steffenjohn.eu
[ EXPERIMENT 001 ]

ICD Classification
Experiment

ICD-10-GM 202622 chapters 13,977 terminal codes
02 THE DECISION TRAILRecorded Jev run
Beam width 3 Click a code to inspect its decisionjev-1.13.0
FINAL PATH4 decisions evaluated
Final pathExplored alternativeOther optionsPercentages = local Jev probabilities
Start
  • 98.0%
    • 100%
      • 100%
        • 100%
        • Show 7 other options at G43Scored · not pursued
          • 0%
          • 0%
          • 0%
          • 0%
          • 0%
          • 0%
          • 0%
      • Show 8 other options at G40-G47Scored · not pursued
        • 0%
        • 0%
        • 0%
        • 0%
        • 0%
        • 0%
        • 0%
        • 0%
    • Show 12 other options at VIScored · not pursued
      • 0%
      • 0%
      • 0%
      • 0%
      • 0%
      • 0%
      • 0%
      • 0%
      • 0%
      • 0%
      • 0%
      • 0%
  • 2.0%
  • 0%
  • Show 21 other options at ICD-10-GMScored · not pursued
    • 0%
    • 0%
    • 0%
    • 0%
    • 0%
    • 0%
    • 0%
    • 0%
    • 0%
    • 0%
    • 0%
    • 0%
    • 0%
    • 0%
    • 0%
    • 0%
    • 0%
    • 0%
    • 0%
    • 0%
    • 0%
Replay decisions04 / 04
Original run time2.36 sGate + all hierarchy calls
Est. API cost USD$0.0003728,848 input tokens · 5 requests
Original run estimate. Replay makes no API calls.$0.042 / 1M input tokens
03 THE RESULT
G43.0

Migräne ohne Aura [Gewöhnliche Migräne]

Review code & coding notes
Path score 99.5%

Code suggestion · not a clinical assessment

INSIDE A DECISION

G43 Migräne

8 available optionsJev probability
G43.0Migräne ohne Aura [Gewöhnliche Migräne]100.0%
↳No matching category0.0%
G43.9Migräne, nicht näher bezeichnet0.0%
G43.8Sonstige Migräne0.0%
↳Insufficient detail0.0%
G43.3Komplizierte Migräne0.0%
G43.1Migräne mit Aura [Klassische Migräne]0.0%
G43.2Status migraenosus0.0%

Local choice probabilities, not clinical certainty.

UNDER THE HOOD

Many small decisions.
One traceable path.

Jev is a System One model from TypeSafe. It returns structured choices and probabilities instead of a generated explanation. This map displays those choices—not hidden reasoning.

01

Check the sentence

A diagnosis gate looks for one confirmed, explicitly documented diagnosis. The demo abstains on symptoms, negation, uncertainty, family history or multiple diagnoses. An ambiguous gate also stops the run.

02

Navigate with a beam of three

At each level, Jev chooses among direct children plus two stop options. Up to three candidate paths survive. Their score is the geometric mean of the edge probabilities, following TypeSafe’s cookbook. The tree shows every retained branch and when it was discarded. Expand other options to see all local probabilities. Catalog descendants marked “Not evaluated” were not scored at that replay step. The highlighted route follows the current leader during replay and the final selection when complete.

03

Suggest a code—or stop

The full official 2026 hierarchy contains 17,177 entries. Its largest branch has 44 children: 46 choices including stops, below Jev’s 255-option limit. A terminal code is a suggestion for review, not a validated billing assignment.

Recordings contain real API responses from synthetic examples. Replay timing is slowed for readability. Run time measures server elapsed time, including the diagnosis gate and all Jev API round trips; it is not pure inference latency. Estimated cost uses the total input tokens reported by successful requests and the published price for that model. Output is free. Hosting and any unreported retry charges are excluded. Replaying a trace makes no API calls.

Path scores are ranking heuristics, not calibrated accuracy. This proof of concept has no clinical validation. The imported hierarchy does not contain the complete tabular inclusion and exclusion notes. Review the source, coding rules and context before any real use. Synthetic input only.

Read the hierarchical classification cookbook