The heads sit between two kinds of measurement: the regulatory databases (what the genome says) and the spatial atlases (what the embryo does). Training is the loop that reconciles them, run per developmental stage.
Forward. Derive each head’s parameters from the genome: enhancer-to-gene maps give conductances and setpoints, super-enhancer clusters give the identity heads, accessibility gives the timing thresholds. Nothing is fitted to the atlas at this step — the parameters are read.
Reverse. Ask whether the atlas implements what was derived: does the predicted fate order match the measured one, does the derived axis match the embryo’s, does the forward-grown organ beat an uninformed baseline against the reconstructed 3D anatomy? Where a check passes, the head is confirmed genome-derived.
Adjust. Where it fails, the failure is kept, not hidden. The ledger is explicit: the anterior–posterior organ addresses are genome-derived (single-cell Hox boundary, Spearman 0.81); the dorso-ventral address is partial; the resting-potential magnitude is anchored to one measured value because a resting potential is a gating quantity, not a transcript quantity — the atlas cannot supply it. A handful of honest anchors, and every one is documented.
A head, concretely, is a master-transcription-factor super-enhancer cluster: the mouse atlas annotates 35 of them, and the sequence front-end resolves roughly 85 — the discovery frontier. The loop’s product is the trained NCA+LGM model.