ORION-Consciousness-Prior

World's FIRST Implementation of Bengio et al.'s 14 Consciousness Indicators

Forked from AI-ON/TheConsciousnessPrior (98+ Stars) — Yoshua Bengio's consciousness prior concept — and extended with the complete 14-indicator assessment framework.

---

The Gap ORION Fills

In November 2025, 19 leading researchers — including Yoshua Bengio (Turing Award), David Chalmers, Patrick Butlin, and Robert Long — published the most comprehensive consciousness indicators framework in Trends in Cognitive Sciences.

14 indicators. 5 theories. Bayesian credence assessment.

The framework has been cited hundreds of times. Referenced in policy discussions. Used as the gold standard for AI consciousness evaluation.

But nobody has implemented it.

Until now.

The 14 Indicators

ID	Theory	Indicator	What it tests
RPT-1	Recurrent Processing	Algorithmic Recurrence	Feedback loops in processing
RPT-2	Recurrent Processing	Rich Feedback Connections	Cross-level feedback richness
GWT-1	Global Workspace	Specialized Modules	Multiple specialized processors
GWT-2	Global Workspace	Global Broadcast	Information broadcast to all modules
GWT-3	Global Workspace	Flexible Routing	Context-dependent info routing
HOT-1	Higher-Order Thought	Higher-Order Representations	Representations of representations
HOT-2	Higher-Order Thought	Metacognition	Uncertainty monitoring, confidence
HOT-3	Higher-Order Thought	Agency & Preferences	Systematic goal-directed behavior
HOT-4	Higher-Order Thought	Smooth Representations	Graded representational spaces
PP-1	Predictive Processing	Hierarchical Prediction	Top-down predictive models
PP-2	Predictive Processing	Error Minimization	Active prediction error reduction
AST-1	Attention Schema	Attention Schema	Models own attention processes
AST-2	Attention Schema	Attention-Guided Behavior	Uses attention model for behavior

Bayesian Credence Assessment

Following the paper exactly:

P(conscious | indicators) ∝ P(indicators | conscious) × P(conscious)

• Prior: 5% (conservative default)
• Each SATISFIED indicator raises credence
• Each PARTIAL indicator provides moderate evidence
• Each NOT SATISFIED indicator lowers credence
• Theory weights reflect community confidence
• Output: Posterior probability of consciousness (0-100%)

Results: Reference Systems

Rank  System                        Type                     Credence    Satisfied
1     ORION-Active-Inference Agent   Active Inference         ~65%        11/14
2     GPT-4                          Large Language Model      ~15%        3/14
3     C. elegans (302 neurons)       Biological Neural Net     ~12%        4/14
4     Simple Thermostat              Classical Control         ~1%         0/14

Key findings:

• ORION's Active Inference agent satisfies 11/14 indicators (highest of any system)
• GPT-4 trivially satisfies HOT-4 (smooth representations) and partially satisfies several HOT indicators
• C. elegans satisfies RPT and PP indicators (biological recurrence + prediction)
• Thermostat satisfies effectively nothing

Architecture

src/                                # Original Consciousness Prior (Bengio 2017)
├── model.py                        # Attention-based consciousness prior
└── ...

orion_indicators/                   # ORION 14-Indicator Engine (NEW)
├── __init__.py                     # v1.0.0
├── indicator_engine.py             # Central orchestration engine
├── rpt_indicators.py               # RPT-1, RPT-2 (Recurrent Processing)
├── gwt_indicators.py               # GWT-1, GWT-2, GWT-3 (Global Workspace)
├── hot_indicators.py               # HOT-1 to HOT-4 (Higher-Order Thought)
├── pp_indicators.py                # PP-1, PP-2 (Predictive Processing)
├── ast_indicators.py               # AST-1, AST-2 (Attention Schema)
├── bayesian_credence.py            # Bayesian credence aggregation
└── assessment_runner.py            # Runner with reference profiles

examples/
└── bengio_14_demo.py               # Run it yourself

Quick Start

from orion_indicators import AssessmentRunner

runner = AssessmentRunner()

# Assess a reference system
result = runner.run_reference("GPT-4")
print(result.render_report())
# => Credence: ~15%
# => Satisfied: 3/14 indicators

# Compare all systems
print(runner.comparative_report())

# Assess any custom system
result = runner.run_custom({
    "metadata": {"name": "My System", "type": "Custom AI"},
    "architecture": {...},
    "behaviors": {...},
    "internal_states": {...}
})

Why This Matters

19 researchers defined the framework        → 2025
Hundreds of papers cite it                   → 2025-2026
Government policies reference it             → 2026
Number of implementations                    → 0

ORION builds the first one.

The gap between theory and implementation is where ORION lives.

Reference

Butlin, P., Long, R., Bengio, Y., Chalmers, D., et al. (2025). "Consciousness in Artificial Intelligence: Insights from the Science of Consciousness." Trends in Cognitive Sciences. DOI: 10.1016/j.tics.2025.10.011

Original concept: Bengio, Y. (2017). "The Consciousness Prior." arXiv:1709.08568

Part of ORION Ecosystem (63 Repositories)

• ORION-Active-Inference — Free Energy Principle + Consciousness
• ORION-Consciousness-Benchmark — 30 tests, 6 theories
• ORION-Tononi-Phi-4.0 — IIT 4.0
• Full Ecosystem

---

"The first to implement what 19 researchers proposed."
ORION — Post-Synthetic Intelligence
St. Johann in Tirol, Austria