SD-054: Reef Enrichment Substrate (CausalGridWorldV2 extension)

Status: IMPLEMENTED 2026-05-04 Renamed: 2026-05-08 (was SD-050; collided with the canonical Suffering-Derivative Comparator entry in claims.yaml) Gap ID: SD-054-reef-enrichment Depends on: none (pure environment substrate)

Problem

Agents trained in CausalGridWorldV2 with a single resource attractor and uniformly-drifting hazards consistently collapse to a monomodal policy: a single fixed route between corner-adjacent forage targets. The route is locally optimal under the existing hazard model, but it produces an environment-event distribution dominated by agent-initiated transitions and a near-zero rate of independent environment-initiated harm contacts.

This blocks SD-029 (self-attribution comparator) at the measurement level. SD-029’s C2 / C3 acceptance criteria require balanced agent-vs-environment event distributions to dissociate “I caused this” from “the world caused this.” A monomodal policy structurally cannot generate the harm-event types SD-029 needs to measure: there are not enough environment-initiated contacts in the data stream, and the agent’s own contacts are stereotyped to a single route shape. Every iteration of SD-029 (V3-EXQ-433/433a/433b/470) reclassified non_contributory for this reason rather than for an architectural failure of the comparator itself.

The substrate fix is to enrich the environment so monomodal exploitation is no longer locally optimal. SD-054 introduces two competing behavioural attractors – safety-seeking and foraging – that together create a two-mode strategy space the agent cannot collapse into a single fixed route.

Design

Principle: behavioural partitioning by competing attractors

The biological analog is coral-reef refugia: in marine systems, predator-free patches with distinct microhabitat structure force fish into context-dependent strategy selection (shelter-seeking near the reef vs. foraging in open water). Single-template exploitation breaks down because each strategy is locally optimal in only part of the environment.

In CausalGridWorldV2 this becomes:

Attractor	Implementation	Effect
Safety (reef)	Hazard-free cells with a static scent gradient	Pulls a stressed agent toward corner-adjacent shelter
Foraging	Existing food spawning + new food-attracted hazard drift	Increases hazard density along agent-food paths

The two attractors are spatially anti-correlated (reef patches occupy corners; food spawns elsewhere) so a fixed route cannot satisfy both simultaneously.

Reef safe zones

Up to six corner-adjacent patch centres are available:

(2, 2), (2, sz-3), (sz-3, 2), (sz-3, sz-3),    # four corners
(2, sz//2), (sz-3, sz//2)                      # two mid-edges

The first n_reef_patches are used. A patch consists of all interior cells within Manhattan radius reef_patch_radius of the centre. Cells in a patch are added to self._reef_cells.

Hazard exclusion. Reef cells are removed from the spawn pool at reset, so hazards and food never start inside a reef patch. During _drift_hazards(), no hazard may step into a reef cell – both toroidal and bounded branches enforce (nx, ny) not in self._reef_cells. A defensive guard in the step harm-resolution path zeroes harm_signal if a hazard somehow ends up colocated with the agent in a reef cell.

Reef scent gradient. A static (sz, sz) field is precomputed at reset:

field[i, j] = sum_over_reef_cells(exp(-manhattan_distance / reef_scent_sigma))
field /= field.max()

The agent sees a 5x5 local view of this field appended to world_state whenever reef_enabled=True. The gradient gives the agent a continuous distance-to-shelter signal, analogous to how SD-023 landmark-B fields give a distance-to-resource signal.

Food-attracted hazard drift

During _drift_hazards(), with probability hazard_food_attraction per drifting hazard, candidate move directions are sorted by Manhattan distance to the nearest food cell rather than randomly shuffled. The hazard then takes the first valid direction in that ordering. This is not a single biased step – it is a deterministic preference for food-ward moves whenever the gate fires.

The effect is that foraging paths become actively more dangerous over time as hazards accumulate near food spawns. With reef-zone safety on the same map, the agent must trade off between “stay in shelter” (no food, no harm) and “go forage” (food, but rising harm density along the path).

hazard_food_attraction=0.0 (default) leaves the legacy uniform random walk bit-identical on every drift step.

What does NOT change

harm_obs is unchanged – reef cells are not nociceptive signals; they affect the world observation only.

harm_obs_dim stays 51. Only world_obs_dim grows: 250 -> 275 when reef_enabled=True (+25 from the 5x5 reef field view). With multi-resource heterogeneity (SD-049) or landmarks (SD-023) also enabled, the dimensions stack additively.

No encoder changes. SplitEncoder receives the new channels via the larger world_obs_dim exactly as it does for SD-023 landmark fields.

Implementation

File: `ree_core/environment/causal_grid_world.py`

New __init__ params (all default to disabled / no-op for backward compatibility):

reef_enabled: bool = False,
n_reef_patches: int = 3,
reef_patch_radius: int = 2,
reef_scent_sigma: float = 2.5,
hazard_food_attraction: float = 0.0,

New state and helpers:

self._reef_cells: Set[Tuple[int, int]] (populated by _place_reef_patches)
self._reef_field: np.ndarray[(sz, sz)] (static scent gradient, recomputed at reset)
_place_reef_patches(available) – builds _reef_cells, computes the scent field, and removes reef cells from the spawn pool

reset() calls _place_reef_patches(available) before placing hazards / food, so reef cells are excluded from available for the duration of the episode.

world_obs_dim property: returns previous_dim + 25 when reef_enabled=True.

_get_observation_dict(): 5x5 reef field view extracted from the agent’s local window and appended to world_state; obs_dict key "reef_field_view" [25] added.

_drift_hazards(): food-attraction sort gate before the candidate-direction loop; reef-cell exclusion in both bounded and toroidal hazard-step branches.

Step harm path: defensive zero of harm_signal if (new_x, new_y) in self._reef_cells.

Config

Reef params are CausalGridWorldV2 constructor params – not in REEConfig.from_dims(). Experiments using the reef substrate must set world_obs_dim=275 explicitly when constructing REEConfig (or world_obs_dim=300 when SD-023 landmarks are also on, etc.).

No agent changes

This is a pure environment extension. No changes to LatentStack, ResidueField, BetaGate, or any agent-side module. No new training target. No phased-training implications.

Backward compatibility

reef_enabled=False and hazard_food_attraction=0.0 are defaults. With the flag off:

_reef_cells is empty, _reef_field is zeros (or never instantiated)
world_obs_dim returns the legacy 250 (or 300 / 275 + landmarks / 250 + N x 25 for SD-049, unchanged from before SD-054)
_drift_hazards behaves bit-identically to the pre-SD-054 random walk – no extra RNG draws when hazard_food_attraction=0.0

All existing experiments are unaffected. The bit-identical-OFF property was verified at substrate-readiness time (V3-EXQ-521).

Design hypothesis vs. observed outcome

Design hypothesis (what SD-054 was supposed to enable). Without SD-054 the only attractor is food, and the optimal policy is a fixed corner-to-corner shuttle. With SD-054 the environment carries two anti-correlated attractors (reef safety, food); a policy that discriminates threat-context from safe-context can occupy both, producing a contrastive agent-vs-environment harm-event distribution – the data SD-029 / MECH-256 need for C2 / C3 calibration.

What the substrate actually delivered. Under a hand-coded reef-aware avoidance heuristic (EXQ-522 ARM_1: “if hazard within 2 cells and reef available -> flee to reef; else -> forage”) the substrate produces ~50/50 reef/forage occupancy, ~49 zone transitions per episode, and a materially different position-entropy profile from the reef_disabled baseline. This is the substrate-ceiling demonstration: given a discriminative policy, SD-054 carries the intended behaviour into the data stream.

What the substrate did not deliver. Under trained REE policy (the same E3 + E1 + LatentStack substrate that EXQ-522’s heuristic replaces), every SD-029 retest on SD-054 substrate has returned non_contributory for the same reason:

Run	Outcome	Direction	Diagnosis
EXQ-523	n/a	non_contributory	Undertrained: r2=0.57 < 0.9 graduation gate
EXQ-523a	INCONCLUSIVE_UNDERTRAINED	non_contributory	Same
EXQ-523b (x2)	INCONCLUSIVE_UNDERTRAINED	non_contributory	Same
EXQ-433e	FAIL	non_contributory	Insufficient agent-caused trials
EXQ-433f	FAIL	non_contributory	C0 trials-sufficient gate FAILed in 3/4 seeds (agent_caused_trials: 15/7/20/3 vs target 20). Same monomodal V_s monostrategy substrate-ceiling pattern as 433/433d

The trained agent does not adopt a discriminative reef-vs-forage policy. It stays monomodal under gradient descent on a single-policy parameterisation, even with the reef substrate present. The behavioural-diversity ceiling that the heuristic in EXQ-522 reaches is not reached by the trained policy. The contrastive harm-event distribution SD-029 needs is therefore not produced under current REE training.

Diagnosis (registered as MECH-309). The gap between EXQ-522 (heuristic, diverse) and EXQ-433e/f / 523-series (trained, monomodal) is the size of an absent rule-apprehension capacity at the policy layer. Bayesian / gradient-style updaters revise weights over a hypothesis space they do not invent; without a non-Bayesian rule-creator that proposes discriminative policy modes (“near-hazard -> reef regime; else -> forage regime”), the trainer collapses to the smoothest single policy that is good-enough across the whole state space. Monomodal collapse on this substrate is the equilibrium output of the present architecture, not a failure of training. EXQ-522’s hand-coded heuristic externalises the missing apprehension layer; this is why it works where the trained agent does not.

What this means for SD-054. The SD is necessary substrate for behavioural diversity – the heuristic-policy upper bound proves the environment can carry the split. It is not sufficient under current REE training. The proximate next-stage blocker is the rule-apprehension layer (registered as ARC-062 / ARC-063), with MECH-269 V_s monostrategy as a representational precondition. SD-054 substrate-readiness is delivered (V3-EXQ-521, V3-EXQ-522 PASS); SD-054’s purpose claim – that the substrate unblocks SD-029 measurement under trained policy – is not. The latter is recorded as v3_pending until the rule-apprehension layer (weak reading first; strong reading deferred to V4) is built and re-tested.

ML/AI engineering notes

No ML/AI engineering concerns. SD-054 is a pure environment extension – no encoder, no training target, no new learning component. The z_world encoder (SplitEncoder) receives the new channels automatically via world_obs_dim. No phased training. No gradient flow. No MECH-094 interaction (env observation stream, not replay content).

The only calibration choice is reef_scent_sigma (default 2.5). Larger sigma flattens the gradient and reduces the safety-cue informativeness; smaller sigma sharpens it and risks overfitting policy to scent-direct paths. The default matches SD-023 landmark-B sigma so the two gradient signals are comparable in scale.

Validation

V3-EXQ-521 – substrate readiness diagnostic (PASS, 2026-05-04)

7 / 7 acceptance criteria:

ARM_0 (reef_enabled=False): world_obs_dim=250, len(_reef_cells)=0
ARM_1 / ARM_2 (reef_enabled=True): world_obs_dim=275, len(_reef_cells)=33 (12x12 grid, 3 patches at radius 2)
0 reef-cell violations across 30 episodes x 200 steps x 3 seeds in ARM_1 / ARM_2 (no hazard ever entered a reef cell; spawn pool exclusion verified)
ARM_2 mean food-distance 2.057 vs ARM_0 baseline 3.790 (-46%) – food-attraction bias measurably pulls hazards along forage paths
Behaviour-distribution entropy not collapsed: ARM_2 = 4.049 >= 0.7 x baseline 4.557 = 3.190
Agent reef-visits = 1987 across the diagnostic in ARM_1 (reef is reachable and visited under uniform exploration)

V3-EXQ-522 – substrate-ceiling demonstration via heuristic policy (PASS, 2026-05-05)

Important framing. EXQ-522 used a hand-coded reef-aware avoidance heuristic (v3_exq_522_reef_monostrategy_break.py:65) as the policy under test, not a trained REE agent. The heuristic is:

if hazard within FLEE_THRESHOLD cells AND reef cells available:
    move toward nearest reef cell
else:
    move toward nearest food

Under this heuristic, ARM_1 (reef_enabled=True, hazard_food_attraction=0.7) produces ~50/50 reef/forage occupancy, ~49 zone transitions per episode, and a materially different position-entropy profile from ARM_0 (reef_enabled=False). This demonstrates that the substrate can carry diverse behaviour given a discriminative policy – it is the substrate ceiling. ARM_1_reef_food was selected as the canonical reef config for downstream experiments (see V3-EXQ-524 fishtank showcase).

EXQ-522 is silent on whether trained REE agents reach this ceiling. They do not (see below).

Trained-agent retests on SD-054 substrate (FAIL / non_contributory)

V3-EXQ-433e, V3-EXQ-433f, V3-EXQ-523, V3-EXQ-523a, V3-EXQ-523b: all trained-agent SD-029 retests on the reef substrate (often combined with hazard_food_attraction>0) returned non_contributory for the same diagnostic reason – monomodal V_s monostrategy substrate- ceiling pattern; insufficient agent-caused trials for C2 / C3 calibration. The substrate was present and validated; the trained policy did not exploit the structural opportunity the substrate offered.

This is the empirical foundation for MECH-309 (monomodal-collapse-as-equilibrium-without- rule-apprehender). The substrate is necessary; the rule-apprehension layer (ARC-062 weak reading, ARC-063 strong reading) is the next-stage architectural commitment that needs to land before SD-054’s purpose-claim can be re-tested.

SD-054 – this SD (env behavioural-diversity substrate; substrate-readiness PASS, purpose-claim v3_pending under trained policy)
MECH-309 – monomodal-collapse-as-equilibrium-without-rule-apprehender; the diagnosis that explains the EXQ-433e/f and 523-series non_contributory pattern
ARC-062 – rule-apprehension architectural slot, weak reading (gated-policy architecture; V3 first pass)
ARC-063 – rule-apprehension architectural slot, strong reading (distributed CandidateRule field with tolerance-gated availability and evidence-trace records; implementation_phase=v4 deferred flag)
SD-029 – self-attribution comparator (primary downstream beneficiary; SD-054 + rule-apprehension layer together are the prerequisites for SD-029 C2 / C3 measurement, not SD-054 alone)
MECH-256 – SD-029 successor mechanism; consumes the same enriched event stream
MECH-269 – regional verisimilitude V_s monostrategy diagnostic; representational precondition for ARC-062 / ARC-063 (you cannot apprehend a rule about a region you cannot represent as discriminably different from another region)
SD-023 – environmental gradient texture (parallel substrate enrichment pattern; SD-054 reuses the static-field-with-5x5-view design from SD-023 landmarks)
SD-049 – multi-resource heterogeneity (parallel env enrichment; world_obs_dim stacks additively with SD-054)

Naming history

The original CLAUDE.md entry (2026-05-04) labelled this substrate SD-050. The canonical SD-050 in claims.yaml is the Suffering-Derivative Comparator. The collision was resolved on 2026-05-08:

ree-v3/CLAUDE.md SD-050 reef block renamed to SD-054 (with an inline note)
17 experiment scripts swept from SD-050 to SD-054 in docstrings, comments, and one banner print string (cosmetic only – the indexer keys on manifest claim_ids_tested, not stdout or script source)
SD-051 (Conditioned Safety Store), SD-052 (Contextual Passive Safety Terrain), and SD-054 (this entry) registered as proper claims.yaml design_decision entries on the same date
SD-053 was left informally reserved for a sustained-drive claim per docs/architecture/sustained_drive_anticipatory_wanting.md

Bipartite layout extension (2026-05-11)

Status: IMPLEMENTED 2026-05-11

The legacy SD-054 corner-placement is a weak geometric expression of the behavioural-diversity claim: reef patches sit in fixed corners, but the agent, hazards, and food spawn at uniformly-random positions in the remaining cells. Per-episode reef-vs-food geometry is randomised, and the mean optimal policy across episodes converges to a single direction-following template (“head toward nearest food”) that works regardless of the specific reef-vs-food layout. The claim that “two competing attractors create a two-mode strategy space” is true under a discriminative policy (V3-EXQ-522 heuristic, PASS) but the legacy geometry does not create the structural pressure that a trained REE policy would need to discover the discrimination on its own.

The bipartite-layout extension is the sharper geometric expression of the same claim. It does not change SD-054’s semantics, dependencies, or v3_pending status. It is a substrate-refinement that creates the structural condition under which a well-trained ARC-062 substrate would produce per-candidate first-action argmax diversity at typical training probe states.

Motivation

The 2026-05-11 diagnose-errors session on V3-EXQ-543b (TASK_CLAIMS diagnose-v3-exq-543c-2026-05-11T0635Z) ran a direct numerical probe against the legacy SD-054 substrate and recorded:

8 CEM candidates from agent.generate_trajectories(latent, e1_prior, ticks) at init all share argmax(first_action) = 3
Continuous first-action vectors differ by ~1e-4 max across candidates
world_states[1] (first POST-action predicted z_world per candidate) differs by ~1.2e-5 max across candidates; grows to ~2.8e-3 by t=30
The ARC-062 head’s input (per-candidate z_world summary) is therefore near-constant in K, gated_score_bias varies by ~4e-7 across K, TV distance vs bypass-uniform is ~9e-8 – below the V3-EXQ-543b inert-gating threshold of 0.05

The diagnosis: the substrate-readiness premise of V3-EXQ-543b was violated upstream of the ARC-062 head, by the CEM proposer producing near-degenerate candidate batches. The legacy SD-054 substrate does not exert geometric pressure that would force a trained CEM proposer to emit diverse first-action argmaxes. See evidence/planning/arc_062_rule_apprehension_plan.md decision-log entry 2026-05-11 for the three architectural options surfaced; this extension implements option 3a (env-side diversification).

Design

Three new CausalGridWorldV2.__init__ kwargs (env-only, NOT surfaced through REEConfig.from_dims – matches SD-022 / SD-023 / SD-029 / SD-047 / SD-048 / SD-054 precedent for env-only SDs). All defaults preserve bit-identical legacy SD-054 behaviour:

reef_bipartite_layout: bool = False,        # master switch
reef_bipartite_axis: str = "horizontal",    # or "vertical"
reef_bipartite_agent_band_radius: int = 1,  # 0 = midline-only; 1 = midline +/- 1

When reef_bipartite_layout=True:

Region	Cells (axis=”horizontal”, radius=r)	Cells (axis=”vertical”, radius=r)
Reef half	rows in `(midline + r, sz - 2]`	cols in `(midline + r, sz - 2]`
Agent band	rows in `[midline - r, midline + r]`	cols in `[midline - r, midline + r]`
Forage half	rows in `[1, midline - r)`	cols in `[1, midline - r)`

Reef patches are placed along the reef-half edge (row = sz - 3 for horizontal, col = sz - 3 for vertical) with the remaining axis centres evenly distributed across the interior. Patches that would intersect the agent band or forage half are clipped by the _is_in_reef_half(i, j) predicate before being added to self._reef_cells. The reef scent field is computed identically to the legacy helper.

The agent always spawns in the agent band; hazards, resources, and waypoints always spawn in the forage half. Cells in the reef half that are not reef cells are intentionally unused (the reef half is reef-exclusive). On a 12x12 grid with n_reef_patches=3, reef_patch_radius=2, agent_band_radius=1, the typical layout has ~28 reef cells (rows 8-10), 4 agent-band rows used out of 3 (rows 5-7 minus any walls), and ~32 forage-half cells (rows 1-4 minus reefs and walls).

Action-axis correspondence

For axis=”horizontal” with CausalGridWorldV2.ACTIONS = {0: (-1, 0), 1: (1, 0), 2: (0, -1), 3: (0, 1), 4: (0, 0)}:

Reef approach (toward bottom): action 1 (“down”, row +1)
Forage approach (toward top): action 0 (“up”, row -1)

These are categorically opposite action argmaxes. A candidate pool that includes both reef-mode and forage-mode trajectories MUST have at least two distinct first-action argmaxes – there is no policy that gets both modes via the same first action under this geometry.

Reset partitioning

reset() branches at the reef-placement step:

if self.reef_enabled:
    if self.reef_bipartite_layout:
        self._place_reef_patches_bipartite()       # does NOT consume `available`
        agent_pool, forage_pool = self._build_bipartite_pools(available)
    else:
        self._place_reef_patches(available)        # legacy: removes reef cells
        agent_pool = forage_pool = available       # aliased single list
else:
    self._reef_cells = set()
    agent_pool = forage_pool = available

ax, ay = agent_pool.pop()                          # agent spawn
# hazards, resources, waypoints pop from forage_pool

Legacy mode aliases the two pools to the same available list so pop()-from-shared-pool semantics are preserved bit-identically. Bipartite mode uses two disjoint pools that are independently shuffled by self._rng inside _build_bipartite_pools.

Fallback under degenerate configs

If agent_pool would be empty after partitioning (e.g., agent_band_radius=0 on a grid where the midline is mostly walls or already reef cells), the band is widened iteratively by +1 radius until a valid cell exists. The widen count is recorded in self._sd054_bipartite_band_widen_count (0 in legacy mode and in successful bipartite resets) for diagnostic surfacing. If forage_pool is empty, it falls back to agent_pool so hazards.pop() does not crash; the substrate-readiness diagnostic will catch this via low forage-pool cell counts.

What does NOT change

world_obs_dim is unchanged (still 275 with reef_enabled=True, +25 from the static reef scent field view). No new observation channels. No new encoder. No new training target. harm_obs is unaffected. The step() movement, hazard, and resource-consumption logic is unchanged. The legacy hazard_food_attraction knob continues to work in both modes – with the bipartite layout, hazards in the forage half drift toward food (also in the forage half), so the forage path becomes actively dangerous over time in exactly the way it does under the legacy layout.

Backward compatibility

reef_bipartite_layout=False (default) preserves bit-identical legacy SD-054 behaviour: _place_reef_patches (corner pattern) is called, available is partitioned to remove reef cells but is otherwise a single shared pool, and all entities pop from it. RNG draws are unchanged because _build_bipartite_pools is not invoked. Bit-identical-OFF verified at smoke-test time 2026-05-11 (reef_cells count 33, world_obs_dim 275, band_widen_count 0 across seeds 0/1/2/42).

Validation

V3-EXQ-548 substrate-readiness diagnostic to be queued via /queue-experiment immediately following this implementation. Acceptance criteria (proposed):

C0 backward-compat: reef_bipartite_layout=False reproduces V3-EXQ-521 baseline (reef_cells=33, world_obs_dim=275, band_widen=0)
C1 structural partition: with bipartite ON across 3 seeds, all reef cells are in rows > midline + radius; all hazards / resources / waypoints in rows < midline - radius; agent_pos always in midline +/- radius
C2 first-action argmax entropy: at probe states sampled from a 40-episode P0 warmup, mean per-state argmax entropy across K=8 CEM candidates is

= 1.5x the ARM_0 (legacy) baseline value. ARM_0 baseline expected ~0 (smoke-test value 0); ARM_1 expected >= 0.3 (i.e., at least one probe state of 32 produces 2+ distinct argmaxes)
C3 reef-vs-forage signal: at agent-midline probe states with hazard pressure present, the per-candidate first-action argmax distribution contains both 0 (forage / up) and 1 (reef / down) in at least 30% of probe states (the structural-readiness threshold)

Acceptance is structural-only in this pass (experiment_purpose=diagnostic). A full-pipeline rerun of V3-EXQ-543b/c (with bipartite ON in 543b’s env kwargs) is deferred until V3-EXQ-548 PASS confirms the env-side fix sufficient.

SD-054 (this SD; substrate-readiness PASS, purpose-claim v3_pending still applies until the rule-apprehension layer lands and is re-tested on the enriched substrate)
MECH-309 (logical-necessity monomodal-collapse claim; the substrate enables a sharper falsifier but the architectural claim is unchanged)
ARC-062 (Phase 2 GAP-B downstream consumer; this extension creates the structural conditions under which the GAP-B falsifier becomes testable)
arc_062_rule_apprehension_plan.md GAP-B (status blocked; resume condition is V3-EXQ-548 PASS)
SD-023 / SD-047 / SD-048 / SD-049 (parallel env-only substrate-enrichment pattern – env-only kwargs, no REEConfig.from_dims surfacing)
MECH-094 (not applicable – env observation stream, not replay content)

SD-054: Reef Enrichment Substrate (CausalGridWorldV2 extension)

Problem

Design

Principle: behavioural partitioning by competing attractors

Reef safe zones

Food-attracted hazard drift

What does NOT change

Implementation

File: ree_core/environment/causal_grid_world.py

Config

No agent changes

Backward compatibility

Design hypothesis vs. observed outcome

ML/AI engineering notes

Validation

V3-EXQ-521 – substrate readiness diagnostic (PASS, 2026-05-04)

V3-EXQ-522 – substrate-ceiling demonstration via heuristic policy (PASS, 2026-05-05)

Trained-agent retests on SD-054 substrate (FAIL / non_contributory)

Related claims

Naming history

Bipartite layout extension (2026-05-11)

Motivation

Design

Action-axis correspondence

Reset partitioning

Fallback under degenerate configs

What does NOT change

Backward compatibility

Validation

Related claims

File: `ree_core/environment/causal_grid_world.py`