SD-013: E2_harm_s Interventional Training

Claim ID: SD-013
Subject: self_attribution.e2_harm_s_interventional_training
Status: IMPLEMENTED 2026-04-10
Registered: 2026-03-28
Depends on: SD-003, SD-011, ARC-033
Blocks: SD-003 z_harm_s counterfactual attribution pipeline (causal_sig validity)

Problem

A forward model trained on observational trajectories learns the correlational distribution P(z_harm_s_next | z_t, a), not the interventional distribution P(z_harm_s_next | do(a)) (Scholkopf et al. 2021). These diverge in confounded states: environments where agent presence correlates with ambient harm independently of the agent’s causal contribution.

In such states, E2_harm_s trained purely observationally encodes the correlation, not the causal mechanism. Consequence: causal_sig = E2(z_t, a_actual) - E2(z_t, a_cf) will underestimate the agent’s causal contribution (both terms are biased upward in confounded states, compressing the difference).

The forward model must be trained to produce divergent outputs for different actions, not merely accurate outputs for the observed action. Without this, the model may learn action-invariant harm representations that produce near-zero counterfactual gaps.

Solution

Add a contrastive interventional loss to E2HarmSForward training. For each training step, sample a counterfactual action a_cf != a_actual from the action space. Compute:

z_pred_actual = E2HarmSForward(z_harm_s_t, a_actual)   [standard MSE target]
z_pred_cf     = E2HarmSForward(z_harm_s_t, a_cf)        [counterfactual]

Add a margin loss that pushes the two predictions apart:

interventional_loss = max(0, margin - ||z_pred_actual - z_pred_cf||_2)

This has zero gradient when the predictions are already >= margin apart, and positive gradient when they are too similar. Combined with the standard reconstruction loss, the model learns both accuracy (fit observed transitions) and action-sensitivity (different actions produce different harm predictions).

Config params (E2HarmSConfig)

Param	Type	Default	Purpose
`use_interventional`	bool	`False`	master switch
`interventional_fraction`	float	`0.3`	fraction of batches adding contrastive loss
`interventional_margin`	float	`0.1`	L2 separation margin

New method: `E2HarmSForward.compute_interventional_loss()`

compute_interventional_loss(z_harm_s, a_actual, a_cf) -> Tensor

Computes the margin loss for one training step. a_cf must be a different one-hot action. The caller is responsible for providing a_cf != a_actual and for applying stop-gradient on encoder inputs (same phased training discipline as the standard forward model).

Architecture Context

SD-013 is a training procedure extension to ARC-033 (E2HarmSForward), not a new module. The architecture is unchanged; only the loss function gains an additional term.

E2HarmSForward (ARC-033) already provides counterfactual_forward(z_harm_s, a_cf) for the SD-003 attribution pipeline. SD-013 ensures the underlying model is properly trained to make that counterfactual meaningful.

The phased training protocol from ARC-033 still applies:

P0: HarmEncoder warmup (z_harm_s learns to represent sensory harm)
P1: E2HarmSForward trains on frozen z_harm_s – standard MSE + interventional margin
P2: Evaluation (harm_forward_r2 + causal_sig_gap)

What This SD Enables

SD-003 causal_sig is meaningful: E2(a_actual) - E2(a_cf) reflects genuine causal sensitivity, not noisy near-zero compression artifacts.
Discriminative pair experiments: INTERVENTIONAL vs OBSERVATIONAL conditions can measure whether action-sensitivity of E2 output translates to better causal attribution.

SD-003 (self_attribution.counterfactual_e2_pipeline)
ARC-033 (harm_stream.sensory_discriminative_forward_model)
SD-011 (harm_stream.dual_nociceptive_streams)
MECH-071 (E2 harm calibration asymmetry)

ML/AI Engineering Notes

Engineering problem: Training a forward model to be action-sensitive, not just accurate.

Technique adopted: Contrastive margin loss (Hadsell et al. 2006; SimCLR-style contrastive divergence). The REE adaptation differs from standard contrastive learning: we push apart predictions for different actions from the same state, not different augmentations of the same sample. The biological grounding (Scholkopf 2021 causal identifiability) motivates why action-sensitivity is required, not just accuracy.

Known failure mode: If interventional_margin is set too high, the model may ignore the reconstruction objective and learn action-dependent but inaccurate representations. Mitigation: keep margin small (0.1 « typical z_harm_s scale) so reconstruction still dominates. The contrastive term acts as a regulariser, not the primary signal.

Backward compatibility: All new params default to no-op (False / 0.0). Existing experiments using E2HarmSForward are unaffected.