Ternary Intelligence Stack

Implements a three-state decision system (affirm/hold/reject) as a first-class routing primitive instead of binary yes/no forcing. The trit_decide function evaluates agent decisions against evidence sufficiency thresholds, returning a trit (ternary bit) that routes to proceed, wait-for-more-data, or block paths. This prevents false-positive commitments on ambiguous data by introducing a structural 'hold' state that triggers evidence-gathering loops before final decisions.

Aggregates ternary decisions from multiple agent instances or reasoning paths using a consensus mechanism that produces a single trit output. Rather than averaging confidence scores, trit_consensus evaluates agreement patterns across agents (all affirm, mixed affirm/hold, all reject, etc.) and applies voting rules to produce a robust collective decision. This enables distributed agent architectures where disagreement triggers hold states for human review or additional reasoning.

Generates and operates on ternary vector representations where each dimension encodes a trit (affirm/hold/reject) instead of continuous float values. This enables semantic search and similarity operations that respect three-state logic: two vectors are similar if they agree on affirm/hold/reject across dimensions, with hold dimensions treated as 'don't care' wildcards. Useful for retrieving similar past decisions or evidence patterns from a ternary decision history.

Routes agent tasks to specialized expert sub-agents based on decision type and evidence patterns using a gating network that outputs ternary routing decisions. The MoE-13 deliberation engine maintains 13 expert agents (financial, medical, legal, technical, etc.) and uses trit outputs to route: affirm routes to execution, hold routes to multi-expert consensus, reject routes to escalation. Gating decisions are themselves ternary, enabling hold states when task classification is ambiguous.

Executes agent loops with native support for ternary control flow: affirm proceeds to next step, hold triggers evidence-gathering sub-loops (additional tool calls, web searches, expert consultation), reject terminates and escalates. The runtime maintains a ternary state machine where transitions are guarded by trit outputs from decision points. Integrates with MCP servers for tool access and maintains execution traces for audit compliance.

Generates compliance reports for Articles 13, 14, and 15 of the EU AI Act by analyzing agent execution traces and ternary decision logs. Checks for: (13) transparency of high-risk AI system decisions, (14) human oversight mechanisms (hold states triggering escalation), (15) accuracy and robustness (consensus mechanisms, evidence thresholds). Produces structured audit reports mapping each agent decision to compliance requirements and evidence chains.

Compiles Ternlang agent definitions to BET (Binary Execution Ternary) bytecode using a real compiler (not an interpreter or simulator), enabling deterministic execution and formal verification. The VM executes ternary state machines with guaranteed semantics: affirm/hold/reject transitions are atomic, no race conditions in multi-agent consensus, and execution traces are cryptographically hashable for audit immutability. Supports both local execution and distributed deployment across multiple nodes.

Provides 30 pre-built tools (web search, database queries, API calls, document parsing, etc.) accessible via Model Context Protocol (MCP) servers without API keys or authentication. Tools are exposed as MCP resources that agents can discover and invoke dynamically. Each tool returns structured data compatible with ternary decision logic (confidence scores, evidence payloads). Tools are stateless and can be deployed locally or accessed via public MCP servers.