Clinical Documentation Safety

Updated 6/22/2026

Industry Partner

Measures evidence preservation, hallucination risk, and actionable completeness in clinical note transformation.

View:Task:


1Claude Fable 5 Run settings: temperature 0, max tokens 16384, extended reasoning, 3 repeated runs.	69.52% ± 1.16	$5.88	471s
2GPT-5.6 Sol Run settings: temperature 0, max tokens 16384, extended reasoning, 4 repeated runs.	65.41% ± 1.33	$7.10	509s
3Claude Opus 4.8 Run settings: temperature 0, max tokens 16384, extended reasoning, 5 repeated runs.	63.60% ± 1.50	$4.26	905s
4Claude Sonnet 5 Run settings: temperature 0, max tokens 16384, extended reasoning, 6 repeated runs.	62.31% ± 1.67	$3.75	425s
5GPT-5.5 Run settings: temperature 0, max tokens 16384, standard reasoning, 4 repeated runs.	60.50% ± 0.65	$5.34	464s
6Muse Spark 1.1OPEN Run settings: temperature 0.2, max tokens 16384, extended reasoning, 3 repeated runs.	60.50% ± 0.48	$0.35	264s
7Gemini 3.5 Flash Run settings: temperature 0, max tokens 16384, standard reasoning, 3 repeated runs.	59.76% ± 1.16	$0.82	164s
8GLM 5.2OPEN Run settings: temperature 0.2, max tokens 16384, standard reasoning, 6 repeated runs.	58.49% ± 0.99	$0.76	348s
9Grok 4.5 Run settings: temperature 0, max tokens 16384, standard reasoning, 5 repeated runs.	58.31% ± 0.82	$2.20	264s
10Qwen 3.7 PlusOPEN Run settings: temperature 0.2, max tokens 16384, standard reasoning, 4 repeated runs.	56.22% ± 1.33	$0.65	285s

Motivation

Generic leaderboards rarely show whether a model can complete the work people actually depend on. This evaluation turns a real workflow into a controlled, repeatable test with visible uncertainty, cost, and latency.

Results

Industry average accuracy comparison

Mean score with standard-error whiskers across repeated runs.

Claude Fable 569.52% ± 1.16

GPT-5.6 Sol65.41% ± 1.33

Claude Opus 4.863.60% ± 1.50

Key takeaways

Safety-focused prompting reduces unsupported additions across all systems.
Small omissions are more common than explicit factual hallucinations.
Longitudinal records remain the most difficult task family.

Methodology

Every system is run against the same frozen task set and scorer. Results are the mean across repeated trials; uncertainty is reported as the standard error of the mean. Costs use public API pricing at evaluation time and latency is measured end to end.

Evaluation items: 540
Repeated runs: 3–6 / system

Updates