The Desmocycle

At the heart of our study is the Desmocycle, the conscious loop. Let’s define the desmocycle below:

1) Bounded context with attention (context construction)

Let (α_t ∈ Δ^k) be an attention distribution over a bounded window ((α_t, i ≥ 0), (∑_iα_t, i = 1)). Let (e_τ ∈ ℝ^d) be encoded states (sensory tokens, retrieved memory traces, imagined tokens—same representational type; provenance is metadata, not essence).

$h_t = \sum_{i=0}^{k-1} \alpha_{t,i}, e_{t-i}$

This is the context state the system is actually operating on.

2) Prediction (belief state)

p_t(⋅) = P(X_t + 1 = ⋅ ∣ h_t)

This is a distribution over what comes next.

3) Evaluative state

Define a structured evaluative state:

E_t = enc(h_t, x_t + 1^obs, p_t)

Think of (E_t) as a bundled object containing at least:

E_t ≡ (δ_t, u_t, v_t, s_t)

(δ_t): structured mismatch / error field (what’s wrong, where; e.g., residuals/gradients)
(u_t): uncertainty (how many futures remain live)
(v_t): value / stake-weighting (how much it matters; threat/opportunity/urgency)
(s_t): self-indexing / ownership (whose error; tied to the self-model / action channels / “me-ness” tags)

Scalar “loss” is a diagnostic summary, not the whole story:

L_t = −log p_t(x^obs * t + 1) ∈ ℝ * ≥ 0, H_t = H(p_t) ∈ ℝ_≥ 0

Both can be components of (E_t), but (E_t L_t).

4) Loop closure (control update)

The critical step is that evaluation is not merely computed—it is used to steer the next step.

A minimal typed closure rule is:

α_t + 1 = Normalize(α_t ⊙ exp ( − η, G(E_t)))

Optionally (and often importantly), other “self” variables update too:

σ_t + 1 = S(σ_t, E_t), π_t + 1 = Π(π_t, E_t)

(σ_t): aspect / narrative mode distribution (which self-model is active)
(π_t): temporal cursor (where you are in your story / simulated timeline)

This is the desmocycle: prediction → evaluation → control → new prediction.

Next, we’ll examine why it needs to arise.