L6 — a self-supervised world model: prediction is forward settling, error is the learning wave

The substrate predicts itself. Prediction is forward settling — the learned asymmetric transition field acts on the present, the L0 attractor cleans up the next state — and learning is the difference wave (an error-gated update). Held-out next-step error falls monotonically 0.978 → 0.126, generative rollout stays faithful to a 12-step horizon, and a forward model beats a reactive store by directionality.

Predict-next is the universal self-supervised training signal, and a large error wave is a novelty / surprise signal. All four milestones pass: the model learns an unlabeled stream from its own error (W1), rolls out plausible continuations to horizon 12 then drifts honestly (W2), grades novelty by error magnitude (W3), and — the headline — a forward model anticipates where a reactive store that saw every transition only returns the present (W4). The honest [O]: error-gating halts at prediction-sufficiency (the operator is directionally, not magnitude-identically, the L2 coupling, cosine ≈ 0.81).

L6 makes the substrate a self-supervised learner: it predicts its own next state and turns its own error into the learning signal, with no external labels (session v0.7, digest 41e81a7f…). This is the free-energy-like core — a generative world model built from the frozen L0 and the L2 trajectory.

Prediction is forward settling; learning is the difference wave

The forward step is the learned asymmetric coupling acting on the present (thresholded), followed by L0 clean-up onto the next attractor. The error — prediction minus outcome — is a difference wave that drives an error-gated delta update. Held-out next-step error falls monotonically from 0.978 to 0.126, sign-stable across stream length (m ∈ {4,6,8}) and learning rate (η ∈ {0.25,0.5,1.0}).

Four milestones, all verified

L6 world-model capabilities
milestoneresultgrade
W1 learn an unlabeled streamheld-out error 0.978 → 0.126, monotone[V]
W2 generative rolloutfaithful to horizon 12, honest drift at step 13[V]
W3 novelty / surprisegraded: familiar ≈ 0, full violation ≈ 0.97, boundary 0.1[V]
W4 forward > reactive (headline)forward anticipates; reactive only returns the present[V]

Why the forward model wins

W4 is the key result and it honours the L5 break. A reactive store that has seen every transition can only return the present state; a forward model anticipates the next. The advantage is therefore directionality / structure, not a missing abstraction — confirmed by controls: on a fixed-point stream the gap vanishes, and an untrained model has no real advantage.

The honest limit, resolved one layer up

[O] prediction-sufficiency, not magnitude-identity. Error-gating halts when predictions are good enough: the learned operator is directionally the L2 transition coupling (cosine ≈ 0.803) but not magnitude-identical. This is recorded as the seed for L7 — and L7 resolves it: for closed-loop control, prediction-sufficiency is all that is needed.