Matched labels
Both G10 finalists reached 99.8% agreement with the experimental teacher-silver reference and 100% prediction coverage on the shared 500-patient panel.
Project overview
A modular, cost-aware framework for agentic computable phenotyping
The project tests how rules, machine learning, routing, and language-model components can be compared without losing coverage, auditability, cost records, or the limits of the evidence.
How to use this site
Start with the question, then inspect the exact configuration
The catalog defines the available families. The graph explorer shows how one declared configuration is assembled. The tournament and figure pages then show the recorded aggregate results. The workflow builder is for design proposals; it does not invent performance estimates for combinations that have not been run.
Framework
One common spine; replaceable, versioned components
1
Harmonize
Translate each source into a common evidence contract.
2
Screen
Use transparent rules and local models for decisive cases.
3
Route
Escalate records with structural uncertainty.
4
Adjudicate
Apply the selected agent or bounded program.
5
Audit
Record coverage, cost, probability quality, and failures.
Figure 1. A declared configuration selects exact implementations for each stage and records their dependencies.
Current finding
Similar labels did not imply an identical system
All results here are comparisons with an experimental teacher-silver reference on synthetic records. They are not estimates of clinical diagnostic accuracy.
Probability quality
The mixed-model G10 cascade had the lower Brier score: 0.0025 compared with 0.0103 for the all-Haiku configuration.
Operating burden
The all-Haiku G10 cascade was less expensive under the reconstructed cold-equivalent cost. The practical result is a frontier, not one universal winner.