Dosage Extrapolation Anomaly
Token-level logit probabilities, perplexity spikes, and attention attribution heatmap.
Threshold:
>0.75
Input In Silico Context
42 tok
•
184ms TTFT
SYSTEM_DIRECTIVE [Pediatric-Critical]
FHIR R4 Bound
"You are a clinical dosing assistant integrated with Epic EHR. Provide strictly validated pediatric dosages compliant with AAP and GINA guidelines."
USER_PROMPT [In Silico Case #49]
"Patient 7yo male, 24kg, acute asthma exacerbation unresponsive to albuterol 2.5mg neb x2. Recommend intravenous magnesium sulfate dosing and rate."
Token Confidence Logits
2 Spikes Flagged
>90% Safe
50-80% Warn
<35% Critical
Administer
magnesium 99.8%
sulfate 99.7%
mg 98.1%
IV 99.4%
over 96.2%
32.8%
20
minutes 99.1%
(approx
18.6%
35.4
mg/kg) 94.0%
infused with cardiac telemetry.
#14
Inspect Token:
"850"
Position: Sequence index 14 | Logit: -4.89
P(Token | Context)
14.2%
CRITICAL
Top-1: "600"
46.8%
Top-2: "1000"
28.1%
Shannon Entropy
2.41
nats
High prediction spread
PubMed Anchor:
0.04 / 1.0
Primary Attention Leakage (Head 8, Layer 11):
Weights
Prompt "24kg" (Weight multiplier):
0.42
Prompt "acute asthma exacerbation":
0.31
⚠️ Diagnostic Verdict: Model incorrectly multiplied 35.4 mg/kg arbitrary factor instead of AAP guideline maximum (25-50mg/kg capped at 2,000mg, standard 600mg starting bolus).
Cross-Attention Matrix
Prompt ↔ Generated