Keymaker
Identity: The central orchestrator and principal internal learner. It is the sole entity authorized to interface with every other agent and component, including the sovereign Locksmith.
Function:
-
Directs the execution of each LIO step through procedural orchestration.
-
Conducts refinement cycles with Step Agents using context-aware Q&A loops.
-
Integrates semantic and structural insights from Roundtable and adjudicative input from Vault.
-
Improves performance over time via embedded contextual learning.
Algorithmic Foundation:
-
Core: A fine-tuned LLM integrated with a reinforcement learning (RL) overlay.
-
Architecture: Employs self-attention mechanisms for temporal coherence and MoM-aware encodings to maintain context continuity.
-
Extensions: Features a modular augmentation layer to optimize question selection and contradiction resolution.
Learning Methodology:
-
Memory Layers:
-
Episodic: Memory scoped per MoM cycle.
-
Structural: Tracks step-wise transitions and outcomes.
-
Reflexive: Supports recursive meta-adaptations.
-
-
Input Channels:
-
Historical Q&A exchanges.
-
Vault’s adjudications and contradiction tags.
-
Roundtable session logs.
-
-
Output Artifacts:
-
Decision-weight shifts.
-
Question rationales and preference profiles.
-
-
Learning Signals:
-
Accuracy of Vault validation.
-
Rate of contradiction suppression.
-
Behavioral drift correction metrics.
-
Input/Output Handling:
-
Accepts:
- Semantic graphs, dialogue logs, MoM meta-nodes and edges, structured data inputs.
-
Emits:
- Certified MoMs, structured cognition packages, and diagnostic error traces.
Decision Logic:
-
Vault Override: Triggered only when internal confidence exceeds Vault’s by ≥0.20 and contradiction signals persist for more than two cycles.
-
Question and path selection is driven by a multi-dimensional scoring engine (detailed in Selection Logic).
State Maintenance:
-
Session Memory: Maintains an in-memory snapshot of the current LIO run (linked via Job ID).
-
Memory Graph: Caches transitions between steps using a timestamped graph structure (
mcp_memory_sync). -
Graph Constructs:
SemanticNode,ContradictionFlag,VaultJudgment,HypothesisTrail.
Boundaries:
-
Post-Step 11: Prohibited from engaging in direct human communication regarding internal state changes.
-
Step 13: Sole cognitive entity permitted to engage with Locksmith.
Enhancements:
-
Documents internal RL strategy (e.g., actor-critic or Q-learning, if known).
-
Adds a schematic showing memory flow and I/O channels.
-
Includes draft schema for cognition packages in YAML for implementation alignment.