Emotion as anti-collapse architecture (ARC-088)
Status: candidate architectural framing (MAP, not new substrate) Registered: 2026-06-09 Source thought: docs/thoughts/2026-06-08_emotion_as_anti_collapse_architecture.md Claim: ARC-088 (docs/claims/claims.yaml)
What this claim is
ARC-088 is a unifying map, not a new substrate. It names an organising principle over affect machinery REE already owns: emotion-like systems are a set of gated, partially independent evaluators whose collective effect is to prevent behaviour from collapsing onto the currently dominant gradient (anti-monostrategy). Affect, on this reading, is the architectural source of structured behavioural diversity — not decoration added to a rational optimiser, and not random exploration noise.
A pure optimiser asks one question: what action scores best? An organism runs many gated, partially independent questions in parallel:
Is this harmful? Is this safe? Did this reduce suffering?
Am I pursuing something Did I merely like the Am I stuck?
I actually want? last thing?
Am I overcommitted? Is this goal stale? Should I explore?
Should I stop? Should I sleep/consolidate? Should I release the policy?
That plurality is not inefficiency. It is how the system avoids becoming stupidly efficient in one direction — the smoothest-single-regime collapse MECH-309 predicts and ARC-065 exists to prevent. Each evaluator is a structured deviation from naive optimisation (“do not chase food through fire”, “do not freeze forever”, “do not act forever when safe”, “do not trust stale goals”, “do not collapse all value into one scalar”), not added noise.
Why it is a map and not a substrate claim
Every constituent evaluator is already a registered REE claim. ARC-088 does not re-implement any of them; it asserts that, taken together, they constitute REE’s anti-collapse / behavioural-diversity architecture, and that an optimise-then-bolt-on-safety design fails precisely because these evaluators arrive too late to be native control systems. This is the same shape as the attention = distributed precision-selection map and ARC-084’s signed-coupling map: REE owns the pieces; the new thing is the explicit organising claim.
epistemic_category: substrate_coherence is set explicitly so promote/demote stays suppressed — consistent with the source thought’s governance note that the intake validates no claim, promotes no substrate, and alters no experiment routing. Conflict-resolution alerts still fire.
What ARC-088 depends on (the already-owned evaluator set)
| Evaluator family (“the organism asks…”) | Owned claim(s) | Role |
|---|---|---|
| Prevent collapse to the smoothest single regime | ARC-065 (+ children MECH-313 noise-floor, MECH-314 structured-curiosity, MECH-320 tonic-vigor) | the behavioural-diversity generation pathway, upstream of both rule pathways |
| Affect must be candidate-differentiated to carve behaviour | MECH-359 (candidate-differentiated affect vector) | a scalar added equally to all K candidates cannot move an argmax; only per-candidate range can |
| Is this harmful? | SD-011 (dual nociceptive streams; SD-010 stream separation) | harm as a native, separable stream |
| Did this reduce suffering? | SD-050 + MECH-302 (suffering-derivative comparator → relief-completion event) | relief as action-contingent harm reduction |
| Is this safe? | MECH-304 (conditioned safety store; MECH-303 contextual) | threat-absence prediction / commitment-release gate |
| Am I pursuing something I want? | SD-012 (homeostatic drive) | wanting / deficit-keyed drive |
| Am I overcommitted? Is this goal stale? Should I release? | ARC-079 (gated goal persistence; disengagement the ungated default) | persistence/release regulation |
These eight anchors are representative, not exhaustive: blocked-agency, curiosity sub-flavours (MECH-314a/b/c), boredom (MECH-330/ARC-067), liking-vs- wanting, and the “should I sleep/consolidate” offline-integration evaluators (MECH-272/273 sleep cluster) all sit under the same umbrella and are reachable through the depends_on chain rather than re-listed.
The 603 lineage as a worked example
The 603 lineage makes the principle concrete. The agent could form goals, experience threat, and suppress freezing — but it did not learn directed escape until the missing affective-learning bridge was identified: relief as action-contingent harm reduction, safety as threat-absence prediction, both feeding a bounded future-action bias. “Optimise away from harm” was not enough; the system needed structured emotional mechanisms that bind events to actions, contexts, release conditions, and future candidate selection. That is exactly the anti-collapse function ARC-088 names.
Architectural consequence (do NOT merge to one scalar)
The design implication is not to collapse all emotions into a single value scalar — the opposite. REE should preserve distinct affective streams while letting them share interfaces where appropriate. Relief, safety, wanting, liking, harm, blocked agency, curiosity, salience, and commitment release are partially independent control systems: they may converge on shared consumers (E3 score bias, commitment gating, residue, offline consolidation) but their learning targets and gating conditions must stay distinct. Reuse/gating audits that touch any of these convergence points should reference ARC-088 to check that distinctness is preserved.
Visitor-facing framing layer (documentation, not a substrate claim)
This section is the public-facing articulation. It is documentation; it registers no additional claim.
Current artificial intelligence often begins from optimisation: define an objective, maximise it, then add safety constraints afterward. REE begins from a different premise. A viable agent is not a single optimiser with a safety wrapper; it is a set of partially independent control systems that prevent any one gradient from taking over behaviour.
Emotion-like systems are central to that architecture. Harm, relief, safety, wanting, liking, curiosity, blocked agency, salience, and commitment release are not decorative feelings. They are gates, biases, interrupts, and learning signals that shape what the agent notices, what it repeats, what it avoids, when it persists, and when it stops.
This is why REE treats emotion as a source of behavioural diversity rather than irrational noise. Emotion prevents behavioural collapse: the agent does not simply chase food through fire, freeze forever, optimise one scalar, or keep acting when safe. It can switch, release, recover, explore, and remember consequence.
A one-line summary for both audiences:
Emotion is not an obstacle to optimisation; it is how optimisation is prevented
from becoming pathological. Emotion systems are anti-collapse architecture.
Scope and guardrails
- V3-grounded framing. The harm (SD-010/SD-011), relief (SD-050/MECH-302), safety (MECH-303/304), drive (SD-012), and behavioural-diversity (ARC-065 stack) pieces are V3 substrate; the candidate-differentiated extension (MECH-359) is V4. ARC-088 spans them as a map.
implementation_phase: v3reflects that the anti-collapse architecture is substantially realised now and the framing is visitor-facing today; the V4 reach is via depends_on, not a permission gate. - No promotion/demotion, no substrate code, no experiment follows from the source thought’s governance note. ARC-088 informs documentation language and future reuse/gating audits; it does not move any claim’s confidence.
- Do not invent redundant substrate claims under this umbrella. New affect work attaches to the relevant owned child (or extends MECH-359), not to ARC-088.