Latent Stack (L-space)

Claim Type: architectural_commitment
Scope: Multi-timescale latent representation; temporally stratified prediction depths
Depends On: INV-013 (cognition is predictive, iterative, multi-timescale), INV-002 (coherence includes temporal binding)
Status: stable
Claim ID: ARC-004

REE uses a multi-timescale latent stack to represent temporally displaced prediction depths operating over two primary representational domains: (1) a shared sensory latent that fuses modality-specific evidence, and (2) a downstream affordance / action latent generated by fast forward prediction.

Subsystem abstract (core claims): ARC‑004 defines the multi‑timescale latent stack and its depth‑stratified roles. Supporting context includes INV‑013 (predictive, multi‑timescale cognition), INV‑002 (temporal binding coherence), and the E‑stack interfaces (E1/E2/E3) that consume and populate the depths.

Minimal implementation interface

Implement each depth (k\in{\gamma,\beta,\theta,\delta}) with:

  • encode_k(o_t, prev_state) → z_k(t)
  • predict_k(z_k(t)) → z_k(t+1)
  • topdown_kplus1_to_k(z_{k+1}(t)) → prior / conditioning signal for depth (k)

Interpretation

The latent stack is organised by timescale, not by sensory modality. Sensory modalities are fused before entering the stack into a shared sensory latent (z_S).

  • (z_\gamma): shared sensory binding / feature conjunction (multimodal evidence + precision). This level represents the completed perceptual state (z_S).

  • (z_\beta): affordance and immediate action-set maintenance. Fast forward predictions from (z_\gamma) expand into a candidate affordance/action latent (z_A).

  • (z_\theta): sequence context, temporal ordering, and short-horizon narrative structure over candidate trajectories.

  • (z_\delta): regime, motivational set, long-horizon context, and default-mode attractor that biases selection but does not overwrite perception.

Downstream processes may not directly overwrite (z_\gamma); influence is restricted to prediction-error updates and precision (gain) modulation.

REE does not require explicit oscillations for v0. Oscillatory phase can be added later as a scheduling / gating mechanism.

Abstracted implication: cross‑horizon coordination needs a timing/gating layer, but not necessarily oscillations. REE can implement this via phase‑like clocks, event‑driven windows, or explicit scheduling queues. Oscillatory evidence supports the need for structured coordination across depths, but does not constrain the implementation.

Conceptually, oscillatory nesting is treated as one possible carrier for cross-horizon coordination: different predictive depths can be coupled (or decoupled) by phase-aligned gating, enabling stable multi-timescale coherence without allowing deeper regimes to overwrite sensory state.

⸻

Temporal displacement and the constructed present

The latent stack does not represent a single “now”. It represents a temporally displaced bundle of predictions across multiple future offsets, stratified by depth.

A unitary present-moment experience is not a property of (z_\gamma) alone. It is constructed downstream when a trajectory through the displaced bundle is selected and committed (see trajectory_selection.md):

  • In committed / task-engaged mode, commitment phase-aligns temporally displaced predictions with motor execution timing, enforcing cross-horizon coherence so the unfolding trajectory is experienced as a single present.
  • In Default Mode, displaced rollouts are explored without commitment or phase-locking to action (see Default_mode.md), and therefore do not construct a unitary present-moment “now”.
  • Rollout terminology: explicit multi-step rollouts are hippocampal; E2 provides forward prediction kernels that seed them.

This preserves perceptual corrigibility while still allowing deep regime predictors ((z_\delta)) to bias selection through precision and scoring rather than semantic overwrite.

This separation preserves perceptual corrigibility: illusions and misperceptions arise from misweighted precision or suppressed error, not from semantic overwrite of sensory state.

Open Questions

None noted in preserved sources.

  • ARC-004
  • INV-013
  • INV-002

References / Source Fragments

  • docs/processed/legacy_tree/docs/architecture/l_space.md
  • docs/processed/legacy_tree/architecture/latent_stack.md

REE is developed by Daniel Golden (Latent Fields). Apache 2.0.