V4 Spec (initial deliberate spec, 2026-05-02)

Initial spec-first phase for V4. Captures the substrate primitives V4 needs to support, the claims that wait for them, and the migration shape from V3. Not a substrate implementation – this is the planning artifact that should precede V4 substrate work.

Purpose

Phase 3 wave 2 of the Option E lit/exp decoupling exposed a specific sub-cohort of claims that cannot be tested with any reasonable V3 enrichment, even with the substrate roadmap (substrate_roadmap.md) fully landed. These claims need a substrate generation that V3 cannot incrementally become. V4 is that substrate.

The companion document for V3-tractable enrichment is docs/architecture/substrate_roadmap.md. That document handles the substrate features V3 can incrementally absorb. This document is for the residue: what V3 cannot become.

For a map of the surrounding V4 planning documents, see docs/architecture/v4_planning_index.md.

V4 prerequisites (from V3)

V4 entry is gated on V3 completion. The current V3 completion gates (per ree-v3/CLAUDE.md):

  1. First-paper gate: SD-012 + EXQ-182a oracle + goal-lift experiment. Demonstrates goal representations are real and influence behavior. SD-012 promoted to provisional 2026-05-02.

  2. Full completion gate: hippocampal multi-step trajectory planning validated (MECH-163 VTA/planned system). Required because V4 social extension requires planning trajectories that affect another agent’s z_harm_a and benefit_exposure over time – structurally inaccessible to 1-step greedy.

V3-EXQ-495 is the queued experiment for MECH-163 (per WORKSPACE_STATE 2026-04-30 entries). Until it lands, V4 substrate work should remain spec-only.

V4 scope (the five primitive additions)

V4-0. Object/entity permanence and affordance schemas

The substrate must represent objects as persistent entities whose identity, location, affordances, and causal powers survive partial observation, occlusion, carrying, placement, and delayed re-encounter. V3 resources, hazards, landmarks, and reef cells are useful task features, but they are not yet entity slots with stable identity and intervention-derived affordance bundles.

This is the bridge between the V3 single-agent world model and V4 social ecology. Other agents are a special, harder case of persistent entity: they have object permanence plus policy, welfare, body state, and goal dynamics. V4 should not start multi-agent work until ordinary entity permanence is measurable.

Claims unblocked:

  • ARC-059 scientist-agent developmental ordering cluster.
  • MECH-275 sleep-phase aggregation for stable schema formation.
  • MECH-276 scientist-agent principle / interventional closure.
  • MECH-277 action-space discovery.
  • MECH-278 object-schema formation via experimental action.
  • V4-1 multi-agent ecology, because other-agent modelling presupposes persistent entity slots.

Implementation surface: the reserved harness is docs/architecture/v4_developmental_harness_spec.md. It specifies a MiniGrid-style object-affordance ladder: carry, push, place, use, affordance bundles, confounder isolation, and sleep-aggregation ablation. The likely implementation surface is a V4-reserved environment/harness first, then the REEAgent changes needed to bind persistent entity slots into z_world, hippocampal anchors, and E1 schemas.

Open design questions:

  • Whether entity slots are explicit objects in the observation pipeline, hippocampal anchors, E1 schema memory, or a coordinated view across all three.
  • How identity survives occlusion and carried-object transformations without hand-engineering the answer into z_world.
  • Which affordance predictions are learned from passive observation versus experimental action.
  • How sleep aggregation stabilises object schemas without overwriting episode-specific evidence.
  • How entity permanence becomes agent permanence when the entity has its own action policy and affective state.

V4-1. Multi-agent ecology

The substrate must represent more than one agent and the causal interactions between them. Each agent has its own z_self / z_harm_a / drive state / commitment chain. Agents perceive each other through the same observation pipeline (potentially with self-other distinguishability cues, depending on V4-2).

Claims unblocked:

  • Q-028 (substrate_ceiling) ethics.axiom_conflict_resolution – preserving self vs preserving other requires there to be an other
  • Q-029 (substrate_ceiling) ethics.loneliness_as_harm – loneliness is unshared suffering; needs other agents to share or fail to share with
  • MECH-102 (substrate_ceiling) violence-as-terminal-error-correction – the “all other channels fail” framing requires multiple coping channels which in social context include negotiation, withdrawal, cooperation, etc.
  • MECH-095 (substrate_ceiling) – agency detection becomes more tractable when “other” is structurally distinct from “environment”
  • INV-064 self-model integration sub-claims that depend on representing other selves

Implementation surface: new env class MultiAgentCausalGridWorldV4 (or similar). Each agent has its own REEAgent instance; the env arbitrates concurrent actions, computes inter-agent observations, handles collisions / cooperative state changes. Substantial – this is a new substrate generation, not a V3 extension.

Open design questions:

  • Synchronous vs asynchronous agent ticks
  • How agents perceive each other (full state? body-state only?)
  • Communication primitive (shared signal channel? action-as-signal?)
  • Cooperative state changes (joint resource consumption? joint hazard?)

V4-2. Self-model integration (DR-10..DR-14 from ree-v3 CLAUDE.md)

Five architectural gaps already enumerated in ree-v3/CLAUDE.md “V4 scope (self-model integration)”:

  • DR-10: z_self in E3 trajectory scoring. score_trajectory() currently evaluates entirely in z_world space; bodily state must modulate viability.
  • DR-11: z_self-domain goal representation. z_goal currently lives in z_world; self-state goals (energy restoration, pain avoidance) cannot be represented.
  • DR-12: E2 prediction error -> E3 confidence modulation. E3 trusts E2 unconditionally; PE-magnitude must signal trajectory unreliability.
  • DR-13: z_self temporal depth. Current z_self is single hidden layer + EMA; needs recurrence or E1 feedback for temporal self-model.
  • DR-14: Environment must dissociate proxy from hedonic content. CausalGridWorldV2 conflates location with reward; MECH-214 addiction failure mode is structurally invisible.

Claims unblocked:

  • MECH-214 goal-referent E1-representability (by DR-11)
  • MECH-215 self-model prerequisite for agentive prediction (by DR-10 + DR-12)
  • INV-064 (overall maturational sequence)

Implementation surface: changes to E3.score_trajectory, GoalState, E2 forward-PE wiring, z_self encoder; env-side: SD-022-style body-state extension generalised. Some of this is V3-tractable (DR-12 is partly addressable in V3), but the cohort coheres around the V4 self-model.

V4-3. Long-horizon dynamics + persistent identity

Episodes that span thousands of agent-substrate cycles, with identity / memory / damage state persisting across what V3 calls “episode boundaries.” Shifts from “an agent in a single episode” to “an agent across a lifetime.”

Claims unblocked:

  • INV-049 waking-vs-offline necessity at full timescale
  • Chronic phenomena (sensitisation, allostasis, sleep debt accumulation): SD-037 orexin override sustained dynamics, MECH-260 bias-suppression long-horizon
  • MECH-214 developmental sequence (the maturational architecture needs longitudinal data, not single-episode probes)

This overlaps the substrate_roadmap.md “Long-horizon regime” item; the V3-side fix is calibration + sleep aggregation + checkpointing the existing substrate. V4 generalisation: persistent agent identity across what was previously a hard reset boundary.

Implementation surface: REEAgent.reset() distinguishes episode boundary vs lifetime boundary; substrate flags for persistent state.

V4-4. Richer action repertoire (V3-lite extension also possible)

V3 has 4 cardinal moves + noop = 5 actions. V4 needs structurally distinct action modes that capture meaningful coping channels:

  • Movement (current 4 cardinal + noop)
  • Communication-analog (signal-emission to other agents)
  • Manipulation-analog (effecting non-locomotion change in environment)
  • Withdrawal-analog (committed disengagement, distinct from noop)

Some of this is V3-tractable (substrate_roadmap.md item #5 is the V3-lite version: extend action space without requiring multi-agent). The V4 version requires multi-agent context for the communication-analog to make sense.

Claims unblocked:

  • MECH-102 full ethical-channels framing (V4 version vs V3-lite from the roadmap)
  • MECH-099 agency attribution under richer action causation
  • Several of the ethics axiom claims (INV-028 / INV-029) where intervention modes matter

V4-bound claim cohort summary

From the Phase 3 wave 2 walk and prior surveys:

claim epistemic_category V4 primitive needed
ARC-059 (V4-roadmap) V4-0 developmental ordering / object-schema harness
MECH-275 (V4-roadmap) V4-0 object-schema sleep aggregation
MECH-276 (V4-roadmap) V4-0 interventional closure
MECH-277 (V4-roadmap) V4-0 action-space discovery
MECH-278 (V4-roadmap) V4-0 object-schema formation via experimental action
Q-028 substrate_ceiling V4-1 multi-agent
Q-029 substrate_ceiling V4-1 multi-agent
Q-038 substrate_conditional V4-2 (ARC-053/055 substrate, related)
Q-039 substrate_conditional V4-2 (ARC-053/055 substrate, related)
MECH-095 substrate_ceiling V4-1 multi-source / V4-4 richer ecology (V4-lite plausible)
MECH-102 substrate_ceiling V4-1 multi-agent + V4-4 channels
MECH-099 (untagged, similar) V4-1 + V4-4
MECH-214 (V4-roadmap) V4-2 self-model
MECH-215 (V4-roadmap) V4-2 self-model
INV-064 (V4-roadmap) V4-2 + V4-3 longitudinal
ARC-031 (V4-roadmap) V4-2 z_self navigation
MECH-118 / MECH-119 (V4-roadmap, gated on Q-022) V4-2 self-model

About 17 claims explicitly wait for V4 substrate or its developmental harness.

What V4 is NOT for

To keep scope honest:

  • Out-of-domain claims (Q-031, Q-032 pharmacology) are not V4- bound. No substrate at any level helps them. These belong as research_anchor / literature_synthesis claim types if/when those are added.
  • Derivational claims (Q-025, Q-026, partly Q-027) are not V4- bound. They want axiom-system work, not substrate.
  • V3-tractable enrichments (substrate_roadmap.md items 1-7) belong in V3. Don’t roll them into V4 just because V4 is also a larger substrate.

Migration shape (sketch)

V4 substrate is additive; V3 substrate continues to exist for V3 claims. Sketch:

  1. Build MultiAgentCausalGridWorldV4 as a new env class. Inherits nothing from V3 substrate; allowed to be its own ecology. Before the multi-agent version, build the V4-0 object/entity permanence harness so persistent entities and affordance schemas are validated in the single-agent case.
  2. REEAgent extension: factor out the per-agent state so a single env instance can own N REEAgent instances each with their own state.
  3. New experiment-script template for V4 (multi-agent loops aren’t compatible with the V3 single-agent template).
  4. Indexer + governance: V4 evidence carries a distinct architecture_epoch (e.g. ree_multi_agent_v1 parallel to V3’s ree_hybrid_guardrails_v1). Existing epoch machinery handles this.
  5. Migration of self-model integration (DR-10..DR-14): some changes (DR-12 PE-confidence) can land in V3 first; others (DR-11 z_goal self-domain) wait for V4-1 multi-agent context.

Process gating

V4 spec work proceeds in phases:

  • Phase A (this document): spec-first. Write down what V4 needs.
  • Phase B0 (can be specified before V3 closes): object/entity permanence harness freeze. Implementation remains V4-reserved unless governance explicitly authorises a V3-scale L0/L1 probe.
  • Phase B (gated on V3 completion): substrate prototyping. MultiAgentCausalGridWorldV4 first, then per-agent REEAgent factor.
  • Phase C: V4-1 contract tests, claim-cohort migration.
  • Phase D: V4-2 self-model integration; DR-10..DR-14 enacted.
  • Phase E: V4-3 long-horizon work.
  • Phase F: V4-4 richer action repertoire (lite version may precede in V3 per substrate_roadmap.md item 5).

Proceeding past Phase A requires:

  • V3 full completion gate cleared (MECH-163 PASS)
  • A claim explicitly authorising V4 substrate work (governance decision)

Status

Phase A draft, 2026-05-02. No substrate work has started. This is the deliberate spec-first artifact preceding V4 substrate decisions.

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