v0.3 · foundation surrogate
**whole-device foundation surrogate** — ONE shared-trunk multi-task neural net (per-domain input adapters → a shared plasma-state trunk → per-domain heads, ~67k params) trained JOINTLY across transport (KYRO), equilibrium (KORE), state (KFLOW) and disruption (KWARD) on the data those specialists use. Benchmarked head-by-head against a MATCHED single-task baseline on the same split: on the well-populated domains the shared representation shows **no negative transfer** (equilibrium slightly positive, state/disruption neutral); transport is **data-starved (16 pts) and reported, not hidden**.
KODEX KMIND is a fast, calibrated surrogate for a device-scale multi-task foundation model on real multi-diagnostic data — reproducing its result at inference speed, so it runs inside a real-time control loop or a design search where the full computation is far too slow to call.
One line, one contract — a prediction, its uncertainty, and whether the input is in-domain.
from kronos_ml import KMIND
model = KMIND() # loads the trained surrogate
y, sigma, in_domain = model.predict(x) # y = prediction
if not in_domain: # out of its trusted region
fall_back_to_full_physics() # KMIND abstains, never extrapolatesEvery KODEX code wraps the shared spine — KHALO for calibrated uncertainty and KGATE for the out-of-domain gate — so it reports how confident it is and abstains rather than extrapolate. Benchmarks are computed on held-out data with a fixed seed; pre-registered misses are kept, not hidden.
Part of the open kronos-ml package (Apache-2.0). Open-access deposits with citable DOIs are listed below.
Open-access deposits, each with a citable DOI: