The numbers

Scored once. On a locked test.

Blind internal test, grouped by chunk and scored once. The headline is the lower of the two systems.

0%
Headline. SNOMED CT on the blind internal test, the lower of the two systems. 95% interval 64.9–75.6.
Three ways to measure the same build
ICD · blind test
72.4%
ICD · held-out replay
68.7%
ICD · out-of-fold
74.9%
SNOMED · blind test
70.6%
SNOMED · held-out replay
68.9%
SNOMED · out-of-fold
67.5%
Out-of-fold numbers were also used to choose components, so the replay (chosen on 60% of chunks, scored on 40%) is the conservative estimate for unseen charts.
0%
Recall in the pool. The right ICD code is a candidate for 94.3% of rows (SNOMED 89.6%). That is the ceiling for everything after retrieval.
0%
Best single scorer. No model alone passes 58% top-1; the 20-seed stacker is what gets to 70%+.
0s / term
Footprint, honestly. About 5 s per term on a 4 GB RTX 2050 with the LLM gate on, 3 GB peak GPU. Built for batch, not for chat.
Error analysis

The bar is 95%. This is where we miss.

∅

Never in the pool

For 5.7% of ICD and 10.4% of SNOMED rows the right code is never retrieved, so no later stage can find it. Retrieval recall caps accuracy.

∞

The long tail

Codes unseen in training are mapped correctly 62.8% of the time on the internal test. That is reasoning, not memorising, and it is the weakest slice.

⌗

Specificity slips

“fx of right distal radius” maps to S52.91XB (open fracture, whole forearm) instead of the closed distal-radius code: a confident, wrong specificity call.

Low-confidence rows are flagged in every upload (confidence under 0.5) so a reviewer sees them first. Numbers come from the TermMap-X+ repository’s own evaluation: internal blind test, 243 ICD and 275 SNOMED diagnosis rows, scored once.

⌂

Stays on your machine

Models run locally or on your own GPU. No hosted API, no keys, no PHI leaving the building.

✓

Always a legal code

Every emitted code is checked against the supplied dictionaries, for both systems, on every row.

◐

Knows what it doesn’t know

A calibrated confidence per system, and the lower of the two, on every row.