What’s training now

Public highlights from the lab. Full runs, metrics, and pipelines live behind the operator login.

Router model · active

Jeff — the d3-edge router

A small language model (0.59B) being trained as a calibrated router for agent orchestration: which lane, which tier, how confident. Labels come from executable unit tests, not vibes.

Contrastive dataset v2026-10-13-4: 618 rows, 147 families
Golden suite: 385 cases, ~4,355 checks green
Gates: ECE under 0.05, held-out win required — no promotion without one

Embeddings · serving

MTNN v5 — hoops player embeddings

The serving champion behind the basketball embedding map: every player-season as a point in 64-dimensional space, where nearness means stylistic kinship.

12,966 player-seasons, 64 dimensions
Single-model serving policy; challengers must beat it held-out

Underwriting · evaluating

eye-ml — property risk models

Machine learning on property underwriting data. The incumbent is a humble L1 GLM — and so far it’s beating the fancy gradient booster on held-out data.

Held-out deviance: L1 204.24 vs LightGBM 319.35
Pure-premium Gini: L1 0.270 vs LightGBM −0.195

Highlights updated 2026-10-14. Numbers are real measurements from held-out evaluations.