V3 / V4 Architecture Transition Boundary
Overview
V3 is the waking + minimal offline decision-making substrate. Core V3 achievements: a functioning commit gate (ARC-016/ARC-030) with symmetric go/no-go drives (MECH-112/116), multi-rate execution (SD-006), harm/reafference separation (SD-010), and — added 2026-04-05 — a minimal sleep-phase infrastructure (SD-017): SWS-analog (hippocampus-to-cortex schema/slot consolidation) and REM-analog (causal attribution slot-filling). Without SD-017, the E3 goal and harm streams cannot produce useful attribution maps and the hippocampus remains behaviourally impoverished regardless of online training quality.
Roadmap change 2026-04-05: Sleep phases are no longer purely V4 scope. SD-017 moves the minimal sleep-phase infrastructure to V3. The full sleep machinery (MECH-120-123: SHY, NREM SWR+spindles, thalamo-cortical spindles, REM precision recalibration) remains V4. See
sd_017_sleep_phase_architecture.md.
V4 adds the full offline consolidation cycle: MECH-120-123 (SHY, NREM, spindles, REM precision recalibration), consolidated memory transfer from E3 to E1, precision prior recalibration, and dynamic modulation of setpoints currently hardcoded in V3.
The boundary is not a clean wall – V3 must be designed with V4 in mind, measuring and scaffolding what V4 will later make dynamic.
V3 Static Setpoints -> V4 Dynamic Mechanisms
These values are fixed constants in V3. In V4, they will be dynamically modulated by sleep phase machinery. V3 must: (a) measure their natural setpoints under normal operation (so V4 calibration is principled), and (b) accept them as config parameters (already done via from_dims()) so V4 can override them.
| V3 Static Setpoint | Current Value | V4 Dynamic Mechanism |
|---|---|---|
commitment_threshold | 0.40 | Post-SWS recalibration (MECH-120); REM reset (MECH-123) |
residue field decay | per-step static | SWS denoising pass (MECH-120) replaces step-wise |
z_goal.decay_goal | 0.005/step | Extended during NREM consolidation (MECH-121 protects goal traces) |
z_goal.alpha_goal | 0.05 | Boosted during consolidation for goal-trace preferential strengthening |
z_delta recalibration | none | SWS attractor flattening (MECH-120), REM prior reset (MECH-123) |
alpha_world | 0.9 | Modulated by attentional state / z_beta arousal |
ThetaBuffer size | 10 | Modulated by z_beta arousal (MECH-093); bidirectional in V4 (MECH-122) |
e3_steps_per_tick | 10 | Modulated by MECH-093 z_beta heartbeat rate |
E1 prediction_horizon | 20 | Extended during REM unconstrained simulation (MECH-123) |
D_eff threshold | 1.5x baseline | Sleep-calibrated: EXQ-075 setpoint becomes V4 dynamic baseline |
z_beta EMA alpha | 0.3 (shared) | Modulated by sleep phase (near-zero during REM aminergic suppression) |
precision_ema_alpha | 0.05 | Recalibrated post-SWS (MECH-120 SNR restoration changes variance baseline) |
novelty_bonus_weight | 0.0 (off) | V4: modulated by post-REM curiosity state (novel prior hypotheses from MECH-123) |
V3 measurement requirement: For each setpoint above, V3 experiments must log the value’s natural range across episodes (min, max, mean, std) so V4 calibration targets are principled, not arbitrary.
V3 Prerequisites for V4
These must be validated in V3 before V4 consolidation is implementable:
| Prerequisite | V3 Mechanism | Validation Experiment | Status |
|---|---|---|---|
| Quiescent replay infrastructure | MECH-092 | EXQ-061 (PASS) | PASS |
| Hypothesis tag (commit suppression) | MECH-094 | – | candidate, untested |
| Multi-rate execution | SD-006 | EXQ-052b (PASS) | PASS |
| ThetaBuffer waking direction | MECH-089 | EXQ-052b (PASS) | PASS |
| ThetaBuffer bidirectional | MECH-122 prereq | Not yet designed | V4 design needed |
| D_eff monitoring with stable setpoint | MECH-113 | EXQ-075 (pending) | pending |
| z_goal persistent representation | MECH-112/116 | EXQ-074b, EXQ-076 (pending) | pending |
| Balanced harm/goal salience | MECH-124 prereq | EXQ-074b outcome | pending |
| z_beta natural setpoint measured | MECH-093 | ongoing telemetry | partial |
| Go/NoGo competitive commit | ARC-030 | EXQ-077 (planned) | planned |
| HippocampalModule multi-step planning | MECH-163 VTA/planned system | TBD – goal-seeded trajectory generation | V3 full completion gate |
| Minimal sleep-phase infrastructure | SD-017 | EXQ-239 proxy (MECH-153 supervised signal); direct test TBD | V3 required (2026-04-05 roadmap change) |
V3 Sleep-Phase Minimum (SD-017) – Added 2026-04-05
V3 now includes a minimal offline infrastructure, architecturally distinct from the full V4 sleep machinery:
-
SWS-analog pass: hippocampus-to-cortex schema/slot consolidation. Triggered periodically (not every episode). Updates context templates in ContextMemory. Does not require z_goal (ARC-038 compliant). This is the slot-formation phase (MECH-166): installs which context distinctions are structurally relevant.
-
REM-analog pass: causal attribution replay. Triggered after SWS-analog (slots must exist before filling). Replays recent trajectory buffer through attribution pipeline. Fills context slots with co-correlational evidence. This is the slot-filling phase (MECH-166).
What this is NOT (still V4): MECH-120 (SHY), MECH-121 (NREM+spindles), MECH-122 (thalamo-cortical spindles), MECH-123 (REM precision recalibration), sleep phase controller, dynamic setpoint modulation.
See sd_017_sleep_phase_architecture.md for full design.
V4 Scope Summary
V4 implements the full offline consolidation cycle, extending the V3 SD-017 scaffold:
- Entry trigger: quiescent period detected (z_beta below threshold, no external input for N steps)
- Sub-phase sequence:
- Phase 0: Sensory gating (spindle analog, MECH-122) – block new input
- Phase 1: SWS denoising (MECH-120) – residue flattening, z_delta recalibration (extends SD-017 SWS-analog)
- Phase 2: NREM replay (MECH-121) – E3/hippocampal -> E1/neocortical transfer (extends SD-017 REM-analog)
- Phase 3: Spindle coordination (MECH-122) – theta channel bidirectional packaging
- Phase 4: REM recalibration (MECH-123) – precision priors reset, E1 unconstrained
- Exit: all setpoints recalibrated; sensory gating lifted; z_beta restored to waking baseline
- Guard: MECH-124 prevention – balanced replay scheduling, MECH-094 tag active throughout
V4 also adds:
- Dynamic setpoint modulation (all items in table above become driven by sleep phase state)
- Bidirectional ThetaBuffer (MECH-122)
- Full sleep phase controller (new module, analogous to mode_manager but for offline cycle)
- Reverse replay for MECH-165 (replay diversity, forward/reverse balance)
- MECH-304 approach attractor extension (V4-1): V3 SD-051 implements only the commitment-release gate (output 1 of the MECH-304 spec). Output 2 – the safety cue becoming an approach attractor in its own right – requires V4 multi-step trajectory planning infrastructure (HippocampalModule V3 completion gate, MECH-163). Without it, a safety-cue approach attractor reduces to a step-wise proximity signal that cannot compete with harm-avoidance commitments over trajectories. See
sd_051_conditioned_safety_store.mdV4-Deferred Extensions section. - MECH-304 contrastive cue-specific learning (V4-2): The V3 EMA prototype in SD-051 conflates any z_world co-occurring with MECH-302 events, including stable contextual features that are not the conditioned safety cue per se. A contrastive store (trained on paired safety-event / non-safety-event z_world samples) would produce sharper cue specificity. Requires a trainable encoder head with phased-training infrastructure (see EXQ-166b/c/d joint-training failure modes). V3 ceiling: prototype may over-generalise safety predictions in environments with stable non-safety features co-occurring with relief. See
sd_051_conditioned_safety_store.mdV4-Deferred Extensions section.
V4 social extension dependency on V3 hippocampal planning (2026-04-02):
V4’s social extension – sharing joys and sorrows (INV-029 benefit gradient), multi-agent welfare planning – requires multi-step trajectory planning over another agent’s benefit/harm gradient. The habit system (SNc/model-free, MECH-163) can approach its own resources but cannot plan paths that affect another agent’s z_harm_a accumulation or benefit_exposure trajectory over time. Only the VTA/hippocampal system (MECH-163) supports this: HippocampalModule generates multi-step proposals; E3 evaluates trajectories that account for shared gradient fields.
Consequence: validating the VTA/hippocampal system in V3 is a prerequisite for V4 social extension, not merely a V4 feature that can be designed later. V3 full completion gate (row above: “HippocampalModule multi-step planning”) must be passed before V4 entry. See roadmap.md “Two-tier V3 completion” and MECH-163.
Current Session Priority Note (2026-03-23)
The go mechanism (ARC-030, MECH-112, MECH-116, MECH-117) is the current priority over completing the no-go mechanism. The no-go architecture (SD-010, MECH-095, ARC-016) is near-complete. The approach drive architecture (z_goal, wanting/liking separation, D1/D2 competitive commit) is actively being built. Experiments EXQ-072-076 address this. The MECH-124 consolidation failure mode further motivates prioritising go: without adequate z_goal salience in V3, V4 consolidation will amplify the imbalance.
Relation to Existing Transition Documents
v2_v3_transition_roadmap.md– V2->V3 transition rationale and SD-004/SD-005 motivation- This document – V3->V4 boundary specification and static setpoint registry
- V5 scope (multi-agent social synchronisation) is noted in
control_plane_heartbeat.md