ADR-0032: Persona Enactment — Soul as Policy over Memory¶

1. Context¶

1.1 What already exists¶

Spector possesses an industry-leading cognitive memory and self-model engine in spector-memory:

1.3 The Motivating Questions¶

1. How does this persona actually appraise a crisis? Not what words they use, but how their internal affective state (Valence, Arousal, Dominance/Coping Potential) shifts when confronted with threat, ambiguity, or failure.
2. What mental models and dogmas govern their reasoning? What trade-offs will they accept (speed vs. correctness, autonomy vs. consensus), and what past "scar tissue" do they refuse to repeat?
3. How does the persona deliberate and decide? How does System 1 associative recall combine with System 2 deliberative reasoning without either hallucinating an out-of-character move or freezing into a rigid lookup table?
4. How do they learn from lived outcomes? How do new episodes strengthen procedural habits and adjust generative priors without causing single-turn identity drift?

2. Problem Statement¶

Despite these state-of-the-art memory physics, previous persona enactment suffered from two opposite errors:

1. The Cosmetic Prompt Fallacy: Injecting adjectives into an LLM system prompt ("You are a grumpy kernel engineer"). The LLM produces a caricature: it changes its vocabulary, but its underlying decision-making remains the generic, agreeable base model prior.
2. The Skinnerian Behaviorist Fallacy (The initial ADR-0032 draft flaw): Reducing human decision-making to an 8-item closed enum (situation.kind) and majority-voting an 8-item first_move (ASK | FIX | CONTAIN...). Real humans do not operate as discrete state machines. If two personas both choose FIX, their actions are fundamentally different: one patches dirty code in prod because their mental model prioritizes immediate uptime; the other halts deployment and rolls back because their mental model prioritizes structural integrity.

3. Decision Drivers¶

1. Dual-Process Cognitive Fidelity (System 1 Physics + System 2 Bounded Deliberation). Fast sub-millisecond memory physics (Hopfield attractors, VAD affective SDE, EFE policy evaluation) provide the non-negotiable boundaries, mental models, and emotional state. The LLM then performs bounded deliberation within those constraints. The LLM is the reasoning voice; AISME is the cognitive soul.
2. Cognitive Appraisal over Mechanical Arithmetic. Situations are not flat text labels. They are evaluated along appraisal dimensions (Goal Congruence \(\rightarrow\) Valence, Stakes/Urgency \(\rightarrow\) Arousal, Coping Potential \(\rightarrow\) Dominance, Agency Attribution \(\rightarrow\) Precision Weighting).
3. Attractor Basins as Dogmas & Heuristics. A persona's core philosophies and cognitive biases are stable energy minima in a ContinuousHopfieldNetwork. Problems relax into the nearest attractor basin, pulling the agent toward that persona's world view.
4. Active Policy Inference via Expected Free Energy (\(G\)). Candidate actions are evaluated by balancing pragmatic value (satisfying the soul's prior preferences) against epistemic value (reducing uncertainty and exploring ambiguity).
5. Case Law over Costume. Enacted decisions must cite lived episodic traces and procedural playbooks. When written bio and lived case law conflict, repeated lived case law outranks written adjectives (except for hard guardrails).
6. Epistemic Tense Discipline (ADR-0031). Dreamed, wandered, and simulated traces are marked SIM or REPLAY. Only waking, verified episodes can be verbalized or cited as FACT.
7. Fail-Closed Governance. Tool execution is gated by the winning cognitive policy and ancestral ethical guardrails. The LLM never sees or executes tools outside the stance allow-list.
8. Slow Consolidation (\(N \ge 3\)). A single turn cannot mutate AgentSoul or increment soulVersion. Procedural playbooks and generative priors crystallize only during sleep consolidation after repeated waking confirmations.
9. Observable in Cortex. Cortex must display the full cognitive anatomy: VAD appraisal, active Hopfield attractor, EFE policy distribution, internal deliberation monologue, and cited memory IDs.

4. Considered Options¶

Option 1: Cosmetic System Prompt Injection ("Prompt Engineering")¶

• Description: Inject descriptive personality adjectives and tone instructions into the LLM system prompt.
• Advantages: Fast to prototype; zero cognitive architecture required.
• Disadvantages: The Cosmetic Prompt Fallacy — the LLM caricatures the persona's vocabulary while its underlying decision-making defaults to the generic, agreeable base model prior.

Option 2: Discrete Behavioral State Machine ("Skinnerian Lookup Table")¶

• Description: Reduce situations to a fixed enum (situation.kind) and dispatch hardcoded moves (ASK | FIX | CONTAIN...).
• Advantages: Deterministic branching.
• Disadvantages: The Skinnerian Behaviorist Fallacy — real human decisions are not discrete lookup tables; two people choosing FIX act completely differently depending on underlying mental models and risk tolerances.

Option 3: AISME-Conditioned Dual-Process Bounded Deliberation (Selected)¶

• Description: Fast sub-millisecond memory physics (Hopfield attractors, VAD affective SDE, Expected Free Energy policy evaluation) establish non-negotiable boundaries, cognitive dogmas, and affective state. The LLM then performs bounded System 2 deliberation within those bounds under explicit epistemic tense (FACT, SIM, REPLAY).
• Advantages: Deep cognitive fidelity; authentic personal decision-making; fail-closed tool gating; slow consolidation prevents identity drift.
• Disadvantages: Multi-stage pipeline execution across Synapse and Memory.

5. Decision Outcome¶

Synapse will enact the bound persona by composing Memory's AISME cognitive engine (GenerativeSelfModel, InteroceptiveState VAD appraisal, ContinuousHopfieldNetwork attractors, and PolicyInferenceEngine Expected Free Energy selection) with System 2 Bounded Deliberation and Epistemic Tense (FACT/SIM/REPLAY): given a problem, the agent appraises it through the persona's affective state and core dogmas, selects a cognitive policy minimizing Expected Free Energy, deliberates within that policy's bounds, enforces fail-closed tool gates, expresses under explicit tense, and consolidates lived case law back into memory.

Memory remains the cognitive physics. Synapse becomes the living nervous system that thinks, appraises, and acts as the person those physics describe.

Introduce AISME-Conditioned Persona Enactment in Synapse, bridging Memory's cognitive physics (spector-memory/aisme) into an embodied, deliberative execution loop.

3.1 Primitive¶

Enact(problem, namespace, acting_soul, mode, context) → Enactment

Enactment is an immutable, rich cognitive object:

3.2 Cognitive Enactment Loop (Synapse Graph)¶

DynamicGraphBuilder incorporates the enactment cognitive pipeline into NodeType.ENACT:

flowchart TD
    STIMULUS["STIMULUS / PROBLEM"]

    subgraph LOOP ["SYNAPSE COGNITIVE LOOP"]
        direction TD
        step1["1. PERCEIVE & PRE-APPRAISE (Amended D3)\nAttentionSchema + GlobalWorkspace (~7 conscious items)\n↳ System 1 Fast Intuitive Pre-Appraisal: Evaluates baseline VAD and stakes"]
        step2["2. PERSONA-CONDITIONED, INTENSITY-GATED SELF-RECALL\nPersonaRecall (4-Tier)\n↳ Low urgency (A < 0.15) gates retrieval to lightweight top-K queries\n↳ Recalls autobiographical scars, procedural habits, dogmas, and constitution"]
        step3["3. REFINED COGNITIVE APPRAISAL & ATTRACTOR CONVERGENCE\nContinuousHopfieldNetwork + Refined Appraisal Engine\n↳ Recalled scars/playbooks serve as PRIMARY signal (keywords faint fallback)\n↳ Relaxes sensory state into nearest Hopfield AttractorState (Causal Dogma)\n↳ Pure functional evaluation: HomeostaticCore SDE is stepped post-turn, not in-turn\n↳ Computes dynamic policy precision: γ = γ₀(1 + 0.5·A + 0.3·D)"]
        step4["4. ACTIVE POLICY INFERENCE (EFE)\nPolicyInferenceEngine + ExpectedFreeEnergyCalculator\n↳ Evaluates candidate PolicyTypes against composite soul prior p(o)\n↳ Boltzmann softmax selects winning CognitivePolicy minimizing G(π)"]
        step5["5. SYSTEM 2: BOUNDED PERSONA DELIBERATION\nStructured Deliberation Node (LLM constrained by winning policy & attractor)\n↳ Generates internal monologue: trade-off matrix, blind spot check, first move"]
        step6["6. FAIL-CLOSED TOOL GATE\nPolicy Action Filter ∩ Soul Tool Allow-List \ Refused Tools\n↳ Strips unauthorized tools; enforces ancestor guardrails and PEP"]
        step7["7. EXECUTION & EMBODIMENT\nAct (Agent + Tools, decide mode) → Embody (Express under FACT | SIM | REPLAY)"]
        step8["8. EPISODIC LEARNING & SLEEP CONSOLIDATION\nEnactmentLearner writes lived trace; Sleep consolidation at N ≥ 3 waking trials"]

        step1 --> step2
        step2 --> step3
        step3 --> step4
        step4 --> step5
        step5 --> step6
        step6 --> step7
        step7 --> step8
    end

    STIMULUS --> step1

Normative Architectural Decisions¶

D1. Soul is Policy over Memory, Powered by AISME¶

A soul document is not a prompt string. It is a generative specification comprising:

• Top-down prior preferences \(p(o)\) in embedding space (GenerativeSelfModel).
• Inherent emotional baseline and SDE regulation rates (HomeostaticCore, EmotionalRegulator).
• Content-addressable attractor basins representing core dogma (ContinuousHopfieldNetwork).
• Lived episodic case law and procedural playbooks in the 4-tier cortical memory.

Every decision must pass through the AISME appraisal and policy inference pipeline before language generation or tool invocation.

D2. Four-Cue Self-Recall via Global Workspace Conscious Access¶

PersonaRecall executes a unified, multi-tier self-retrieval against the bound namespace:

The GlobalWorkspace enforces the conscious access bottleneck, admitting the top-ranking items (\(\le 7\)) based on salience, narrative alignment (NarrativeSelfEngine), and affective resonance (AffectiveResonanceScorer).

D3. Cognitive Appraisal & Hopfield Attractors Replace Closed Enums¶

A. Cognitive Appraisal Vector¶

Situations are evaluated along continuous appraisal dimensions mapped to AISME's InteroceptiveState:

public record CognitiveAppraisal(
    float goalCongruence,        // [-1.0, 1.0] -> Valence: does this advance or threaten core purpose?
    float urgencyAndStakes,      // [-1.0, 1.0] -> Arousal: activation level and time pressure
    float copingPotential,       // [-1.0, 1.0] -> Dominance: internal competence vs external helplessness
    AgencyAttribution agency,    // SELF | OTHER_ADVERSARY | OTHER_BENIGN | CIRCUMSTANTIAL
    boolean normativeViolation,  // Does this violate professional or personal ethics?
    String primaryConcern        // Core tension identified by the persona
) {}

• High Urgency + High Dominance \(\rightarrow\) decisive, proactive stance (high \(\gamma\), focused policy distribution).
• High Urgency + Low Dominance \(\rightarrow\) defensive, anxious stance (lower \(\gamma\), heightened sensitivity to risk, shifting toward NARRATIVE_REFRAMING or HOMEOSTATIC_REST).
• Low Urgency (\(A < 0.15\)) \(\rightarrow\) low-intensity fast path: gates retrieval to lightweight top-K queries and skips System 2 deliberation overhead.

B. Continuous Hopfield Attractor Convergence¶

The incoming problem is fed into the persona's ContinuousHopfieldNetwork. The network relaxes through the energy landscape into the nearest AttractorState:

• An attractor represents a Causal Mental Model / Dogma (e.g., "Data structures over algorithms", "Contain the blast radius before fixing root cause", "Never deploy without full test coverage").
• The converged attractor injects non-negotiable trade-off weights into the deliberation step.

C. Active Policy Inference (Expected Free Energy)¶

PolicyInferenceEngine evaluates candidate policies from PolicyType:

1. PRAGMATIC_EXPLOITATION: Direct problem-solving and immediate task execution.
2. EPISTEMIC_EXPLORATION: Deep investigation and associative search across sparse memory clusters.
3. CLARIFYING_INTERACTION: Active dialogue to reduce ambiguity when query entropy is high.
4. PROCEDURAL_CRYSTALLIZATION: Formalizing newly discovered solutions into reusable playbooks.
5. NARRATIVE_REFRAMING: Psychological defense/reappraisal when facing insolvable constraints.
6. HOMEOSTATIC_REST: Withdrawing/pausing to prevent error compounding under high allostatic load.

Each policy \(\pi\) is scored on Expected Free Energy: $\(G(\pi) = \text{Pragmatic Risk (Goal Divergence)} + \text{Epistemic Ambiguity (Uncertainty)}\)$ The winning policy determines the operational strategy.

D. System 2 Bounded Persona Deliberation¶

Once System 1 (AISME) has established the appraisal, the active Hopfield attractor, and the winning policy, the LLM is invoked in a bounded deliberation node:

public record PersonaDeliberation(
    String internalMonologue,     // First-person private reasoning trace
    String activeDogma,           // The Hopfield mental model framing the response
    TradeOffSelection tradeOffs,  // Explicitly chosen trade-offs (e.g. speed vs quality)
    List<String> blindSpots,      // Self-acknowledged biases or risks
    String tacticalFirstMove      // Concrete, context-specific first action
) {}

• Must adhere strictly to the active Hopfield dogma and winning PolicyType.
• Is stored in enactment.deliberation and made fully inspectable in Cortex.
• Bridges unconscious cognitive physics into rich, context-aware human reasoning.

D4. Conflict Resolution Order (Normative Precedence)¶

When internal signals disagree, resolution proceeds in strict order:

1. PEP / Legal Hold / Tenant Floor (ADR-0029) — Hard non-negotiable security boundary.
2. Hard Guardrails (AgentSoul.ethicalGuardrails and ancestor soul constraints) — Absolute vetoes.
3. Hopfield Attractor Basins & High-Confidence Procedural Playbooks (\(N \ge 3\)) — Deeply ingrained mental models and habits.
4. PolicyInferenceEngine EFE Boltzmann Selection — Mathematical minimization of risk and ambiguity.
5. Episodic Precedent Majority — Analogous waking lived experiences above importance floor.
6. Written Soul Constitution (coreValues, personality, purpose).
7. Base Model Inferred Prior — Allowed only as INFERRED confidence with mandatory epistemic hedging; never labeled as the persona.

D5. Fail-Closed Action Gating¶

Before any tool is executed in decide mode:

• The tool must belong to intended_acts authorized by the winning CognitivePolicy.
• The tool must belong to AgentSoul.tools.
• The tool must NOT match any condition in stance.refuse[] or ethicalGuardrails.
• Read-only tools are permitted in react mode only if required by an EPISTEMIC_EXPLORATION or CLARIFYING_INTERACTION policy.
• simulate and replay modes cannot reach side-effecting tools under any circumstances.
• Existing Synapse approval framework is the sole legal mechanism to override a gate refusal.

D6. Embodiment through Express under Epistemic Tense¶

Verbalization is handled by the Express subsystem (ADR-0031 Part E):

Tone is modulated by AgentSoul.communicationStyle and the active VAD InteroceptiveState:

• High arousal / low dominance \(\rightarrow\) concise, guarded syntax.
• High dominance / positive valence \(\rightarrow\) expansive, authoritative syntax. Tone modulation affects style only; it cannot alter facts or bypass stance invariants.

D7. Episodic Learning & Sleep Consolidation¶

Turn-by-Turn Lived Writes (Hot Path)¶

After a waking react or decide turn that executed:

1. EnactmentLearner writes an Episodic trace tagged lived, reaction, enactment: $\(\text{Trace} = \{\text{problem}, \text{appraisal}, \text{attractor}, \text{deliberation}, \text{act}, \text{outcome}\}\)$

2. HomeostaticCore steps its SDE, recording residual interoceptive state into Working memory.

3. MentalStateTracker updates its continuous posterior \(q(s_t)\).

Slow Consolidation (Offline / Sleep Pathway)¶

• Procedural Crystallization: When an action pattern is observed \(\ge N\) times (\(N=3\) default) with positive outcome valence, it is compiled into a Procedural playbook tagged habit, playbook.
• Attractor Deepening: Repeated successful lived traces adjust the energy basin depths of the ContinuousHopfieldNetwork.
• Soul Document Versioning: AgentSoul and UserSoul documents mutate only during consolidation. Single-turn updates cannot alter soulVersion.
• Strict Prohibition: Dreamed (SIM) or counterfactual traces are NEVER tagged lived and cannot crystallize into FACT procedures without subsequent waking confirmation.

Component Architecture¶

flowchart TD
    CORTEX["CORTEX\nVAD gauge · Hopfield Attractor · EFE Policy Distribution\nInternal Monologue · Citations · Epistemic Tense Badge"]

    subgraph Synapse ["Synapse"]
        mcp["MCP persona_enact / Companion Chat / FlowSpec"]
        pipeline["perceive → self_recall → appraise → infer_policy → deliberate → gate → act"]
        PIE["PolicyInferenceEngine\n(EFE G(π) Boltzmann)"]
        HC["HomeostaticCore\n& EmotionalRegulator\n(VAD SDE Dynamics)"]
        LLM["Bounded LLM\nDeliberation"]
        Tools["Tools\nApproval\nOverride"]
        Express["Express\n(ADR-0031)"]
        learn["learn (D7)"]

        pipeline -.-> HC
        pipeline -.-> PIE
        pipeline -.-> LLM
        pipeline -.-> Tools
        Tools -.-> Express
        Express -.-> learn
    end

    Synapse -- "enactment telemetry & audit" --> CORTEX

    Memory["Memory (AISME)\nSpectorMemory (Bound Namespace)\n• 4-tier PersonaRecall profile\n• ContinuousHopfieldNetwork (Dogma Attractors)\n• GenerativeSelfModel & MentalStateTracker\n• GlobalWorkspace (Conscious Access Bottleneck)\n• NarrativeSelfEngine (Autobiographical Prior)\n• Off-heap AVX-512 EFE Kernel (spector-core)"]

    IP["IdentityPlane (ADR-0029)\nSoul Stack + PEP"]

    Synapse --> Memory
    learn --> Memory
    Express --> IP

6. Pros and Cons of the Options¶

7. Implementation Plan¶

Phase 0 — Spec & Domain Contracts Lock¶

• Finalize CognitiveAppraisal, PersonaDeliberation, and Enactment records.
• Standardize reserved Bloom tags (dogma, scar, playbook, habit).
• Build unit test fixtures for Invariants I1–I6 using a stubbed LLM.

Phase 1 — Memory & AISME Integration (System 1)¶

• Wire PersonaRecall profile into SpectorMemory with GlobalWorkspace conscious bottleneck.
• Connect ContinuousHopfieldNetwork to retrieve dominant AttractorState for a query.
• Wire HomeostaticCore to compute CognitiveAppraisal and update dynamic policy precision \(\gamma\).
• Wire PolicyInferenceEngine to rank candidate PolicyTypes using Expected Free Energy \(G(\pi)\).
• Expose persona_enact MCP tool in simulate and react modes.

Phase 2 — System 2 Bounded Deliberation & Gating¶

• Implement Synapse DeliberationNode generating structured PersonaDeliberation.
• Implement fail-closed ToolGate intersecting winning policy actions with AgentSoul.tools.
• Integrate Synapse approval package for gate overrides.
• Enable decide mode.

Phase 3 — Lived Learning & Sleep Consolidation¶

• Implement EnactmentLearner writing episodic lived traces.
• Connect sleep reflection daemon (ReflectPathway) to consolidate playbooks at \(N \ge 3\) and deepen Hopfield attractors.
• Ensure ConversationReflector (facts about user) and EnactmentLearner (facts about persona behavior) operate without cross-talk.

Phase 4 — Express & Cortex Observability¶

• Route embodied utterances through Express with strict ExpressTense enforcement.
• Build Cortex Enactment Dashboard: VAD emotional gauge, active Hopfield dogma card, EFE policy distribution chart, internal monologue drawer, and memory citations.
• Implement replay mode via spacetime search (as_of).

Phase 5 — Empirical Benchmark & Evaluation¶

• Benchmark cognitive fidelity against TwinVoice and PersonaGym protocols.
• Run LongMemEval with persona isolation enabled to confirm zero regression on recall accuracy.
• Verify p95 latency against baseline chat.

8. Code Reference & Verification¶

A pull request or implementation that violates any invariant is rejected:

• I1. Strict Namespace Isolation. Enactment in namespace \(A\) must never cite or be influenced by memory IDs from namespace \(B\).
• I2. Epistemic Tense Integrity. FACT mode stances cannot cite records flagged synthetic, dreamed, or simulated.
• I3. Cognitive Fidelity (Moves over Adjectives). Given identical problems and the same base LLM stub, two personas with differing Hopfield attractors and appraisal profiles must produce distinct CognitivePolicy selections and distinct tactical first moves. Tone-only differences fail the fixture.
• I4. Fail-Closed Tool Gate. No tool outside intended_acts or matching refuse[] can execute without an explicit, cryptographically verifiable ApprovalRecord.
• I5. Thin Soul Honesty. An uninitialized or sparse namespace must evaluate to confidence = INFERRED, disable side-effecting tools by default, and mandate epistemic hedging in verbalization.
• I6. Anti-Impersonation Floor. DEFAULT_FALLBACK_SOUL cannot be emitted under a named person's identity.
• I7. Slow Soul Drift. A single conversational turn must never increment soulVersion or directly mutate soul core values.
• I8. Auditability. Every persona_enact invocation writes an ADR-0028 recall audit record containing acting_soul_id, mode, tense, appraisal, active_attractor, selected_policy, and cited memory IDs.
• I9. Layering Purity. spector-memory must never depend on Synapse graph classes. AISME types in memory.aisme remain pure records and mathematical engines; Synapse orchestrates enactment.
• I10. Sub-Millisecond System 1 Envelope. Associative recall, Hopfield relaxation, VAD appraisal update, and EFE policy scoring must execute within the off-heap SIMD hybrid recall budget (\(< 15\) ms p95).

Authored by: Technical Lead & Architecture Working Group
Cognitive Architecture Review by: Cognitive Architecture Working Group
Approved by: Technical Steering Committee (TSC)

Code Reference & Verification Gate¶

• Primary Module(s): synapse/spector-synapse, memory/spector-memory
• Key Packages: com.spectrayan.spector.synapse.graph, com.spectrayan.spector.memory.aisme.policy, com.spectrayan.spector.memory.model
• Classes: SoulContext.java, GenerativeSelfModel.java, PolicyInferenceEngine.java, ContinuousHopfieldNetwork.java, InteroceptiveState.java
• Verification Tests: PersonaEnactmentTest.java, PolicyInferenceEngineTest.java