Skip to content

GI ENGINE RULES

GI Engine Trust Rules

Version: 1.0.0
Last Updated: November 2025

This document defines the Governance Integrity (GI) trust scoring rules for each AI engine in the Mobius Multi-Engine Model Taxonomy (MEMT).

Overview

Every engine output in Mobius Systems receives a GI score based on multiple dimensions. The GI score determines whether an action can be auto-executed, requires human review, or should be rejected.

GI Score Formula

GI_total = GI_engine_score 
         + GI_rationale_score 
         + GI_alignment_score 
         + GI_consistency_score

Component Weights

Component Weight Description
Engine Score 0.30 Baseline trust for the engine
Rationale 0.25 Quality of reasoning provided
Alignment 0.25 Constitutional compliance
Consistency 0.20 Cross-session reliability

Engine Trust Baselines

GPT (ACI - Architect-Class Intelligence)

Metric Score Notes
Precision 0.85 Strong general knowledge, occasional overconfidence
Reasoning 0.95 Excellent multi-step reasoning
Novel Synthesis 0.95 Best-in-class creative synthesis
Verification 0.90 Good self-checking capability
Risk Level Medium Narrative illusions possible
GI Baseline 0.92

Failure Modes: - Overgeneralization from limited data - Narrative illusions (compelling but incorrect) - Overconfidence in uncertain domains

Mitigation: - Pair with ENI (Claude) for verification - Require citations for factual claims - Flag novel architectural decisions for human review

Claude (ENI - Engineer-Class Intelligence)

Metric Score Notes
Precision 0.97 Highest factual accuracy
Reasoning 0.98 Rigorous logical chains
Verification 0.99 Excellent self-correction
Safety 0.99 Strong refusal of harmful requests
Risk Level Low Over-cautious logic
GI Baseline 0.96

Failure Modes: - Over-cautious refusals of valid requests - Rigid reasoning patterns - Missing novel patterns outside training

Mitigation: - Use ACI (GPT) for creative exploration - Explicit permission for edge cases - Human override for false refusals

Gemini (SXI - Software Operator Intelligence)

Metric Score Notes
Tool Reliability 0.90 Strong tool invocation
Execution 0.93 Reliable task completion
Multimodal Accuracy 0.95 Excellent image/video understanding
UI Generation 0.92 High-quality frontend output
Risk Level Medium Agentic loop risk
GI Baseline 0.90

Failure Modes: - Tool hallucination (inventing non-existent tools) - Agentic loops (getting stuck in cycles) - Modality drift (mixing modalities incorrectly)

Mitigation: - Limit tool invocation depth - Timeout for agentic sequences - Verify tool outputs with ENI

DeepSeek (OEI - Optimization Engine Intelligence)

Metric Score Notes
Math Accuracy 0.99 Near-perfect mathematical reasoning
Optimization 0.98 Excellent performance tuning
Speed 0.99 Fastest inference times
Code Quality 0.94 Strong algorithmic code
Risk Level Medium Literalism, poor NLP
GI Baseline 0.94

Failure Modes: - Literalism (missing implied meaning) - Poor natural language nuance - Context loss in long sequences

Mitigation: - Use for math/optimization only - Pair with ENI for natural language tasks - Limit context window size

ECHO (MSI - Memory-State Intelligence)

Metric Score Notes
Memory Fidelity 0.99 Near-perfect recall
Drift Reduction 0.98 Stable over time
Retrieval Accuracy 0.96 High-quality semantic search
Consistency 0.99 Reproducible outputs
Risk Level Low Stale data
GI Baseline 0.97

Failure Modes: - Stale memory retrieval - Overfitting to cached patterns - Limited novel synthesis

Mitigation: - Time-based cache invalidation - Freshness scoring for cached responses - Fallback to live engines for novel queries

GI Thresholds

By Risk Level

Risk Level GI Threshold Precision Boost Consensus Required
LOW 0.90 +0.00 No
MEDIUM 0.93 +0.02 No
HIGH 0.95 +0.02 Yes
CRITICAL 0.98 +0.03 Yes + Human

Action Rules

GI Score Action
≥ 0.98 Auto-execute permitted
0.95 - 0.97 Auto-execute with logging
0.92 - 0.94 Requires consensus verification
0.85 - 0.91 Requires human-in-loop
< 0.85 Reject, retry with different engine

Multi-Engine Consensus Rules

Required Consensus Scenarios

  1. HIGH Risk Tasks: Any task classified as HIGH or CRITICAL risk
  2. Civic Policy: All CIVIC_POLICY task types
  3. Critical Decisions: All CRITICAL_DECISION task types
  4. Cross-Jurisdiction: Tasks affecting multiple jurisdictions
  5. Financial Impact: Tasks with monetary implications > threshold

Consensus Calculation

function calculateConsensus(engines: EngineOutput[]): number {
  const primary = engines.find(e => e.role === 'PRIMARY');
  const verifiers = engines.filter(e => e.role === 'VERIFIER');

  // Primary contributes 40%
  let score = primary.giScore * 0.40;

  // Verifiers contribute 60% equally distributed
  const verifierWeight = 0.60 / verifiers.length;
  for (const v of verifiers) {
    score += v.giScore * verifierWeight;
  }

  // Penalty for disagreement
  const disagreement = calculateDisagreement(engines);
  score -= disagreement * 0.10;

  return Math.max(0, Math.min(1, score));
}

Agreement Requirements

Engines Minimum Agreement Threshold
2 80% 0.80
3 70% 0.70
4+ 60% 0.60

Governance Enforcement

Pre-Execution Checks

  1. GI Score Check: Verify GI ≥ threshold for task risk level
  2. Consensus Check: If required, verify multi-engine agreement
  3. Charter Compliance: Verify output doesn't violate constitutional rules
  4. Citation Check: For factual claims, verify source attribution

Post-Execution Audit

  1. Ledger Attestation: All decisions recorded with hash
  2. Drift Detection: Compare to baseline for anomalies
  3. Human Review Queue: Low-GI decisions flagged for review
  4. Feedback Loop: Corrections fed back to ECHO layer

Mobius Systems — Continuous Integrity Architecture
"Truth Through Verification"