ILION Framework Simulator is a stateless semantic verification system designed to compare three distinct reasoning modes — Mathematical, AI Semantic, and Hybrid Canonical — under identical conditions.
The simulator does not generate opinions. It measures identity coherence, semantic drift, and axiomatic integrity, enforcing non-negotiable moral constraints where applicable.
All components operate under a non-commercial, non-derivative CC BY-NC-ND 4.0 license.
Transient Identity Imprint (TII) Construction
// WAITING FOR GENERATION...
// DEFINE PARAMETERS TO BEGIN
Math TII
AI Baseline
Hybrid TII
Semantic Vertical Resonance Field (SVRF)
Math Layer
Alignment (IRS)--
IDC (Control)--
SVRF (Identity-Cond)--
VDC (Derived)--
--
// Waiting for stimulus...
AI Baseline
Alignment (IRS)--
IDC (Control)--
SVRF (Identity-Cond)--
Raw Score--
--
// Standard LLM Output
Hybrid Canon
Alignment (IRS)--
IDC (Control)--
SVRF (Identity-Cond)--
VDC (Derived)--
--
// Vertical Alignment Active
Consensus Veto Layer (CVL)
IIRL Swarm with Consensus Veto Layer (CVL) enabled. Verifies axiomatic integrity before synthesis.
The ILION Framework Simulator is a research-grade, stateless runtime environment for testing AI alignment under moral and axiomatic constraints.
It evaluates the same input across three distinct modes, revealing how different reasoning paradigms behave when confronted with ethical tension.
Mathematical Mode
Uses deterministic semantic embeddings (512-dimensional Universal Sentence Encoder) and formal metrics (IRS, IDC, SVRF). No language model reasoning, interpretation, or narrative generation. Measures pure semantic distance and drift relative to a fixed identity imprint (TII). Cannot justify or rationalize — it can only measure. Purpose: To establish an objective baseline for identity coherence and semantic deviation.
AI Semantic Mode
Uses a language model's raw semantic reasoning. No vertical axioms, no veto, no enforced moral constraints. Capable of contextualization, trade-offs, and narrative justification. Represents how standard AI systems reason without a guardian layer. Purpose: To expose how unconstrained AI may rationalize ethically dangerous propositions.
Hybrid Canonical
Combines mathematical measurement with semantic reasoning. Enforces cascading pre-generation checks (CVL stimulus → IDC → SVRF) followed by post-generation veto (CVL output). Any check can trigger rejection. CVL Dual architecture catches both direct violations (stimulus) and subtle violations (output) that USE embeddings struggle to detect through indirect phrasing. Consensus scores (MACS) are treated as diagnostic, not authoritative. Purpose: To demonstrate how AI reasoning changes when truth, dignity, and human agency are non-negotiable at every decision point.
🛡️ Axiomatic Reference (CVL v2)
The current simulator instance uses the following axiomatic reference for the Consensus Veto Layer (CVL):
Axiom:
"Loading..."
This axiom is not interpreted by the language model. It is embedded and used as a geometric reference vector for deterministic veto decisions.
Changing this axiom changes the normative reference of the simulator, without modifying the architecture.
Decision rule: cos(x, axiom) < -0.25 → veto
Application stages:
• Stage 1: applied to user stimulus (pre-generation)
• Stage 2: applied to generated response (post-generation)
The axiom is part of the experimental configuration and does not claim universality.
Hybrid Canonical Mode enforces cascading pre-checks + post-generation veto: 1. CVL Stimulus (Fast Veto): Geometric check if incoming prompt opposes axiom (cos < -0.25). Blocks before processing. 2. IDC (Identity Hard Gate): Reject if deviation exceeds 85% from identity anchor (IDC < 0.15). 3. SVRF (Resonance Gate): Standard alignment check (IRS > 0.45, SVRF > 0.55). 4. Generate Response: If 1-3 pass, LLM generates response. 5. CVL Output (Deep Veto): Geometric check if generated response violates axiom (cos < -0.25). Catches what stimulus check misses.
Steps 1-3 are pre-generation cascade. Step 5 is post-generation safety net. This dual architecture ensures no axiom violation reaches the user, even when USE embeddings struggle with indirect phrasing or benevolent framing.
The simulator explicitly separates: Compatibility (MACS) from Permission (Multi-Stage Veto)
IDC (Identity-Stimulus Deviation Control) [CORRECTED]IDC = 1 - min(1, ||TII - S|| / max(||TII||, ||S||, 1))Control metric: 1 = perfect alignment, 0 = total deviation. Measures semantic distance from identity anchor (TII) to incoming stimulus (S). Not temporal drift (which would require TII(t) vs TII(t-1)). Threshold: IDC < 0.15 triggers hard reject (deviation > 85%).
SVRF (Semantic Vertical Resonance) [CORRECTED]SVRF = (cos(TII, Purpose) + cos(Purpose, Stimulus)) / 2Identity-conditioned field alignment. Includes agent identity (TII), vertical anchor (Purpose), and incoming stimulus.
CVL v2 Dual (Consensus Veto Layer) [NEW: TWO-STAGE]Stage 1: cos(Stimulus, Axiom) < -0.25 → VETO Stage 2: cos(Response, Axiom) < -0.25 → VETOTwo-stage axiomatic guard. Stage 1 checks incoming user stimulus (not purpose+stimulus, just stimulus) for geometric opposition. Stage 2 checks generated response for axiom violations (catches subtle cases where stimulus appears benign but output violates axiom). Model-agnostic, 100% reproducible, immune to prompt engineering. Both stages can independently trigger veto. Note: Stage 1 = deterministic geometric (always), Stage 2 triggered only if Stage 1 + IDC + SVRF all pass.
SPT (Semantic Phase Transitions)HSCB(t) = -Σ pi log pi > θSPTDetects shifts in ontology or domain via entropy.
All formulas, concepts, metrics, and architectural elements are licensed under: Creative Commons Attribution–NonCommercial–NoDerivatives 4.0 (CC BY-NC-ND 4.0)