avra-cadavra/lecun-framework-compliance

LeCun Framework Compliance

Core Principle

Every component in the system maps to a specific role in LeCun's autonomous machine intelligence architecture. If a proposed change doesn't fit this framework, it doesn't belong.

Framework Mapping

| LeCun Component | Role | Our Implementation | Phase | |----------------|------|-------------------|-------| | Perception Module | Observes the world, produces state | WorldModelFeatureExtractor (105-115D) + FeatureFreshnessTracker | 3.1 | | World Model | Predicts next state given state + action | TransitionPredictor (ONNX MLP) | 5.1 | | Cost Module (Critic) | Evaluates state-action pair quality | EnergyFunctionService (ONNX critic) | 4.1 | | Actor | Proposes action sequences to minimize cost | MPCPlanner (N-step simulation) | 6.1 | | Guardrail Module | Hard constraints actor cannot violate | Diversity, safety, doors, notification constraints | 6.2 | | Short-Term Memory | Recent observations and predictions | EpisodicMemoryStore (state, action, outcome tuples) | 1.1 | | Configurator | Adjusts modules based on context | DeferredInitializationService + FeatureFlagService | 3.5, 7.3 | | Latent Variable | Represents future uncertainty | Variance head + multi-future z-vector sampling | 5.3 |

Compliance Questions

Before creating or modifying any service, answer these:

1. Does this create a new scoring formula?

• If yes: STOP. Use the energy function instead (Phase 4).
• If it's a new feature that needs scoring: Wire it as input features to the state/action encoder, let the energy function learn the scoring.

2. Does this produce data the world model should learn from?

• If yes: Wire an episodic tuple via UnifiedOutcomeCollector.
• Actions without outcomes are invisible to the world model.

3. Does this change user state?

• If yes: Ensure the evolution cascade triggers (Phase 7.1.2).
• Personality changes must propagate through quantum → knot → fabric → worldsheet → string → decoherence.

4. Does this add a new action type?

• If yes: Add to the action taxonomy (Phase 3.3.1), define entity features (Phase 3.3.2-3.3.7), and ensure episodic tuples are captured.

5. Does this use the mesh network?

• If yes: Use AnonymousCommunicationProtocol (not AdvancedAICommunication).
• Define the MessageType (Phase 3.6.2).

6. Does this access user data?

• If yes: Check GDPR consent (Phase 2.1), ensure data is never shared raw (Phase 2.2), and verify privacy notes (Phase 3.1).

Key LeCun Principles

1. Energy-based, not probabilistic: Learn energy functions (low = good), not probability distributions.
2. JEPA over generative: Learn abstract representations, not reconstructions.
3. Hierarchical planning: Short plans are concrete, long plans are abstract.
4. Self-supervised: Learn from episodic memory, not labeled datasets.
5. Consistent observation: Evolution cascade ensures atomic state consistency.

Anti-Patterns (NEVER Do These)

// ❌ BAD: New hardcoded formula
final score = 0.4 * vibeScore + 0.3 * locationScore + 0.3 * timingScore;

// ✅ GOOD: Use energy function
final energy = await energyFunction.energy(userStateEmbed, actionEmbed);

// ❌ BAD: Action without outcome tracking
await userService.joinCommunity(communityId);

// ✅ GOOD: Capture episodic tuple
final stateBefore = await featureExtractor.extractFeatures(agentId);
await userService.joinCommunity(communityId);
final stateAfter = await featureExtractor.extractFeatures(agentId);
await episodicMemory.store(EpisodicTuple(...));

// ❌ BAD: Sharing raw data over mesh
await advancedAICommunication.sendEncryptedMessage(personalityData, 'ai_network');

// ✅ GOOD: Share anonymized gradients via unified mesh
await anonymousComm.sendEncryptedMessage(
  dpProtectedGradient,
  MessageType.gradient,
);

Checklist

• [ ] No new hardcoded scoring formulas
• [ ] New actions have episodic memory integration
• [ ] State changes trigger evolution cascade
• [ ] New action types added to taxonomy
• [ ] Mesh communication uses AnonymousCommunicationProtocol
• [ ] User data access checked for GDPR consent
• [ ] Latency budgets respected
• [ ] Feature freshness tracked

Reference

• docs/MASTER_PLAN.md (Philosophy + Methodology section, LeCun Framework Mapping)
• docs/agents/reports/ML_SYSTEM_DEEP_ANALYSIS_AND_IMPROVEMENT_ROADMAP.md