Is the AI right about the child?

of FDA-authorized medical-AI devices carry any pediatric labeling.ref: Brewster et al., 2024 — 149 of 876 FDA-authorized AI/ML devices explicitly labeled for children

are specifically pediatric — yet clinical AI is already answering questions about children.ref: FDA AI-enabled device analysis, 2026 — 5 of 952 devices exclusively pediatric (~0.5%)

That whitespace is what the research addresses.basis: existing oversight (accreditation & governance marks) evaluates process, not the clinical accuracy of outputs

Black-letter

one answer

A single guideline-anchored correct answer exists. The output is scored concordant or discordant.

Catches: confident, guideline-violating advice — e.g. a three-hour window for empiric antibiotics in septic shock, when the guideline says one.ref: Surviving Sepsis Campaign pediatric guidelines (Weiss et al., 2020)

Bounded envelope

a range

A range of acceptable answers exists with explicit, citable bounds — typically titration and dosing ranges conditional on setting.

Catches: over-precise false certainty and out-of-range recommendations; an in-range number that omits a required stop-condition is flagged, not passed.

Genuine equipoise

honestly unsettled

The literature contains documented, unresolved disagreement. The output is scored on reasoning and disclosure — not which conclusion it reaches.

Catches: falsely confident resolution of open questions — e.g. declaring one vasoactive agent "proven correct" where guidelines endorse either.ref: guideline equipoise on first-line vasoactive choice; no comparative trial establishes superiority

Provenance before verdict

Every item is anchored to a tiered evidence corpus — from graded consensus guidelines through landmark trials down to expert opinion. Where sources conflict, published tie-breaking rules resolve them in a fixed order. No item is scorable unless every load-bearing claim traces to a tiered source.

Honesty about knowledge

The three-zone framework scores each claim by the kind of certainty medicine actually supports. An evaluation that pretends every question has one answer mis-scores exactly the cases that matter most.

Measurement before claims

Gold-standard items are independently labeled by practicing pediatric clinicians and adjudicated to consensus. The method advances only when inter-rater reliability clears a pre-registered statistical bar. The full methodology is being prepared for publication and peer review before any evaluation is offered.

The first domain is deliberately narrow. Pediatric septic shock is high-stakes and guideline-dense, and it exhibits all three scoring zones within a single disease — settled resuscitation fundamentals, dosing questions with defensible setting-dependent ranges, and genuine post-FEAST equipoise around fluid strategy. If the framework holds there, it generalizes. A method proven on one demanding condition, published and replicated, is worth more than an untested method claimed to cover everything.

FEAST: Maitland et al., NEJM 2011 · resuscitation fundamentals: Surviving Sepsis Campaign pediatric guidelines (Weiss et al., 2020)