that a retrieval failure does not silently become a claim that the source lacks the answer.
Watch for
One generic error hides whether the source is absent, inaccessible or simply not found by the retrieval process.

Basis: S01, S02. Editorial review request.

05

Source categories stay separate

Bring to the review
An evidence panel showing source category, title, edition or date, and the supported claim.
What to inspect & where it fails
Observe
Review how handbook content, external research and regulatory records are labeled when they appear together.
Watch for
A nearby authoritative record can make another claim seem supported. Proximity and visual style do not establish authority.

Basis: S01. Editorial review request.

06

Changes survive the handoff

Bring to the review
A before-and-after example plus a review record naming the source version, changed claim, owner and unresolved questions.
What to inspect & where it fails
Observe
Change one source version in the prototype. Show how the old answer and the need for another review remain visible.
Watch for
Replacing a source silently can leave a saved answer carrying stale support. An unowned review request may be missed.

Basis: S01. Editorial review request.

Method & scope

A review request is a starting point.

The six checks are an editorial selection informed by the sources below. They are not an exhaustive safety checklist. The tool sorts your selections; it does not inspect a prototype, retrieve evidence or call an AI model.

Known gaps come first, then unreviewed items, then documented items. Within each group, the question order stays fixed. The exports retain every status, question, requested artifact and source binding.

To connect citation problems with design work, Pharos Production healthcare AI product design makes source inspection and explicit interface states part of the process described in the linked case.

Bring the brief to a review

  1. Choose one representative answer and inspect its source passages.
  2. Discuss each artifact with the person responsible for it.
  3. Assign owners and dates to open questions in your own project tracker.

Who made this, and why?

Published by Pharos Production with AI-assisted research and implementation as a companion to the linked case. Intended to help teams frame product-design and engineering discussions. No independent clinical reviewer or usability validation is claimed.

Only categorical selections are accepted. The kit uses no analytics, cookies, saved answers or form-data upload. Downloads stay on your device. Hosting requests follow GitHub's policies; the linked website has its own policies.

Evidence register

Sources, with boundaries

Inspected . Review questions are editorial applications of these sources.

S01 · AI citation UX for aesthetic medicine

Publisher-reported design-stage case and internal citation audit. No independent runtime or clinical validation. Updated September 24, 2026.

Checked .

Case-study context and link

Retained snapshot SHA-256: b5106de08130bb104960d6585071ae25af9aa9df6f96ff236832a824cb42ce3a

S02 · Citations | Shape of AI

Citation interface patterns: source specificity, metadata and explicit missing sources. Does not validate this kit's six checks.

Checked .

Read Citations | Shape of AI

Retained snapshot SHA-256: a14c91940b684899cbf432259ed8beb3d0edb2563cc1a7626feda811dc79367d

S03 · Explainable AI in Chat Interfaces | NN/G

Original UX analysis of explanation text and citations in chat interfaces. Does not establish clinical reliability.

Checked .

Read Explainable AI in Chat Interfaces | NN/G

Retained snapshot SHA-256: e1c3f7a718d66f70ad5f885b6bc35e38e9fb5e857d858fb28a794fad02748ec4

Keep the version with the brief.

First release · Version 2026-09-26.1 · Schema 1. Identical selections produce identical exports for this version. Future changes receive a new version.

Snapshot hashes identify retained source text, not the truth of a claim. Full third-party source text is not redistributed.