Offline Phase Architecture

This file documents REE’s offline consolidation phases.

2026-04-05 roadmap update: A minimal sleep-phase infrastructure (SD-017) is now V3 scope. The four full sub-phases (MECH-120-123) remain V4. See sd_017_sleep_phase_architecture.md for the V3 design.

V3 Minimal Offline Phases (SD-017)

Phase	Name	Function	Status
SWS-analog	Schema/slot consolidation	Hippocampus-to-cortex replay; installs context templates (slot-formation, MECH-166)	V3 required
REM-analog	Causal attribution replay	Fills context slots with co-correlational evidence; slot-filling phase (MECH-166)	V3 required

Sub-phases

Phase	Claim	Biological Analog	V3 Prerequisite
SWS denoising	MECH-120	Synaptic homeostasis (Tononi SHY)	MECH-094 tag
NREM replay	MECH-121	SWR + spindles + slow oscillation	MECH-092 replay
Spindles	MECH-122	Thalamo-cortical spindle bursts	MECH-089 ThetaBuffer (bidirectional)
REM recalibration	MECH-123	REM precision prior (Hobson/Friston)	MECH-094 tag, ARC-016

Phase Ordering Rationale and V4 Rewiring Specification (INV-045)

The V4 sub-phase sequence is not arbitrary. Each phase is in its position because of what must be true before it can run. This makes the ordering a V4 engineering constraint, not a design choice — and each phase directly specifies the rewiring the REE substrate requires.

Phase	Claim	Failure mode addressed	Why this position	Rewiring required
0: Sensory gate	MECH-122	New input corrupts in-progress schema installation; context templates shift during consolidation	Must be first — nothing downstream can run while perception is live	Input gate on E1 latent stack; ThetaBuffer paused for new observations
1: SHY normalisation	MECH-120	Replaying into a landscape dominated by recent high-salience experiences reinforces the dominant trace rather than consolidating diverse content	Must precede replay — homeostasis must flatten attractors before replay repopulates them	Normalisation pass over E1/E2/E3 weights decaying toward mean; dominant attractor suppression
2: NREM schema replay	MECH-121 + MECH-165	You cannot fill context slots that do not exist (INV-044); world model only represents what was done, never what was possible	Must precede REM — schema installation produces the stable attractors that attribution replay fills; replay diversity (MECH-165) must enter the schema now	Hippocampus-to-cortex directed replay; balanced scheduler (forward + reverse + non-dominant paths); ContextMemory templates updated
3: Spindle coordination	MECH-122	E1 state cannot be consolidated into E3 in reverse direction; ThetaBuffer is waking-only in V3	After schema is installed, bidirectional packaging transfers E1 updates for long-horizon integration	ThetaBuffer gains reverse-direction mode; theta-packaged E1 updates transferred to hippocampus
4: REM recalibration	MECH-123	Next waking cycle starts with priors calibrated to the previous episode; early evidence is over/under-weighted	Must be last — resetting precision priors before replay would change how that evidence is weighted during replay	z_beta suppressed (aminergic suppression); E1 runs unconstrained (no harm gate); commitment_threshold and precision_ema_alpha recalibrated from episode natural range

Key derivation: This sequence was not taken from sleep biology. It was derived from asking what an agent needs to reliably know what its actions mean across contexts. The biological NREM→REM sequence converges on the same order because the failure modes are the same. That convergence is evidence for both the biological account and the computational one.

Two findings from the Friday AM literature pull clarify the internal structure of MECH-122 and confirm that Phase 0 and Phase 3 are genuinely distinct sub-mechanisms rather than two names for one thing:

1. Sensory gating and content packaging are temporally dissociated (Yang et al 2018, J Neurosci)

Thalamic (mediodorsal nucleus) firing is suppressed around spindle-uncoupled ripples — this is the sensory gating event. The suppression begins ~0.4 s before ripple onset, preceding the replay event. Thalamic firing is elevated around spindle-coupled ripples — this is the content packaging event. The two modes are sequential within a single consolidation episode, not concurrent: gate first (uncoupled ripples, thalamic step-aside), then package (spindle-coupled ripples, thalamic participation in transfer).

REE implication: Phase 0 (sensory gate) corresponds to the pre-spindle thalamic suppression phase; Phase 3 (spindle coordination) corresponds to the spindle-coupled transfer phase. The two-phase architecture in this table is biologically grounded, not just an engineering choice.

2. The spindle burst is neocortex-initiated, not hippocampus-initiated (Ngo, Fell & Staresina 2020, eLife)

Directionality analyses (Granger causality) in human intracranial EEG show a neocortex→hippocampus influence during spindle-ripple coupling, not hippocampus→neocortex. Long-duration ripples produce stronger coupling.

REE implication: Phase 3 is best understood as a neocortical-hippocampal-neocortical reactivation loop — E1 initiates the spindle-equivalent burst, this triggers E3 replay, and E3 content is returned to E1 within the same burst window. The ThetaBuffer bidirectional mode (Phase 3 rewiring) should therefore model E1 as the burst initiator, with E3 replay activated in response. This is consistent with the Phase 3 rewiring spec (“theta-packaged E1 updates transferred to hippocampus”) but the return path (E3 content back to E1) is the primary consolidation payload; the E1 trigger is the mechanism that opens the window.

Failure Mode

MECH-124: consolidation-mediated option-space contraction (Walker PTSD analog). Prevention requires balanced replay scheduling and MECH-094 tag throughout.

Offline Phase Architecture

V3 Minimal Offline Phases (SD-017)

Sub-phases

Phase Ordering Rationale and V4 Rewiring Specification (INV-045)

MECH-122 mechanistic refinements (2026-04-24 lit-pull)

Failure Mode

See Also