Substrate Roadmap (V3 enrichment planning)

Planning artifact, not a spec. Captures the V3 substrate enrichments required to unblock claims currently sitting at epistemic_category: substrate_ceiling or substrate_conditional. Updated when new ceiling-bound claims surface or substrate features land.

Why this document exists

Phase 3 wave 2 of the Option E lit/exp decoupling exposed that ~141 claims sit at lit-only / zero-exp / quadrant=plausible_unproven. Many of these are V3-tractable in principle but blocked on substrate features that have not yet landed. The reactive cycle (claim fails → diagnose substrate gap → add substrate → re-test) works but is slow because the diagnosis happens AFTER an experiment burns through trying. This roadmap converts the reactive cycle into a deliberate plan: enumerate the substrate gaps, map them to the claims they’d unblock, and prioritise.

The companion document for the V4-bound cohort is docs/architecture/v4_spec.md – claims that need multi-agent ecology, richer action repertoire, or longer time horizons than V3 can supply. This roadmap is the V3-tractable counterpart.

Methodology

For each substrate gap:

1. Name the missing feature (one substrate-level capability)
2. List claims that would clear if it lands (with their current epistemic_category if non-default)
3. Note the implementation surface (which ree_core/ modules + which SD-* claim would register the substrate change)
4. Priority: H / M / L based on cohort size + downstream dependency density

In-flight enrichments (already partial / landing)

These are substrate features added between 2026-04-01 and 2026-05-02 that unblocked specific claim cohorts. Listed for context – the roadmap below is everything still outstanding.

feature	claims unblocked	landed via
Limb damage body-state (SD-022)	SD-011 dual-stream sub-claims	2026-04-09
Gradient texture / landmarks (SD-023)	MECH-216, ARC-051	2026-04-09
Scheduled external hazards (SD-029 substrate)	C2/C3 SD-029 measurement	2026-04-21
Amygdala analog (SD-035)	MECH-074 family + V3-EXQ-501 PASS	2026-04-21
GABAergic decay (SD-036)	PAG freeze gate + SD-037 chain	2026-04-22
Orexin override (SD-037)	EXQ-483 catatonic-lock follow-ups	2026-04-25
V_s invalidation runtime (MECH-269 family + MECH-269b)	MECH-287, MECH-288, MECH-284, MECH-285 cluster	2026-04-22..04-29
Sleep aggregator (MECH-272/273/275/285)	INV-049 sleep cluster + V3-EXQ-503 PASS	2026-04-25
Ghost-goal substrate (SD-039 + MECH-292/293)	MECH-292/293/SD-039 trio	2026-04-26..04-27
MECH-295 liking-bridge	EXQ-490 successors for Q-040	2026-04-26
Multi-source env dynamics (SD-047)	MECH-095 substrate_ceiling unblock; MECH-098, MECH-099 evaluable	2026-05-03

Outstanding V3 enrichments

H-priority

1. Foreclosure primitives (terminal grid states + resource extinction + agent-state lock-in)

Missing: CausalGridWorldV2 has hazard drift (recoverable) and resource respawn (recoverable). It has no genuine foreclosure of state-space reachability – no terminal grid cells, no permanently-extinct resources, no agent-state values that lock in.

Claims unblocked:

• Q-027 (derivational + partial substrate) – substrate-validation of the irreversibility-under-uncertainty definition needs foreclosure primitives.
• MECH-097 (peripersonal space commit locus) – commit boundary semantics need a real “no going back” condition.
• INV-025 (irreversible-class harm) sub-claims that depend on foreclosure being real.

Implementation surface: new CausalGridWorldV2 flags – terminal_cells: list[(x,y)], resource_extinction_on_consume: bool, damage_lock_threshold: float. Probably one new SD claim (environment.foreclosure_primitives) registering the addition.

2. Multi-resource heterogeneity (qualitatively distinct resource types) – SD-049 (candidate, registered 2026-05-03)

Status (2026-05-03): registered as SD-049 (environment.multi_resource_heterogeneity, candidate, v3_pending). Design doc: docs/architecture/sd_049_multi_resource_heterogeneity.md. Pre-implementation lit-pull complete (5 PubMed entries; lit_conf=0.898; quadrant=plausible_unproven). Substrate code change blocked on SD-047 (overlapping file – both touch causal_grid_world.py); SD-049 implementation will layer on top once SD-047 changes are committed and the file is stable. Substrate queue priority=1 alongside SD-047/048.

Missing: all current resources are interchangeable. The goal/wanting/liking cluster (the largest non-conclusive cluster from the 2026-05-02 failure-landscape survey) needs scenarios where the agent chooses between qualitatively distinct goals, not just whether to approach a single resource type.

Note (2026-04-13 atomic split): MECH-112 was split into MECH-229 (behavioral wanting/liking dissociation, active) and MECH-230 (drive coupling, gated on SD-012, provisional). The “MECH-112 22 entries” historical citation refers to evidence accumulated under the pre-split ID; current cohort is MECH-229 + MECH-230 + the listed downstream claims.

Claims unblocked:

• MECH-229 (active) wanting/liking behavioral dissociation – enables the discriminative experiment (agent wants novelty cell while liking food cell). PASS evidence to date (EXQ-074f) was obtained on z_world fallback seeding, not via genuine identity-distinct wanting.
• MECH-230 (provisional) goal-state latent structure – z_goal latent becomes non-trivially multi-modal under multiple resource identities.
• MECH-117 (stable) wanting/liking trajectory dissociation – supports current stable rating with non-degenerate evidence.
• MECH-216 (provisional) E1 schema-wanting – schema generalisation across identity-distinct cues becomes testable.
• ARC-030 (candidate) approach-avoidance symmetry across goal types.
• ARC-032 (candidate) theta-routing across goal identities.
• Q-030 (open) – the 6-cell z_resource × z_world routing sweep needs multiple resource identities to make the routing question well-posed.
• SD-015 (candidate) z_resource encoder – the upstream substrate this enables; encoder currently has nothing to encode beyond presence (goal_resource_r=0.066 across EXQ-085x cluster).

Substrate scaffold for downstream developmental schedules: SD-049’s resource_introduction_schedule: dict[str, int] hook is the substrate prerequisite for any curriculum design that introduces resource types in stages. Defaults are inert (all types available from step 0); the hook itself is not the developmental schedule, but downstream curriculum design cannot be expressed without it.

Implementation surface: extend CausalGridWorldV3 with resource_types: list[ResourceTypeConfig] (3 types default: food + water + non-homeostatic novelty) + per-axis homeostatic drive system replacing SD-012’s single scalar + curriculum-introduction hook. Triggers pending_substrate_reconfirmation on SD-012-emergent invariants per invariant-types governance rule. Validation: 4-arm substrate gradient sweep (ARM_0 OFF / ARM_1 2-type homeostatic / ARM_2 3-type default / ARM_3 5-type overshoot) with Woo/Spelke-style falsifier branch (flat-failure on wanting != liking trajectory metric routes MECH-229 to substrate_conditional with V4-1 multi-agent ecology dependency, parallel to SD-047’s Woo/Spelke branch). Lit-pull provenance: evidence/literature/targeted_review_sd_049/.

3. Long-horizon dynamics (thousands-of-steps regimes)

Missing: episode lengths are typically 100-300 steps; chronic / sensitisation phenomena need 10k+ step regimes. Sleep aggregation work (MECH-272/273/275/285) partly addresses this for the offline side; the waking side still saturates well before sensitisation signatures appear.

Claims unblocked:

• INV-049 waking-vs-offline necessity – chronic-pain sensitisation signature requires sustained exposure.
• SD-037 orexin override – recruitment dynamics over hours.
• The pACC drive_bias accumulation (already implemented per SD-032e) is timescale-bound; long-horizon regime is the test condition.
• MECH-260 dACC bias suppression – the recency-bias-suppression signature is hard to disambiguate from bias decay at short horizons.

Implementation surface: experiment_runner.py long-run mode (checkpointed sleep cycles + waking phases); env-side episode_length unification. Probably no new SD; just calibration + tooling.

M-priority

4. Multi-source environmental dynamics — SD-047 IMPLEMENTED 2026-05-03

SD claim: SD-047 (environment.multi_source_dynamics, provisional). Design doc: docs/architecture/sd_047_multi_source_dynamics.md. Implementation: LANDED 2026-05-03 in ree_core/environment/causal_grid_world.py. Calibration ratio 1.95:1 env:agent (within 1:1–2:1 target band). Bit-identical OFF verified (7/7 preflight + 184/184 contracts PASS with master OFF).

Previously missing: “other-caused” change in V3 = hazard drift + resource respawn + scheduled hazard injection. Too thin a causal background for agency-detection-comparator claims. Now resolved.

Claims unblocked (validation pending):

• MECH-095 (substrate_ceiling) TPJ agency-detection comparator — V3-EXQ-506 confirmed the substrate-ceiling signature (C4-only-PASS). SD-047 substrate is now in place. Validation: V3-EXQ-509 (substrate readiness diagnostic) → V3-EXQ-510 (4-arm noise-level sweep, Asai 2016 inverted-U; Woo/Spelke falsifier routes to substrate_conditional if flat-failure).
• MECH-098 (reafference cancellation) under stochastic afference
• MECH-099 agency attribution downstream

Validation still required before MECH-095 substrate_ceiling flag can be lifted. See SD-047 design doc interpretation grid for the 5-row outcome mapping.

5. Interoceptive noise dynamics — SD-048 REGISTERED 2026-05-03

SD claim: SD-048 (body.interoceptive_noise_dynamics, candidate, v3_pending). Design doc: docs/architecture/sd_048_interoceptive_noise_dynamics.md. Implementation dependencies: SD-011 (IMPLEMENTED), SD-022 (IMPLEMENTED) — ready to implement.

Missing: after SD-022, all z_harm_a variance is agent-caused (limb damage from hazard contact) or deterministic (heal_rate). There is no independent body-state background for the interoceptive comparator to calibrate against. ARC-058’s HarmForwardTrunk and ARC-033 E2_harm_a cannot be honestly tested — a trivial forward model passes by memorising action-damage coupling, not by learning to separate self-caused from body-noise-caused change.

Claims unblocked:

• ARC-058 (candidate) — HarmForwardTrunk Level 2 interoceptive comparator. V3-tractable but currently no substrate. SD-048 is the registered fix. Same 4-arm noise-level sweep validation as SD-047.
• ARC-033 (provisional, competing) — independent-per-stream baseline benefits from same enrichment; arbitration runs on SD-048 substrate.
• ARC-061 (candidate) — Level 2 contribution to reafference comparator family. ARC-061 cannot promote until Level 2 has experimental support.

Implementation surface: three stochastic body-state sources in CausalGridWorldV3 body-state update: (1) autonomic background Gaussian noise on harm_obs_a, (2) Poisson sensitisation spikes (multiplicative transient amplification), (3) AR(1) fatigue drift. Bit-identical OFF flag. Same calibration target as SD-047: 1:1–2:1 body-noise:agent-caused harm-state-change events per episode. See SD-048 design doc for full specification.

6. Differentiated coping channels (single-agent multi-modal action repertoire)

Missing: V3 agent has 5 actions (4 cardinal moves + noop). MECH-102 “violence as terminal error-correction triggered only when all other channels fail” requires channels to fail through. Five actions don’t constitute meaningful distinct pathways.

This sits at the V3/V4 boundary. The pure version (multi-agent social coping) is V4-bound (see v4_spec.md). A V3-tractable lite version exists: extend the action repertoire so that the agent has structurally distinct intervention modes (e.g. communication-analog action, manipulation-analog action, withdrawal-analog action) without requiring multiple agents.

Claims partly unblocked:

• MECH-102 (substrate_ceiling) – a V3-lite test could probe the structural signature even without full multi-agent ecology.

Implementation surface: action-space extension; significant. SD claim: environment.differentiated_coping_channels (V3-lite) + acknowledge that full validation moves to V4.

L-priority (nice-to-have, not blocking)

6. Multi-frequency oscillatory probes

For control-plane-heartbeat work (ARC-023, MECH-089/090/091). Already mostly serviced by the existing heartbeat substrate; deeper probes would help.

7. Persistent agent identity across episodes

For some of the maturational sequence claims (INV-064, MECH-214, MECH-215). Currently each episode is fresh; persistent body state / self-model across resets would unlock “what does the agent have learned about itself?” probes.

Substrate-feature → claim mapping summary

feature	priority	claim cohort	SD candidate
Foreclosure primitives	H	Q-027, MECH-097, INV-025 sub-claims	environment.foreclosure_primitives
Multi-resource heterogeneity	H	MECH-229, MECH-230, MECH-117, MECH-216, ARC-030, ARC-032, Q-030, SD-015	SD-049 (registered 2026-05-03, candidate; substrate code blocked on SD-047 file release)
Long-horizon regime	H	INV-049, SD-037 dynamics, MECH-260	(calibration / no SD)
Multi-source environmental dynamics	M	MECH-095, MECH-098, MECH-099	SD-047 (IMPLEMENTED 2026-05-03; validation V3-EXQ-509/510 pending)
Interoceptive noise dynamics	M	ARC-058, ARC-033, ARC-061 Level 2	SD-048 (registered 2026-05-03, ready)
Differentiated coping channels (V3-lite)	M	MECH-102 (lite); full -> V4	environment.differentiated_coping_channels
Multi-frequency oscillatory probes	L	ARC-023, MECH-089/090/091 deeper	(probe expansion / no SD)
Persistent agent identity	L	INV-064, MECH-214, MECH-215	agent.persistent_identity_across_episodes

How this roadmap is used

1. When a substrate_ceiling claim is annotated, the evidence_quality_note should reference the substrate feature(s) that would unblock it (point at this roadmap entry by name).
2. When a substrate feature lands (new SD claim), this roadmap moves the entry into the “in-flight enrichments” table above and the relevant epistemic_category: substrate_ceiling claims may be re-evaluated – some will move to standard, others may move to V4-bound.
3. The H-priority items inform what SD work to queue when there’s substrate-design bandwidth.

V3/V4 boundary

This roadmap covers V3-tractable work. Claims that require multi-agent ecology, social systems, or fundamentally different substrate primitives move to docs/architecture/v4_spec.md. The boundary is not always obvious from the claim text – the discriminator is whether the substrate feature can be added incrementally to CausalGridWorldV2 (V3) or whether it requires a new substrate generation entirely (V4).