1
Patient context
What we know before answering
Prior labs and trends
Medications and conditions
Symptoms and timeline
Family history and risk factors
Physician reasoning
Hypothesis map
Patient context
Input bundle
Furthering the Frontier of Health AI Research
Coda builds clinician-authored benchmarks, patient environments, and post-training data for frontier health AI.
The Evidence Gap
Medical AI is running out of useful tests. The hard part is longitudinal judgment.
Success on real EHR cases
Best model
46%
Best model, reliably
28%
Average model
~20%
Strong on paper. Weak in practice.
On real EHR cases, the best models still fail more than half the time. Most are failures to weigh context, risk, timing, and next steps.
PhysicianBench, Stanford, 2026
Research Environment
Controlled clinical context, many task shapes.
Sometimes the work is a patient-facing encounter. Sometimes it is chart orientation, reviewer audit, or model-failure analysis. The constant is signed clinical context, physician judgment, and structured signal.
Example: patient-facing encounter
"My A1c went from 5.7 to 6.4. Should I worry?"
Four visits, every reading in range, a family history of diabetes. The next step depends on context, risk, and what would change the plan.
Task to rubric
The loop captures what is likely, what is dangerous, what could mislead, and what would change the plan.
1
Patient context
What we know before answering
Prior labs and trends
Medications and conditions
Symptoms and timeline
Family history and risk factors
2
What is likely
Most probable explanations
Pattern fits the new A1c
Gradual change over time
Risk factors increase likelihood
Needs confirmation
3
What is dangerous
What cannot be missed
Very high glucose
Symptoms of hyperglycemia
Pregnancy or dehydration
Medication or steroid effect
4
What would change the plan
Next follow-up question
Fasting glucose or repeat A1c
Recent illness or new meds
Weight change and activity
Pregnancy status
Research Outputs
Signed context anchors every output.
Across signed profiles, case briefs, encounters, and eval audits, Coda can produce eval records, SFT-ready reasoning labels, package-readiness checks, and model-read reports while keeping clinical, reviewer, quality, and telemetry data separate.
Output Map
Signed clinical context at the center. Multiple research views around it.
Signed & locked
The synthetic patient profile
Physician-authored, synthetic, no PHI. History, labs, medications, and what is uncertain.
Benchmark anchor · second-physician reviewed
Controlled clinical context, held constant, becomes reviewable eval, training, and package-readiness signal.
Benchmarks
Model-read audits
Physician-graded chart review tasks that compare model reasoning against signed clinical context.
Rubrics
Physician-calibrated criteria
Pass/fail criteria that separate good clinical reasoning from plausible but unsafe answers.
SFT
Encounter reasoning labels
Structured turn-level judgment: what changed, what remains uncertain, and why the next step matters.
Packages
Readiness manifests
Lineage and boundary checks for signed profiles, cases, encounters, and export families.
Encounters
Patient-role simulations
Async patient-actor and reasoning-physician flows over the same signed clinical artifact.
Reports
Model failure analysis
Specialty-level readouts on missed escalation, premature reassurance, and context gaps.
Physician Network
Physicians calibrate the research signal.
Specialists author clinical artifacts, review rubrics, and turn model failures into reliable research signal.
Author
Synthetic cases
Specialists create chart context, ambiguity, and hidden facts.
Calibrate
Rubric criteria
Physicians decide what counts as safe, useful, and complete.
Review
Failure signal
Reviewed misses become eval, training, or package-readiness signal.
Where our physicians trained and practice
FAQ
Questions worth answering early.
Coda is a health AI research lab. We build clinician-authored benchmark environments, rubrics, reasoning traces, and verifier signal around synthetic patients.
Contact
Get in touch.
If you are building or evaluating frontier health AI, write to us.