🧠 Predictive Two‑State Memory — Theory

The engine no longer "blends and hopes." A small Echo‑State reservoir builds a latent memory from the lagged past, and an online readout predicts the live present. Memory coherence = predictability of the present from the past.

The coherence iteration

Each frame runs predict → score → learn:

• Encode lag frame & live frame to whitened feature vectors.
• Drive reservoir h with the lag features (leaky update).
• Predict live features from h built only from the past.
• Coherence = exp(−MSE) on that one‑step‑ahead error.
• Update the readout (NLMS) — the recursive "iteration."

Four leakage guards

• Causal order: predict & score before folding the present into memory. No look‑ahead.
• Fading memory: reservoir spectral radius < 1 (echo‑state property) → stale state provably decays. No unbounded memory.
• Holdout audit: 1‑in‑5 frames score but never train. If train‑vs‑val gap spikes → LEAK ALERT.
• Source isolation: features only ever come from raw camera frames, never the rendered composite → no feedback smear.

The reservoir/readout math is real online machine learning; the crystal/telepathic vocabulary is the project's theme.