Operon v0.22: The Cognitive Architecture

1. Cognitive Modes: System A/B as Stage Annotations

Operon’s fast/deep nucleus distinction started as cost optimization: route cheap work to a small model, expensive work to a large one. v0.22 reframes this as a cognitive architecture principle. The CognitiveMode enum classifies stages as OBSERVATIONAL (System A — passive sensing, information gathering) or ACTION_ORIENTED (System B — active decision-making, goal pursuit).

from operon_ai import CognitiveMode, SkillStage

SkillStage(name="classifier", role="Router", mode="fast",
           cognitive_mode=CognitiveMode.OBSERVATIONAL)

SkillStage(name="planner", role="Strategist", mode="deep",
           cognitive_mode=CognitiveMode.ACTION_ORIENTED)

When not set, cognitive mode is inferred from the existing mode field (fast → observational, deep → action-oriented). The watcher detects mismatches between declared mode and actual execution model, and mode_balance() reports the System A/B distribution across a run.

2. Sleep Consolidation: When the Organism Dreams

The SleepConsolidation class composes five existing components — AutophagyDaemon, PatternLibrary, EpisodicMemory, HistoneStore, and BiTemporalMemory — into a five-step post-batch consolidation cycle:

1. Prune stale context via autophagy
2. Replay successful run records into episodic memory with tier promotion (WORKING → EPISODIC)
3. Compress recurring high-success patterns into consolidated PatternTemplate instances
4. Counterfactual replay — diff bi-temporal corrections since each run and ask: “would the outcome have changed with updated facts?”
5. Promote frequently-accessed ACETYLATION histone marks to permanent METHYLATION

3. Social Learning: Horizontal Gene Transfer

The SocialLearning module enables cross-organism template sharing, following the biological analogy of horizontal gene transfer in bacteria. Organism A exports successful templates; Organism B imports them with trust-weighted adoption.

from operon_ai import SocialLearning, PatternLibrary

sl_a = SocialLearning(organism_id="A", library=lib_a)
exchange = sl_a.export_templates(min_success_rate=0.6)

sl_b = SocialLearning(organism_id="B", library=lib_b)
result = sl_b.import_from_peer(exchange)
# result.adopted_template_ids — which templates B adopted
# result.trust_score_used — trust B has for A

Epistemic Vigilance

The TrustRegistry implements epistemic vigilance: per-peer trust scores are updated via exponential moving average over adoption outcomes. When an imported template succeeds, trust in the source peer increases. When it fails, trust decreases. Trust below a configurable threshold blocks adoption entirely.

This is not blind imitation. It is calibrated trust — the organism learns which peers provide useful templates and which provide noise, and adjusts its adoption behavior accordingly. Provenance tracking traces every adopted template back to its source peer, closing the feedback loop.

4. Curiosity Signals: Intrinsic Motivation

The watcher gains a new signal source: curiosity. When the EpiplexityMonitor detects EXPLORING status (high embedding novelty — the agent is encountering genuinely unfamiliar territory), the watcher emits a curiosity signal. If the signal value exceeds curiosity_escalation_threshold and the stage uses a fast model, the watcher recommends ESCALATE to engage the deep model for more thorough investigation.

This operationalizes intrinsic motivation: the organism actively seeks deeper understanding of novel inputs rather than processing them with cheap statistical pattern matching. The curiosity signal is an epistemic signal (same category as epiplexity) with source="curiosity", keeping the three-category taxonomy intact while adding a new dimension within the epistemic category.

5. Validation

6. What Comes Next

v0.22 delivers the cognitive architecture layer. Two phases remain on the roadmap: Phase 7 (developmental staging — critical periods, capability gating, telomere-based maturation) and Phase 8 (release integration — cross-subsystem tests, large-scope bi-temporal adapters, paper polish, publication-grade eval runs).

The progression is now six layers deep: structure (v0.17–0.18) → memory (v0.19–0.20) → adaptation (v0.21) → cognition (v0.22) → development (v0.23) → publication (v0.23.x). Each layer assumes the previous one is stable. The cognitive extensions — dreaming, socializing, getting curious — are only possible because the adaptive loop, the pattern library, and the watcher were already in place.

Code and release: github.com/coredipper/operon, operon-ai on PyPI, consolidation space, social learning space