The Explaining-Away Penalty
in C. elegans
299 neurons running live in your browser. Real OpenWorm connectome. Leaky integrate-and-fire dynamics, proprioceptive feedback, chemotaxis — no pre-computed playback. The same explaining-away penalty I(D;M|Y) > 0 that governs transformers and quantum circuits, running in a biological neural circuit on your machine.
Three scales, two directions, one framework prediction
The explaining-away penalty I(D;M|Y) exists in this circuit and follows framework predictions at every scale tested. At sensory boundaries, external grounding reduces the penalty (three-point fix). At internal junctions, external drive correlates D and M downstream, raising it (Structure Theorem). Both directions predicted; both measured.
Permutation control: shuffling gap junction current (M) at each dual junction → I(D;M|Y) ≈ 0.001 bits. Real coupling: 0.75 bits. That is a 670× ratio, Wilcoxon signed-rank p = 0.004. The penalty is structural, not an estimator artifact.
Why scalar Pe failed and vector Pe works
The worm sim broke the scalar Pe formula. sim_v9f showed scalar collapse captures only 4.3% of the joint MI ceiling because the three dimensions carry XOR-style synergy (II = −0.631 bits). The fix (§219): vector Pe per axis, zero fitted constants.
Scalar Pe (superseded)
Vector Pe (§219)
II(O;R;α) = −0.631 bits — strong negative interaction information. The three dimensions are synergistic: knowing any two tells you almost nothing about the third. This is the signature of obstructions that are irreducibly 3-dimensional with respect to a Cencov-canonical Y (§218).
v0 sanity to v3.6 molecular Fantasia
Seven substrates, one Fisher metric
The explaining-away penalty I(D;M|Y) > 0 has been confirmed on seven independent substrates. Cencov's uniqueness theorem (1972) guarantees the Fisher metric is the only invariant metric on a statistical manifold — the penalty is substrate-independent by mathematical necessity. No technology substitution routes around it.