SD-030: E2_self Motor-Proprioceptive Forward Model Comparator

Claim ID: SD-030 Subject: self_attribution.comparator_z_self Status: CANDIDATE (V4-scoped) Registered: 2026-04-18 Depends on: SD-005 (self/world latent split), MECH-256 (general comparator mechanism) Instantiates: MECH-256 on the z_self stream Parent roadmap: self_attribution_per_stream.md

Problem

V3 operates a single forward-model comparator on z_harm_s (SD-029, ARC-033). That instantiation addresses agency attribution for nociceptive consequences – “did I cause this pain?” – which is the highest-priority stream for moral attribution and the one where biology most clearly demands a working comparator (descending pain modulation, self-inflicted pain attenuation).

It does not address the two other reafferent streams on which the biological forward-model architecture is at least as well-established:

  • z_self (motor-proprioceptive) – “did I cause the movement I just felt?” The cerebellar internal-model loop. Attenuation of self-generated tactile input (Blakemore 1998), force-matching underestimation (Shergill 2003), saccadic corollary discharge (Sommer & Wurtz).
  • z_world (causal-footprint) – “did I cause the world-state change I just observed?” Covered by SD-031.

SD-030 exists to make the topology explicit: the comparator mechanism (MECH-256) applies uniformly across streams; z_self is the stream with the strongest biological grounding for forward-model machinery. Without SD-030 on the roadmap, motor-proprioceptive self-attribution would have no substrate claim, and the cerebellar internal-model loop – the canonical forward-model system in neuroscience – would have no home in the REE architecture.

V3 does not run SD-030 because V3 does not yet have a first-class z_self stream with its own forward model. E2 currently operates on z_gamma (combined self-world), with a split into z_self and z_world per SD-005 that has not yet been instantiated with dedicated stream-specific E2 substrates. SD-030 is deferred to V4 alongside the z_self substrate materialisation.

Solution

E2SelfForward: a dedicated learnable forward model f(z_self_t, a_t) -> z_self_{t+1}. Structural analogue of E2HarmSForward (ARC-033): residual-delta architecture, stop-gradient on z_self inputs during forward-model training, small learning rate, phased training (encoder warmup -> frozen-encoder forward model).

Single-pass comparator at inference:

predicted_z_self = E2SelfForward(z_self_{t-1}, a_actual)
residual_z_self  = z_self_observed - predicted_z_self
causal_sig_self  = f(residual_z_self, context)

No counterfactual query. The prediction on the actually-taken action is the reference; the residual is the agency signal.

Why the structure mirrors ARC-033 rather than adding a new mechanism class. MECH-256 asserts that the comparator is stream-agnostic computationally. SD-030 is therefore an instance, not a variant. The engineering differences from ARC-033 are parameter-level (z_self has different dimensionality, different autocorrelation profile, different action-conditional structure), not architectural.

Architecture Context

E2SelfForward is distinct from E2FastPredictor in the same way E2HarmSForward is: it operates on its own latent stream, trained with its own loss, with stop-gradient on the input latent to prevent encoder drift. In V4, three parallel forward models coexist:

  • E2HarmSForward (ARC-033) – on z_harm_s, V3-active
  • E2SelfForward (SD-030) – on z_self, V4
  • E2WorldForward (SD-031) – on z_world, V4

MECH-257 asserts that each of these substrates is read in two modes: retrospective comparator (attribution) and prospective rollout-scoring (evaluation). SD-030 inherits the dual-mode reading. The V3 evaluator on z_world (score_trajectory in E3) is the prospective mode of a not-yet-existing E2WorldForward; the same pattern applies to E2SelfForward.

Training protocol (phased, inherited from ARC-033):

  • P0: z_self encoder warmup (body-state supervision; currently a single-layer MLP per SD-005, DR-13 flagged this for expansion in V4)
  • P1: E2SelfForward trains on frozen z_self (stop-gradient on z_self_t inputs)
  • P2: Evaluation (forward_r2, self_causal_gap on scripted self-vs-external perturbation trials)

What This Enables

Motor-proprioceptive self-attribution. The REE analogue of the question “did my body just move because I willed it, or was I pushed?” The residual distinguishes self-generated from externally-imposed body-state changes.

Extends descending modulation beyond the harm stream. V3’s SD-021 pathway attenuates z_harm_s during commitment because the comparator expects cancellation (self-inflicted pain attenuation). The z_self stream has an even stronger case for the same pattern: voluntary movement produces predictable proprioceptive input, which is attenuated relative to externally-induced movement. The cerebellar-S1/S2 cancellation (Blakemore 1998) is the canonical biological signature.

Precondition for MECH-215 capacity estimation. DR-12 in the V3/V4 boundary notes that E3 currently trusts E2’s rollout unconditionally. When the forward model on the agent’s body is degraded – injury, fatigue, reduced motor capacity – the residual should increase (the prediction fails even though the action was self-generated). A working E2SelfForward is the substrate through which E3 can learn “this self-transition is unreliable” and discount trajectories accordingly.

Precondition for z_self-domain goals. DR-11 flags z_self-domain goal representation (“I want my energy restored”, “I want to not be in pain”) as structurally inaccessible to V3. A working E2SelfForward is required before E3 can plan trajectories in z_self space – the evaluator mode of MECH-257 on this substrate is the mechanism that scores candidate action sequences by their predicted self-state trajectories.

Module Location (V4 target) 

# Planned location (mirrors ARC-033 pattern):
# from ree_core.predictors.e2_self import E2SelfForward, E2SelfConfig
#
# Enabled via LatentStackConfig.use_e2_self_forward (False default).
# Config dataclass E2SelfConfig standalone in e2_self.py.
#
# Integration point in agent.py:
#   latent_state.z_self_pred = e2_self(latent_state.z_self_prev, action_onehot)
#   self_residual = latent_state.z_self - latent_state.z_self_pred

No V3 implementation exists. V4 milestone: following SD-005 z_self expansion (DR-13: recurrent z_self encoder or E1 feedback into z_self), register a dedicated ree_core/predictors/e2_self.py module.

ML Engineering Notes

Inherited from ARC-033 pattern:

  • Stop-gradient on z_self inputs to forward model during P1 (prevent encoder drift – identity collapse is the default failure mode on autocorrelated streams).
  • Residual-delta architecture (not direct next-state prediction). z_self is strongly autocorrelated (body state changes slowly relative to observation cadence); a direct predictor collapses to identity. Predicting the delta forces the network to learn change structure.
  • Small learning rate (5e-4 initial, matching E2HarmSForward; may need tuning for the z_self dim and typical motion-induced delta magnitude).
  • Phased training required: P0 encoder warmup with body-state supervision, P1 frozen-encoder forward model training, P2 evaluation.
  • Interventional training (SD-013 analogue): once P1 baseline works, consider margin loss forcing ||E2_self(z_self, a_i) - E2_self(z_self, a_j)||

    = margin for a_i != a_j. Prevents observational confound where ambient correlations let the model predict z_self_{t+1} without using the action.

  • MECH-094: not applicable (waking forward model, not replay content).

New to SD-030 relative to ARC-033:

  • Temporal depth constraint (DR-13). If V4 z_self encoder is still a single-layer MLP, the forward model inherits a flat input. Recurrence in z_self (LSTM or E1-feedback into z_self enrichment) should precede SD-030 implementation. Otherwise E2_self has nothing to predict from.
  • Body-frame action encoding. Actions in CausalGridWorldV2 are 4-direction one-hot. For z_self, the relevant representation is the motor command in body frame (which limb, which direction, what force). V4 environment extension may need richer action encoding before E2_self can saturate its capacity.

Biological Grounding

Cerebellar internal models (Wolpert, Miall, Ito). The canonical forward-model locus. Wolpert & Miall 1996 (Neural Networks): the cerebellum maintains an internal forward model of the body that predicts sensory consequences of motor commands. Ito 2008 (Nat Rev Neurosci): cerebellar microcomplexes as learned forward models.

Self-generated tactile attenuation (Blakemore 1998, Nat Neurosci). When subjects self-tickle vs are tickled by another, the self-generated input is perceived as less intense. fMRI shows reduced S1/S2 activation for self-generated tactile input. The attenuation requires predictability – delayed or trajectory-perturbed self-touch reinstates the full response. This IS the E2_self comparator running on the tactile substream of z_self.

Force-matching underestimation (Shergill 2003, Science). Subjects asked to match a force applied to their finger systematically over-apply force when generating it themselves. The forward model cancels the expected sensory consequence of self-generated force, so the self-generated force feels weaker than the external reference at the same physical magnitude. Schizophrenia patients with passivity experiences show reduced attenuation (Shergill 2005, Am J Psychiatry) – consistent with Frith 2000’s comparator-failure framing.

Saccadic corollary discharge (Sommer & Wurtz). Superior colliculus -> MD thalamus -> parietal cortex corollary discharge pathway. Parietal neurons remap receptive fields in anticipation of each saccade; the remapping IS the forward prediction. Lesions to MD thalamus disrupt remapping, producing behavioural deficits in saccade-target-directed tasks. Oculomotor is a sub-domain of z_self in the REE mapping.

Songbird HVC corollary discharge (Crapse & Sommer 2008 review). During singing, efference-copy from HVC-RA motor path projects to auditory cortex and cancels expected auditory feedback. Birds with lesioned corollary discharge cannot maintain song quality because they cannot distinguish self-generated from external sound. Self-vocalization is a distinct z_self sub-stream.

Electrosensory cancellation (Bell, ELL). Weakly electric fish cancel the reafferent signal from their own electric organ discharge via granule-cell negative images in the electrosensory lobe. The cancellation is adaptive (learned) and disrupted by transecting the corollary-discharge input. Perfect-model-organism demonstration of the comparator pattern on a sensory stream with an agent-generated reafferent signal.

Biological grounding for SD-030 is the strongest of the three per-stream instantiations. ARC-033 has good nociceptive-agency evidence (ACC/insula); SD-031 has moderate evidence (parietal outcome-attribution, MD thalamus); SD-030 has five distinct preparations across mammals, birds, and fish all converging on the same mechanism. This is the core of why the comparator was generalised from SD-003 (z_world-only, counterfactual) to MECH-256 (stream-agnostic, single-pass).

Validation Experiment (V4)

Proposed design (not queued for V3):

Phase P0: z_self encoder warmup. Supervised on body-state targets from the environment (energy, limb_damage, residual_pain per SD-022). Verify z_self stream is discriminative before forward-model training.

Phase P1: E2SelfForward training. Frozen z_self encoder, stop-gradient on z_self_t inputs. Metric: self_forward_r2 > 0.85 on held-out trajectories. Identity-collapse check: r2 on action-shuffled control must drop below 0.3 (i.e., the model is using the action, not just autocorrelation).

Phase P2: Attribution evaluation. Scripted trials with self-initiated movement vs externally-imposed perturbation (environment extension: a “pushed” condition that modifies body state independently of the agent’s action). Residual magnitude must be substantially lower for self-initiated than for pushed. Metric: self_causal_gap = mean(||residual||_pushed) - mean(||residual||_self) > 0.1.

Interventional extension (SD-013 analogue): after P1 baseline, enable margin loss and retest. Prediction: ambient correlations in body state compress the causal signal under pure observational training; interventional training restores it.

Ablation: with SD-030 disabled, trajectory selection in E3 cannot distinguish “reliable body transition” from “unreliable body transition”. V4 experiment is a goal-directed task where fatigue/injury degrades a subset of actions; agents with working E2SelfForward + DR-12 confidence modulation should avoid the degraded actions; ablated agents should not.

  • MECH-256: general comparator mechanism (parent)
  • MECH-257: dual-function gated readout (comparator + evaluator on same substrate)
  • SD-005: self/world latent split (prerequisite – z_self must exist as a first-class stream)
  • SD-029: z_harm_s comparator (V3-active sibling)
  • SD-031: z_world comparator (V4-deferred sibling)
  • ARC-033: E2_harm_s forward model (reference implementation template)
  • SD-013: interventional training extension (applied later)
  • DR-11, DR-12, DR-13: V4 boundary items this SD unblocks
  • MECH-215: self-model prerequisite for agentive prediction (downstream)
  • MECH-214: goal-referent E1-representability for z_self domain (downstream)

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