SD-017: Minimal Sleep-Phase Architecture (V3)

Claim ID: SD-017 Subject: sleep_phase.minimal_sleep_infrastructure_v3 Status: candidate Registered: 2026-04-05 Depends on: MECH-092, ARC-038, ARC-007, SD-014 Blocks: ARC-045, MECH-166, MECH-167 (and transitively all claims requiring attribution mapping convergence)

Roadmap Change

Previous position: All sleep/consolidation phases were V4 scope. (v3_v4_transition_boundary.md previously stated: “V3 is fully online, no sleep/consolidation cycle.”)

New position (2026-04-05): A minimal sleep-phase infrastructure is a V3 requirement.

Why the change: Hippocampal attribution mapping – the core E3 function that enables contextual goal-directed navigation – requires offline consolidation phases to converge. Without them:

Context representations remain globally undifferentiated (cosine_sim -> 1.0) regardless of training duration or supervised signal strength (Law 2016: stabilisation requires ~3 interleaved sessions of multi-context exposure, not achievable online alone)
Slot structure cannot be installed before slot-filling (Sanders et al. 2020: prior must be learned before posterior inference is informative – MECH-166)
The agent navigates as if all contexts are the same – monostrategy, the “fish swimming the same route” failure

This is not a V4 edge case. It appears in V3 and blocks the E3 goal and harm streams from producing useful attribution estimates.

The Three Information Phases

Phase	Biological Analog	Function	REE Mechanism
Waking	Online encoding	Gather what is out there	E1 world model, hippocampus raw experience logging, MECH-092 quiescent replay
SWS-analog	Slow-wave sleep, hip -> cx	Form the slots – schema abstraction, which distinctions matter	Hippocampus-to-cortex replay pass updating context templates
REM-analog	REM reactivation	Fill the slots – causal attribution from co-correlational ordering	Attribution replay, filling context slots with evidence estimates

The information direction switch between phases does real computational work:

SWS-direction (hippocampus -> cortex): schema installation. The hippocampus proposes which context distinctions are structurally relevant; the neocortex (E1, E3 context memory) receives and stabilises these as attractors.
REM-direction: filling. With stable context attractors now installed, recent trajectory evidence is replayed and attributed to the correct context slots, building co-correlational estimates (which outcomes co-occur with which contexts).

Without the direction switch, online encoding fills both roles simultaneously and does neither well.

V3 Minimal Implementation

This is distinct from the full V4 sleep machinery (MECH-120-123). V3 requires:

SWS-analog pass

Triggered: at end of a training epoch or after K episodes (configurable)
Mechanism: hippocampal module runs a schema-consolidation pass in the hippocampus-to-cortex direction
- Distinct from MECH-092 (waking quiescence replay, which is hippocampus-internal)
- Updates context templates in ContextMemory (the “slot structure”)
- Does NOT require z_goal (ARC-038 compliant: no active goal during consolidation)
Output: differentiated context attractors in ContextMemory (cosine_sim < 0.95 target)

REM-analog pass

Triggered: after SWS-analog (slots must exist before filling)
Mechanism: causal attribution replay
- Replays recent trajectory buffer through the attribution pipeline
- Fills context slots with co-correlational evidence (which harm events, which benefit events, which action sequences co-occurred with which context)
- Ordering matters: earlier events in a trajectory contribute causal precedence to later outcomes
Output: updated attribution maps – residue field geometry shaped by context-specific evidence

What this is NOT

Not MECH-120 (SHY synaptic homeostasis with normalisation) – that remains V4
Not MECH-121 (full NREM SWR+spindle pipeline) – that remains V4
Not MECH-122 (thalamo-cortical spindle bursts) – V4
Not MECH-123 (REM precision prior recalibration) – V4
Not MECH-092 (waking quiescent replay) – that is the V3 prerequisite, already implemented

SD-017 is the minimal infrastructure that makes MECH-092 useful: without schema installation first, the quiescent replay has no stable context structure to organise into.

Relationship to Existing Claims

Claim	Relationship
MECH-092	V3 prerequisite (waking quiescent replay, already implemented); SD-017 extends it with bidirectional offline phases
ARC-038	Waking consolidation mode (no z_goal, map geometry update) – the SWS-analog pass is the offline instantiation of this principle
ARC-045	Asserts bidirectionality is necessary; SD-017 is the design decision that implements it
MECH-166	Slot-formation/filling temporal separation – grounded by SD-017’s two-phase structure
MECH-165	Replay diversity (forward/reverse balance) – V4 extension; requires MECH-166 first
MECH-120-123	Full sleep machinery – V4; SD-017 is the V3 scaffold they extend

Evidence Grounding

Law et al. 2016 (Hippocampus 26:1560): context representations require ~3 interleaved sessions to stabilise – grounds the necessity of periodic offline phases for schema formation
Sanders, Wilson & Gershman 2020 (eLife): context discrimination bounded by learned prior quality – grounds the slot-formation-before-filling requirement (MECH-166)
Bengio et al. 2009 (ICML): staged/easy-to-hard training necessary for non-convex representation learning – grounds Phase 0 -> SWS -> REM staging (ARC-042)
Ego-Stengel & Wilson 2010 (Hippocampus): disrupting waking rest ripples impairs subsequent spatial learning – grounds necessity of offline phases (not just online training)

Bayesian Reasoning Framework (INV-044)

SD-017 is not merely a training convenience – it implements the architectural prerequisites for approximate Bayesian contextual inference.

Prior construction (SWS-analog = slot-formation): The SWS-analog pass builds P(context) – the prior over context types. Before this pass, the system has no stable schema of which contextual distinctions are structurally relevant. The hippocampus-to-cortex propagation installs context attractors: compressed representations abstracted from the full episodic record.

Posterior inference (REM-analog = slot-filling): The REM-analog pass performs P(context

evidence) – posterior inference given the installed prior. With context attractors stable, trajectory evidence can be attributed to the correct slot. Without the prior, the posterior is undefined – all evidence competes for globally undifferentiated representations.

Why co-computing fails (INV-044): An online system attempting both functions simultaneously never constructs a stable prior. Every new episode shifts the schema, preventing attribution from converging. The result is cosine_sim -> 1.0 regardless of training duration – exactly the MECH-153 failure mode documented by Law et al. 2016.

Bidirectional flow as generative/recognition duality (ARC-045):

Hip->Cx direction (SWS-analog): generative model installation – hippocampus proposes structure, neocortex stabilises as attractors
Cx->Hip direction (REM-analog): recognition pass – prediction error signals update the posterior

This maps onto the generative/recognition duality of hierarchical Bayesian models (predictive coding, Friston 2005-2010) without requiring explicit probability distributions.

What this is NOT:

Not formal Bayesian inference (no explicit P(context) distribution)
Not approximate message passing or belief propagation
Not Bayesian parameter estimation

It is structural isomorphism with approximate Bayesian contextual inference: the information geometry required for context-sensitive attribution to emerge. SD-017 moves ree-v3 from purely discriminative/reactive (observe -> respond) to a system that maintains and updates a structured generative model of context.

Homeostatic accumulators as empirical Bayes (MECH-167): The slow EMA accumulators (z_harm_a, drive_level) implement empirical Bayes over need states. The accumulated baseline is the running prior; deviations are precision-weighted signals. INV-044 applies: the accumulator must converge before deviations are interpretable.

V3 Build Order Impact

SD-017 must be implemented before the following V3 goals become achievable:

Reliable context discrimination (MECH-153/ARC-042) – EXQ-239 tests the online proxy; SD-017 is the full solution
Attribution mapping convergence (E3 residue field organised by context) – currently unachievable without slot-formation
Goal and harm stream integration with context – z_goal seeding and z_harm_a accumulation need context identity to produce interpretable attribution

Until SD-017 is implemented, ree-v3 is behaviourally an online agent whose hippocampus has the machinery for context-sensitive attribution but cannot activate it – the maps stay flat.