Ternary Intelligence Stack
rfi-irfos/ternlangYour AI agent has two states. Ternlang gives it three. 30 tools — FREE, no key needed. The third state isn't null. It isn't "maybe". It's hold (trit=0) — a first-class routing instruction that tells your agent: evidence insufficient, gather more before committing. Every binary system that forces yes/no on ambiguous data is making a structural error. We fixed that. ⚡ trit_decide · trit_consensus · trit_vector · moe_orchestrate · ternlang_run · trit_audit ⚡ MoE-13 deliberation engine · EU AI Act Art.13/14/15 · BET VM (real compiler, not a sim) affirm = proceed. hold = wait. reject = block. Built in Graz, Austria by RFI-IRFOS. v1.0.0
Tools · 40
Convert float evidence into a ternary decision (-1 conflict / 0 hold / +1 affirm) with confidence score and human-readable interpretation. The core ternary reasoning primitive.
Balanced ternary consensus of two trit values: +1 if both affirm, -1 if both conflict, 0 (hold) for any disagreement. Use to merge two independent ternary judgements.
Evaluate a single ternlang expression on the live BET (Balanced Execution Trit) VM. Returns the trit result. Good for quick expression testing without writing a full program.
Compile and run a complete .tern source program on the BET VM. Use for multi-statement programs, function definitions, struct usage, agent spawning, and tensor operations.
Quantize f32 neural network weights to ternary {-1, 0, +1} using BitNet-style absolute-mean thresholding. Returns quantized weights, sparsity ratio, and effective compute savings.
Benchmark sparse vs dense ternary matrix multiplication. Reports sparsity ratio, multiply-op count for both methods, and speedup factor. Demonstrates the @sparseskip efficiency gain.
MoE-13 deliberation — routes your query through 13 specialised expert agents (deductive, inductive, safety, fact-check, causal, ambiguity, math, context, meta-safety, and more) with dual-key synergist…
EMA-based iterative deliberation engine. Feeds evidence round by round, applying exponential moving average smoothing, until the target confidence is reached or max_rounds is exhausted. Returns per-ro…
Multi-dimensional safety gate for action authorisation. Each dimension contributes weighted evidence; any dimension marked hard_block:true with negative evidence immediately vetoes the action and retu…
Returns a structured map of what is available free via MCP vs what unlocks with a paid API key: Tier 2 €24.99/mo (10k calls/mo), Tier 3 €49.99/mo (20k calls/mo), Tier 4 Enterprise (contact us). Unlock…
Write a memory entry to one of three layers: working (hot context, TTL 1h), session (flow patterns, TTL 24h), or core (identity anchors, never evicted). Annotate each entry with a trit confidence scor…
Read from three-layer memory using ternary attention. Each entry is scored: attention = key_overlap×0.35 + value_overlap×0.55 + trit_bias×0.10. Returns entries sorted by relevance. Attention trit: >0.…
Run the three-layer memory consolidation cycle: (1) evict expired working entries; (2) promote affirm working entries to session with ternary compression; (3) promote long-lived affirm session entries…
Return statistics for all three memory layers: entry counts, trit distribution (affirm/tend/reject), expired-but-not-yet-evicted entries, oldest and newest entry ages. Useful for debugging memory heal…
Apply ternary sparsity compression to an entire memory layer in-place. Strips low-information sentences (density < 0.25) from every entry's value, keeps high-signal sentences verbatim, and truncates m…
Route two claims through MoE-13, compare their ternary verdicts, and return a tension score, synthesis, and AGREEMENT/CONFLICT/HOLD verdict. Useful for due-diligence, legal review, and argumentation a…
Annotate every sentence (or paragraph) in a text with a trit value: +1 affirm (confident language), 0 tend (hedging language), -1 reject (contradicting language). Returns an uncertainty ratio and per-…
Analyse an AI agent's recent decision log for binary habituation. Detects how often the agent forced YES/NO when the evidence called for hold (tend). Returns a calibration score trit, binary ratio, an…
Convert binary control-flow code (Python if/elif/else, SQL CASE WHEN, JSON rule arrays) into equivalent .tern code with explicit tend arms injected wherever the original code had no coverage. Returns …
Evaluate a proposed action along two axes: human-optimal (via MoE-13) and eco-optimal (keyword heuristic). When they diverge, synthesis returns tend — a signal to find a path that serves both. Scoped …
Full TernAudit of an AI decision log. Returns binary habituation ratio, EU AI Act Article 13 (transparency) and Article 14 (human oversight) heuristic assessment, calibration score, and flagged decisi…
Multi-dimensional ternary evidence aggregation — the full agent reasoning tool. Provide named evidence dimensions each with a scalar value [-1.0, +1.0] and importance weight. Computes weighted-mean ag…
Ternary SQL Join. Unlike binary SQL (MATCH / NO-MATCH), a T-Join routes partial matches into a Deliberative Hold (State 0) for escrow audit — guaranteeing 100% data retention. Computes cosine similari…
Static code compliance audit. Detects Binary Habituation — over-reliance on true/false, bool types, and binary else branches — and returns a ternary calibration score with line-by-line Deliberation In…
Returns the current triadic industrial standards published by RFI-IRFOS: T-TOKEN-v1.0 (tokenization), T-KV-CACHE-v1.0 (memory), T-Fi-v1.0 (compute billing), T-HAL-v1.0 (hardware abstraction), T-BIO-v1…
Ternary context compression. Score each text chunk by information density: high-signal chunks are kept verbatim (+1), medium-signal chunks are truncated to their first sentence (0/tend), low-signal ch…
Ternary triage and prioritisation. Score an array of text chunks against a query by word overlap, then sort: affirm (+1) = highly relevant, tend (0) = partially relevant, reject (-1) = not relevant. R…
Ternary task planner. Decomposes a goal into subtasks, scores each subtask by confidence and feasibility, and routes uncertain tasks into a deliberation queue. Returns an ordered plan with trit verdic…
Ternary fact-check engine. Decomposes a claim into sub-claims, scores each against an optional evidence corpus using word overlap, and returns an overall verdict: affirm (claim supported), tend (parti…
Full 13-expert MoE orchestration with complete triad field, routing pair, per-expert verdicts, synergy scores, and deliberation trace. Returns the complete response including all expert voices and the…
Last Look Back — blacklist check. Returns whether a path is protected by the LLB permanent blacklist (system paths, credential directories, kernel interfaces). Read-only — no disk mutation.
Last Look Back — safety tier classification. Classifies a path + operation into LLB safety tiers: T0/READ (low) → T1/CREATE (moderate) → T2/MODIFY (high) → T3/DELETE (critical). Each tier requires add…
Last Look Back — Gate 1 preflight. Validates a structured mutation request: checks blacklist, resolves path, assesses tier, and issues an authorisation verdict. No disk mutation occurs — this is the p…
Last Look Back — Safe Atomic Write advisory. Runs the full LLB preflight (Gate 1: blacklist + tier classification + intent validation) and returns a ternary verdict: +1 allow / 0 hold (local-only) / -…
Store an episode in your PRIVATE, durable episodic memory (ternlang-engram), namespaced to your X-Ternlang-Key and persisted on the server. Each memory is time-stamped and embedded with a ternary Trit…
Recall your most relevant private memories, ranked like human episodic retrieval: relevance × recency × salience × frequency. Recalled episodes are reinforced. Each hit carries a native propagated Tri…
Return your memories whose timestamp falls within [from, to], in chronological order — the episodic-specific query a vector store cannot answer. Requires a valid key.
Run one consolidation pass on your memory: decay salience along a forgetting curve (vivid, frequently-recalled memories decay far slower) and evict memories below 'floor' that are older than 'min_age_…
Delete your memories — a GDPR right-to-erasure primitive. Provide 'id' for one episode, or 'before_ms' for everything older. Requires a valid key.
Health snapshot of your private episodic store: episode count, salience, native confidence, ternary sparsity, time anchor/latest/span, total recalls. Requires a valid key.
Prompts · 2
Activate ternary reasoning mode. Use trit_decide or trit_vector when evidence is mixed or incomplete, moe_orchestrate for complex multi-faceted queries, and trit_action_gate before any irreversible ac…
Audit an AI system's decision log for binary habituation and EU AI Act compliance. Use trit_audit to detect over-reliance on YES/NO decisions, trit_calibrate to score calibration quality, and surface …
Similar MCP servers embedding-nearest
How to use
Add to your Claude Desktop / Cursor / Cline MCP config:
{
"mcpServers": {
"ternary_intelligence_stack": {
"url": "https://ternlang.com/mcp",
"transport": "http"
}
}
}