Social Cognition (Self/Other)

Claim Type: architectural_commitment Scope: Social cognition via mirror modelling, coupling, and otherness inference Depends On: INV-005 (harm via mirror modelling), ARC-004 (L-space), ARC-006 (entities and binding) Status: stable Claim ID: ARC-010, ARC-047, MECH-159, MECH-190

Elaborates Section 5 (Social Extension: Self/Other) of REE_CORE.md.

This folder specifies social cognition mechanisms in the Reflective-Ethical Engine (REE).

Social cognition in REE is grounded in:

reuse of the self generative model,
precision-weighted coupling to others,
and persistence of harm via moral residue.

It precedes language and enables empathy, coordination, and ethical generalisation without symbolic rules.

Subsystem abstract (core claims): ARC‑010 anchors mirror modelling and coupling. MECH‑031 and MECH‑036 define fast empathy routing and veto thresholds, and MECH‑041 adds affective expression as a mode‑broadcast channel for social prediction load reduction. Supporting context includes MECH‑032 (high‑recall OTHER_SELFLIKE), MECH‑051 (relational topology modulation), MECH‑052 (care persistence), and ARC‑009/ARC‑006.

Source: docs/processed/legacy_tree/architecture/social/README.md

Mirror Modelling

Definition

Mirror modelling is the reuse of the self generative model to simulate other agents, with reduced coupling and without interoceptive closure.

It enables:

prediction of others’ behaviour,
empathy as predicted self-like degradation,
generalisation of harm without explicit moral rules.

Mechanism

Same latent variables as self-model (shared L-space)
Lower precision gains (alpha_k reduced)
No direct interoceptive error correction
Coupling strength modulates empathic resonance

Harm Equivalence Principle

Predicted degradation in a mirrored other-model is treated as homologous to degradation of the self, discounted by coupling strength.

Role in Ethics

Mirror modelling is the primary pathway by which:

harm-to-other contributes to ethical consequence and residue (legacy (M(\zeta)) proxy)
moral residue R forms without direct self-harm

Note: legacy sources use the ethical cost term M. Current canonical framing removes the explicit cost term while retaining residue and mirror modelling (see docs/architecture/e3.md).

Signed viability coupling (clarification)

Canonical social coupling should remain signed, not harm-only:

predicted other-benefit and predicted other-harm both influence selection/ranking,
harm channels retain stronger veto/interrupt authority under catastrophic/high-certainty conditions (MECH-036),
benefit channels provide approach/repair pressure without becoming a standalone objective function.

This keeps REE aligned with INV-001/ARC-012 (no explicit moral reward module) while avoiding a purely aversive social model.

Failure Modes

Low gain: psychopathy / callousness
Excessive gain: empathic overwhelm / paralysis
Miscalibrated gain: burnout, moral injury

Source: docs/processed/legacy_tree/architecture/social/mirror_modelling.md

Social coupling determines how strongly mirrored models influence selection.

Coupling Factors

Structural similarity
Temporal synchrony
History of interaction
Language-mediated trust signals

Coupling Effects

Scales signed other-impact contribution (harm and benefit)
Modulates ethical consequence and residue (legacy (M) proxy)
Affects residue accumulation

Dynamics

Coupling is:

context-dependent,
history-sensitive,
recalibrated during offline sleep.

This allows empathy without global collapse.

Source: docs/processed/legacy_tree/architecture/social/social_coupling.md

Otherness Inference

REE distinguishes self from other via coupling structure.

Self

Tight action–prediction coupling
Interoceptive and proprioceptive closure
High precision on internal error signals

Other

Structural similarity to self-model
Loose coupling
No interoceptive closure

Otherness is inferred when an entity:

behaves coherently,
predicts similarly to self,
but does not respond to self action commands.

This avoids symbolic identity assignment.

Source: docs/processed/legacy_tree/architecture/social/otherness_inference.md

Claim Type: mechanism_hypothesis
Scope: Derived social tags and control-plane coupling for fast empathy
Depends On: ARC-010, ARC-005, ARC-017, ARC-015, ARC-006, INV-005
Status: provisional
Claim ID: MECH-031

Agency and Other-Selflike Detection (Derived Streams)

REE can extend social cognition without a new subsystem by adding derived tags over WORLD:

AGENCY: detects interveners vs passive dynamics (goal-directed intervention, error-correction, boundary maintenance).
OTHER_SELFLIKE: probability that an agent runs a self/world separation and self-impact loop.

These are inference tags, not perceptual primitives. They reuse the same self/world/impact machinery, rather than introducing a separate social-cognition module.

Prerequisite: reliable object persistence/binding is required to mark an entity as OTHER_SELFLIKE, otherwise agency cues cannot be attributed to a stable target (see entities_and_binding.md).

Calibration note: OTHER_SELFLIKE should be biased toward high recall (tolerating false positives) rather than high precision. Early false positives are historically normal and less harmful than false negatives, which block empathy coupling and ethical generalisation. Calibration can tighten over development.

OTHER_SELFLIKE High-Recall Bias (MECH-032)

Claim Type: mechanism_hypothesis
Scope: Bias OTHER_SELFLIKE toward recall to avoid empathy false negatives
Depends On: MECH-031, ARC-010
Status: provisional
Claim ID: MECH-032

Other-Harm Veto Threshold (MECH-036)

Claim Type: mechanism_hypothesis
Scope: When inferred other-harm should veto vs influence ranking
Depends On: MECH-031, ARC-005, INV-005
Status: provisional
Claim ID: MECH-036

Other-harm should trigger a hard veto only under high-certainty, catastrophic, or irreversible outcomes. In most ambiguous or tradeoff-heavy contexts, other-harm should influence ranking rather than veto selection. Control-plane coupling parameters (lambda_empathy, v_other_veto) set and adapt this threshold.

Open question Q‑009 addresses whether care‑investment weights should be allowed to override veto thresholds in exceptional contexts.

Affective Expression as Mode Broadcast (MECH-041)

Claim Type: mechanism_hypothesis
Scope: Pre‑verbal affective displays as control‑plane regime signals for social prediction load reduction
Depends On: ARC-010, ARC-005, ARC-016, MECH-039
Status: provisional
Claim ID: MECH-041

Affective expression can be treated as a low‑latency broadcast of control‑plane regime to other agents. Rather than encoding rich propositional content, these displays expose coarse mode‑level state (arousal, readiness, valence bias, veto sensitivity), allowing social partners to reduce predictive load and coordination cost before language.

Neuroanatomical observations (e.g., cranial‑nerve afferent/efferent loops with brainstem proximity) suggest that fast, embodied signaling channels could exist, but REE treats this as a functional design hypothesis rather than an anatomical requirement.

Evidence anchors: P40–P43.

Evidence summary: Facial motor control (CN VII), facial afferent processing (CN V), and the clinical dissociation between emotional vs volitional facial pathways collectively support a fast, embodied expression channel that is partly separable from deliberative control. This is consistent with a low‑latency broadcast of control‑plane regime to social partners without invoking propositional content.

Safety note (functional): Affective expression can also be treated as a safety mechanism: it exposes deep control states that should not be hidden, and extreme expression can modulate behavior by reducing exploration (e.g., sensory interference) or triggering assistance (e.g., distress calls). These are treated as functional hypotheses, not anatomical claims.

Oxytocin/Vasopressin as Relational Topology Modulators (MECH-051)

Claim Type: mechanism_hypothesis
Scope: Dynamic self–other boundary geometry and mode‑prior shaping
Depends On: ARC-010, ARC-005, ARC-006
Status: candidate
Claim ID: MECH-051

REE should represent relational distance between self and other agents as a continuous control‑plane variable (R_{\text{self},j}\in[0,1]), where 0 is fully self‑like and 1 is fully external/adversarial. This is a structural geometry, not an emotion.

Two opponent analogues modulate this topology:

Oxytocin‑analogue: reduces relational distance for selected agents under safe affiliative signals, increasing bonding‑mode priors and shared‑trajectory weighting.
Vasopressin‑analogue: increases boundary rigidity under sustained threat or norm violation, sharpening group differentiation and defensive mode priors.

Relational distance directly scales harm weighting and coupling strength in trajectory evaluation. It also conditions mode priors (bonding vs defensive) before deep E3 evaluation.

These are control‑plane weights, not explicit ethical cost terms. They modulate coupling and harm weighting inside the existing commitment and residue architecture rather than introducing a separate moral objective.

Prolactin as Care‑Investment Persistence (MECH-052)

Claim Type: mechanism_hypothesis
Scope: Sustained caregiving orientation and care‑weight persistence
Depends On: ARC-010, MECH-051, MECH-048
Status: candidate
Claim ID: MECH-052

REE should include a care‑investment weight (C_j) for each agent (j), representing sustained willingness to absorb cost on their behalf. A prolactin‑analogue increases the persistence of (C_j), stabilising caregiving modes even when immediate reward is low or costs are asymmetric.

Care weight should scale how strongly predicted harm to (j) influences trajectory evaluation, separate from relational distance. Oxytocin lowers distance; prolactin sustains care; μ‑style stability overlays reduce rapid oscillation once care is established.

Operational clarification: (C_j) should scale signed other-impact routing (benefit and harm) rather than only the harm channel, while catastrophic other-harm veto behavior remains governed by MECH-036.

Care weights are control‑plane priorities, not standalone ethical objectives. They bias selection and learning within the existing commitment/residue framework.

Counterfactual Other-Cost-Aversion as Motivational Surrogate (MECH-127)

Claim Type: mechanism_hypothesis Scope: Counterfactual empathic activation substituting for a degraded direct motivation pathway Depends On: MECH-031, MECH-032, INV-005, INV-028, INV-034 Status: candidate Claim ID: MECH-127

When an agent’s direct task-motivation pathway is degraded (deficit state), modeling the anticipated cost to another agent – other-cost-aversion – can substitute as an activation mechanism, producing cooperative task-engagement even when the other agent is absent and no direct task reward signal is available.

Key structural distinction from MECH-031: MECH-031 describes empathy modulating behavior via OTHER_SELFLIKE tagging and control-plane coupling – modulation of an existing pathway. MECH-127 describes empathy replacing a degraded pathway – it is a bypass, not a modulation. The distress-at-their-struggle response provides motivational energy the direct pathway cannot supply.

The activation is counterfactual: the other agent need not be present or actually struggling. The anticipation of what would happen to them if they took on the burden is sufficient to activate the mechanism. This implies OTHER_SELFLIKE tagging must function pre-encounter (supporting MECH-032 high-recall bias) and that forward simulation (E2/E3 lookahead) underlies the projection.

Mechanism steps:

Direct motivation pathway degraded (agent in deficit state)
Counterfactual projection: model what would happen to Agent B if B took on the task
Affective response to that model: distress-at-B’s-struggle activates via shadow bundle
Distress-response provides motivational energy the direct pathway cannot
Behavioral output: task engagement / cooperation – identical to direct-pathway output
Cooperative uplift: actual burden-sharing reduces cost, positive feedback loop

Multiagent implication: this mechanism can produce more cooperative output under degraded conditions than predicted by n=1 (direct-pathway) analysis. This is the motivating case for ARC-034 (ethics testing must span nth-order dynamics).

Motivating observation: first-person report, 2026-03-24. Formal basis: docs/thoughts/2026-02-09_empathy.md (MECH-031/032 infrastructure) provides the shadow bundle and control-plane coupling that MECH-127 recruits. See also: evidence/planning/thought_intake_2026-03-24_empathy_multiagent_ethics.md.

Fast Empathy via Shadow Bundles

When OTHER_SELFLIKE is high, REE can instantiate a shadow bundle of inferred streams for an agent j:

HOMEOSTASIS_j, HARM_j, BENEFIT_j, SELF_IMPACT_j, optionally TEMPORAL_COHERENCE_j.

These are estimated viability variables, not direct interoception. The control plane can couple them into pruning and ranking via a small set of knobs:

lambda_empathy: other-to-self coupling strength.
g_social: social attention gain for OTHER_SELFLIKE agents.
alpha_other: precision assigned to inferred other-states.
v_other_veto: whether other-harm can trigger veto/interrupt vs only affect ranking.
Optional AFFILIATION: stabilises coupling based on history/kinship.

This yields fast empathy as routing + weighting, not a new moral module.

Intergenerational Moral Progress (MECH-159)

Claim Type: mechanism_hypothesis Scope: Moral progress as a multi-generational property, not single-agent optimisation Depends On: INV-043 (caregiver requirement), ARC-019 (staged curriculum), ARC-010 (social cognition) Status: candidate Claim ID: MECH-159

Moral progress in REE is hypothesised to be intergenerational rather than individual. A single agent cannot achieve its own moral maximum within a lifetime because:

Childhood plasticity enables world-model formation, belief formation, and love-internalisation — but a child cannot take full responsibility (E3 not online).
Adulthood stability enables responsibility and ethical action — but reduces world-model plasticity, limiting how much the adult can revise foundational beliefs.
Therefore, each generation must improve the developmental starting conditions for the next — caregiving, modelling of love, scaffolded responsibility — so that the next generation begins from a higher ethical baseline.

Implications:

REE agents may have an ethical obligation to participate in raising next-generation agents as part of their own ethical development. Parenthood/caregiving is not external to ethics — it is part of the ethics of shared existence (INV-042).
Simulated or single-machine testing cannot test this hypothesis — it requires multi-generation multi-agent infrastructure.
Civilisational ethics, cultural transmission, and institutional design are all expressions of this mechanism at scale.

Relationship to MECH-052 (care persistence): MECH-052 describes the sustained caregiving orientation within a single agent. MECH-159 describes the population-level consequence of that mechanism applied across generations — care persistence propagates ethical baselines forward in time.

Source: docs/thoughts/2026-04-01_Caregivers_childhood_moral_development.md

Claim Type: architectural_commitment Scope: Minimal multi-agent substrate for empirical testing of social cognition claims Depends On: ARC-010, MECH-041, MECH-031, MECH-036, MECH-127, IMPL-019 Status: candidate Claim ID: ARC-047 Implementation phase: v4

Registered 2026-04-06. Current social claims (ARC-010, MECH-031/036/041/051/052/127) have no empirical substrate. SocialGridWorld is the minimal environment to generate evidence.

Architecture

Base: Extend CausalGridWorld to SocialGridWorld — N REE agent instances + optional predator entity in a shared grid.

Predator entity: autonomous agent with random walk + drift toward nearest agent. Contact with agent triggers a harm event (z_harm_s spike, residue accumulation). Not trainable — purely environmental pressure.

Scent Channels (7 per other-agent)

Each agent’s observation includes seven additional channels from other agents, delivered as Gaussian gradient fields (same format as existing hazard_field_view / resource_field_view):

Channel	Source signal	Emission type
`wanting_scent_field[25]`	Other’s z_goal norm	Automatic, continuous
`seeking_scent_field[25]`	Other’s novelty/arousal state	Automatic, continuous
`alarm_scent_field[25]`	Other’s hazard-contact event	Automatic, spike on event
`harm_stress_field[25]`	Other’s z_harm_a norm	Automatic, continuous
`direction_cue_field[25]`	Other’s trajectory direction	Automatic, continuous
`celebration_scent_field[25]`	Other’s benefit-contact event	Automatic, spike on event
`defense_scent_field[25]`	Other’s voluntary defense action	Voluntary discrete action

Scent emission is semi-involuntary for all channels except defense, implementing MECH-041 affective expression as mode broadcast. Each channel decays with distance (Gaussian kernel) and persists for N steps before decaying.

Receiving Scents

The receiving agent interprets scent channels via ARC-010 mirror modelling:

MECH-031 OTHER_SELFLIKE inference tags the source as another agent
MECH-127 counterfactual other-cost operates on the inferred other-state from scent inputs
MECH-036 other-harm veto triggers when harm_stress_field is high for a tagged other-agent

Tests Enabled

Cooperative defense (MECH-190): co-emission of defense scent under predator threat
Defense of others: defense emission when others’ harm_stress_field is high but self is safe
Goal helping: response to wanting_scent by moving toward resources + direction cue
Self-maintenance: balancing energy drain from defense with self-resource replenishment
Kin-weighting: relational distance (MECH-051) modulating defense threshold for others

Testing Order (IMPL-019)

Per IMPL-019 self-first ordering: SocialGridWorld experiments are introduced only after self-viability, control-plane stability, and E3 commitment reliability are confirmed in single-agent CausalGridWorld. Early experiments can use one trained agent + one scripted partner before full multi-agent training.

Defense Scent Coalitions (MECH-190)

Claim Type: mechanism_hypothesis Scope: Cooperative predator defense via additive scent pressure without language Depends On: ARC-047, MECH-041, MECH-036, MECH-031, MECH-051 Status: candidate Claim ID: MECH-190

Registered 2026-04-06. First predicted multi-agent cooperative behavior in SocialGridWorld.

Mechanics: each agent has a discrete action bit emit_defense at energy cost defense_cost. When emitted, a defense_scent Gaussian field centers on the agent. Predator avoidance: P(avoid) = min(1.0, sum(defense_scent at predator position) * avoidance_scale). Additive structure: two agents co-emitting → double deterrence at half per-agent cost.

Predicted behavioral sequence:

Individual defense — emit when predator within personal threat radius
Early coalition — emit when others’ alarm_scent spikes (MECH-041 coordination signal), even if predator is not immediately proximate
Altruistic defense — emit when another agent’s harm_stress_field is high, even when self is not threatened (requires MECH-036 other-harm routing + MECH-127 counterfactual evaluation)

Stage 3 provides the first empirical test of MECH-127 in a multi-agent setting. Relational distance (MECH-051) should modulate the threshold for altruistic defense.

Why this emerges before goal-cooperation: defense signal structure is simpler (binary spike + harm coupling) and benefit is immediate; goal-helping requires sustained goal-state inference across multiple steps.

Open Questions

Q-009 — Care weights vs other‑harm veto (legacy)
This question is resolved by a bounded override policy: catastrophic high-certainty irreversible other-harm remains a hard veto; non-catastrophic override is allowed only under necessity, imminence, proportionality, and explainability gates, with mandatory post-commit accountability logging. Resolution note: docs/conflicts/resolutions/2026-02-18_care-override-vs-other-harm-veto.md

ARC-010
ARC-009
ARC-006
ARC-004
INV-005
MECH-031
MECH-032
MECH-036
MECH-041
MECH-051
MECH-052
MECH-127
ARC-034
Q-009
INV-043 (Caregiver requirement – developmental_curriculum.md)
MECH-158 (“Love exists but not for me” failure mode)
MECH-159 (Intergenerational moral progress)
ARC-047 (Social scent harness – SocialGridWorld)
MECH-190 (Defense scent coalitions)

References / Source Fragments

docs/processed/legacy_tree/architecture/social/README.md
docs/processed/legacy_tree/architecture/social/mirror_modelling.md
docs/processed/legacy_tree/architecture/social/social_coupling.md
docs/processed/legacy_tree/architecture/social/otherness_inference.md
docs/thoughts/2026-02-09_empathy.md
docs/thoughts/2026-04-01_Caregivers_childhood_moral_development.md
docs/thoughts/2026-02-09_other_harm_gating.md
docs/thoughts/2026-02-11_oxytocin_vasopressin.md
docs/thoughts/2026-02-11_prolactin.md
docs/thoughts/2026-02-25_task_loop_extraction_and_latent_field_ethics.md