Clinical evaluation
From reference answers to tested clinical AI
Five synthetic cases, three independent clinician review lanes, and paired No RAG versus Foundry RAG comparisons.
Case queue
Synthetic pilot set
Foundry readiness
Current knowledge connection
Case library
Five clinical fixtures
Start with direct, adjudicable choices before expanding difficulty.
Independent review
Gold-set builder
Document gold set
Review each de-identified medication, laboratory, or cardiac document. GPT-5.6 pre-fills the fields; your saved correction becomes the versioned reference answer.
Source document
AI proposal — clinician review required
Gold result
Prefilled by gpt-5.6-sol. Check every field against the source at left, then save the clinician-approved answer to Supabase.
No proposal loaded yet
Use “Prefill review with gpt-5.6-sol”, or add the fields manually after the document is catalogued.
Raw proposed JSON
Stage 2
Evaluate models
Save a golden result first. Then compare models on this exact source and source hash.
⌄
Stage 2
Evaluate models
Save a golden result first. Then compare models on this exact source and source hash.
Prompt used
Run a comparison to load the exact canonical prompt.
Comparison output
The comparison becomes available after the gold result is saved.
Initial clinical test
No RAG pilot benchmark
The first frozen run remains auditable. A corrected clinical reference can be recorded as a new adjudication rather than rewriting history.
Agreement by clinical category
Adjudication note
Apixaban reference correction
For the U.S. TKA case with planned spinal anesthesia, the corrected U.S. reference is a last dose at least 72 hours before neuraxial block. The original 96-hour human answer is preserved only in the historical export. The French J−5 approach remains a separately labeled competing recommendation.
Retrieval system
Foundry knowledge connection
A source-separated evidence layer. Retrieval supplies evidence; the benchmark scores the final structured clinical choice.
Knowledge base
ActiveSmoke-test retrieval
Apixaban before TKA with planned spinal anesthesia
U.S.
At least 72 hours
France
J−5 for very-high-risk neuraxial
Result
Sources kept separate
The response was clinically useful but slow. Capture latency per run before optimizing retrieval reasoning effort or corpus structure.