SD-019: Affective Harm Non-Redundancy Constraint

Claim ID: SD-019
Subject: harm_stream.affective_nonredundancy
Status: IMPLEMENTED 2026-04-10
Registered: 2026-04-08
Depends on: SD-011, SD-010, ARC-016
Blocks: SD-020 (requires non-redundant z_harm_a before PE-weighting makes sense)

Problem

SD-011 established the dual-stream split (z_harm_s sensory-discriminative, z_harm_a affective-motivational) in principle. EXQ-241/241a showed that z_harm_a remained too close to z_harm_s functionally (D3 reversal: z_harm_a predicted the sensory target better than z_harm_s because both received the same spatial signal).

The SD-011 harm-history input (second source) reduces the correlation structurally, but does not actively push the streams apart. Without an explicit non-redundancy constraint, nothing prevents z_harm_a from collapsing to a delayed/smoothed copy of z_harm_s.

Solution

Add a cosine-squared penalty between z_harm_s and z_harm_a to the training loss:

penalty = cosine_similarity(z_harm_s, z_harm_a)^2
loss += harm_nonredundancy_weight * penalty

cos^2 penalises both positive and negative alignment symmetrically. Gradient is zero when the streams are orthogonal; maximal when they are parallel (redundant).

ARC-016 coupling

When harm_nonredundancy_precision_scale > 0, the penalty is scaled by normalised E3 precision:

precision_norm = min(e3.current_precision / 500.0, 2.0)
penalty *= (1 + harm_nonredundancy_precision_scale * precision_norm)

Higher precision (more confident state) → stronger non-redundancy enforcement. Biological basis: in a stable, well-understood environment, the two harm streams should be maximally informative (least redundant); in a volatile/uncertain environment, some redundancy is tolerated (precision is low).

Config params (REEConfig)

Param	Type	Default	Purpose
`harm_nonredundancy_weight`	float	`0.0`	loss weight (0 = disabled, backward compat)
`harm_nonredundancy_precision_scale`	float	`0.0`	ARC-016 coupling strength

New agent method: `compute_harm_nonredundancy_loss(latent_state)`

Returns the penalty scalar. Returns 0 when:

harm_nonredundancy_weight=0.0 (default)
Either harm stream is absent from the latent state

Architecture Context

Operates on z_harm_s (from HarmEncoder) and z_harm_a (from AffectiveHarmEncoder) after encoding. No new modules — purely a loss-level constraint.

If z_harm_s dim != z_harm_a dim, the method projects to the smaller dimensionality (first min_dim elements) before computing cosine similarity.

What This SD Enables

Structurally valid z_harm_a before SD-020 (precision-weighted PE) can be added
EXQ-319 SD-022 dissociation test (HIGH_DAMAGE vs FRESH) depends on streams being distinct
SD-020 becomes scientifically meaningful only after SD-019 reduces baseline redundancy

Biological Grounding

Barlow (1961) redundancy reduction hypothesis: efficient neural coding minimises redundancy between channels. A-delta (sensory-discriminative) and C-fiber (paleospinothalamic, affective-motivational) pathways project to distinct cortical regions (S1/S2 vs ACC/anterior insula) via distinct thalamic nuclei (VPL/VPM vs intralaminar CM/PF). Cross-stream redundancy is biologically implausible for functionally distinct channels serving different computational purposes.

ML Engineering Notes

Technique: decorrelation loss (Barlow Twins 2021 uses a related approach). Standard contrastive approach would push representations apart in norm-space; cos^2 is more appropriate here because it is scale-invariant and penalises alignment, not distance.

Known failure mode: aggressive decorrelation collapses useful shared structure. Mitigation: cos_sim^2 penalises correlation, not dissimilarity — the streams can have the same norm while being orthogonal. harm_nonredundancy_weight=0.0 default ensures backward compat.

SD-019 (harm_stream.affective_nonredundancy)
SD-011 (harm_stream.dual_nociceptive_streams)
SD-020 (harm_stream.affective_surprise_pe — depends on SD-019)
ARC-016 (cognitive_modes.control_plane_regimes — precision coupling)
MECH-219, Q-036