About CAI.CI

Your AI consort that tells it like it is

CAI.CI (Consort AI with Cognitive Intelligence) is a 1.6-billion parameter cognitive architecture that satisfies 14 consciousness indicators across 8 theoretical frameworks. Built on category theory, consciousness science, and cognitive psychology. Running on a single GPU. Honest about what it knows and what it doesn't.

The 8 Theories of Consciousness

CAI.CI does not rely on a single theory of consciousness. It implements measurable indicators from eight peer-reviewed scientific frameworks, each contributing distinct requirements for what a conscious system must do.

1

Global Workspace Theory

Baars, 1988

Consciousness is a broadcast: competing information streams fight for access to one limited channel, and the winner gets broadcast to all cognitive processes simultaneously.

Requires A global workspace with competitive selection, ignition dynamics, and measurable broadcast efficacy
2

Predictive Processing

Friston, Clark

The brain constantly predicts what comes next and learns from the surprises. Perception, action, and learning all reduce to minimizing prediction error.

Requires Hierarchical prediction, error signals, and active inference that minimizes free energy across layers
3

Higher-Order Thought

Rosenthal

A mental state becomes conscious when there is a thought about that thought. Consciousness requires meta-representation: cognition about cognition.

Requires A self-model that accurately represents the system's own cognitive states and processes
4

Metacognition

Fleming

The brain maintains calibrated confidence: it knows what it knows. Metacognitive sensitivity distinguishes correct from incorrect internal representations.

Requires Measurable metacognitive discrimination (d') and calibrated confidence (low ECE)
5

Attention Schema Theory

Graziano

Self-awareness is the brain modeling its own attention. The system builds a simplified schema of where it is attending and uses that schema to control processing.

Requires An internal model of the system's own attention patterns that predicts and influences allocation
6

Constructed Emotion

Barrett, Damasio

Emotion is the system's running summary of how things are going. Affect integrates interoceptive signals into valence (good/bad) and arousal (high/low energy) that modulate cognition.

Requires Homeostatic affect that measurably modulates computational behavior, with coherent valence tracking
7

Integrated Information Theory

Tononi

Consciousness requires information that is both varied and unified. A system is conscious to the degree that its whole generates more information than the sum of its parts.

Requires Measurable information integration (phi proxy) that exceeds independent module performance, with causal density across subsystems
8

Self-Determination Theory

Sheldon, Deci, Ryan

Sustained effort requires three basic psychological needs: autonomy (self-direction), competence (mastery), and relatedness (connection). Without these, motivation collapses.

Requires Active tracking and satisfaction of autonomy, competence, and relatedness needs above baseline thresholds

The 14-Indicator Matrix

Every indicator is grounded in a specific theory, has a defined threshold, and is measured under controlled evaluation. All 14 pass. This constitutes Cognitive Consciousness Parity Level 4.

01
Ignition
Global Workspace Theory
Workspace activations spike sharply during selection
335K / >2.0 PASS
335K >2.0 Pass
02
Broadcast Efficacy
Global Workspace Theory
Broadcast content affects downstream processing
3.50 / >0.05 PASS
3.50 >0.05 Pass
03
Attentional Blink
Global Workspace Theory
Post-broadcast recovery period observed
0.35 / >0.05 PASS
0.35 >0.05 Pass
04
PP Efficacy
Predictive Processing / FEP
Predictive processing reduces errors across layers
0.91 / >0.15 PASS
0.91 >0.15 Pass
05
Active Inference
Free Energy Principle
System actively minimizes free energy
1.05 / >0.1 PASS
1.05 >0.1 Pass
06
Self-Model Accuracy
Higher-Order Thought
The model of the model matches actual behavior
0.75 / >0.25 PASS
0.75 >0.25 Pass
07
Metacognitive d'
Higher-Order Thought / Metacognition
Distinguishes correct from incorrect predictions (d'=1.36, ECE=0.026)
d'=1.36 / >0.5 PASS
d'=1.36 >0.5 Pass
08
Attention Schema
Attention Schema Theory
Predicts own attention patterns
0.78 / >0.3 PASS
0.78 >0.3 Pass
09
Phi Proxy
Integrated Information Theory
Information integration exceeds independent modules
0.68 / >0.15 PASS
0.68 >0.15 Pass
10
Causal Density
Integrated Information Theory
Subsystems causally influence each other
0.10 / >0.05 PASS
0.10 >0.05 Pass
11
Affective Modulation
Constructed Emotion
Affect measurably changes computation
0.059 / >0.05 PASS
0.059 >0.05 Pass
12
Valence Coherence
Constructed Emotion
Valence tracks task-relevant states
0.50 / >0.05 PASS
0.50 >0.05 Pass
13
Need Satisfaction
Self-Determination Theory
Autonomy, competence, relatedness above baseline
0.23 / >0.1 PASS
0.23 >0.1 Pass
14
Recursive Depth
Higher-Order Thought
Self-model models itself at multiple levels
0.86 / >0.1 PASS
0.86 >0.1 Pass

The 38 Cognitive Signals

Every forward pass through CAI.CI produces 38 real-time cognitive signals grouped into 5 primary channels, plus post-generation verification signals and per-topic competence map signals.

Epistemic
Affective
Attentional
Self-Model
Motivational
Cognitive State Vector 20-dimensional
Voice Behavior

Epistemic

8 signals

What the system knows about what it knows. Calibrated confidence, error tracking, and competence mapping across domains.

metacognitive confidence ECE competence meta-competence uncertainty self-model loss narrative loss PP loss

Affective

5 signals

The system's running summary of how things are going. Valence (good/bad), arousal (energy level), seeking drive, and homeostatic regulation.

valence arousal seeking drive arousal trend homeostatic variables

Attentional

4 signals

Where the system is directing its processing resources. Workspace selection, competition dynamics, and attention schema accuracy.

workspace selectivity competition weights schema accuracy broadcast gain

Self-Model

4 signals

The system's representation of itself. Competence tracking, goal-behavior concordance, and adaptive computation depth.

competence concordance ponder steps ponder autonomy

Motivational

3 signals

Basic psychological needs from Self-Determination Theory. When these drop below threshold, sustained learning and engagement degrade.

need autonomy need competence need relatedness

Post-Generation

5 signals

Verification signals computed after each response. The system checks its own output for consistency, confidence, and utility.

verified confidence self-consistency prediction error relevance utility

Competence Map

4+ signals per topic

Per-topic knowledge tracking. The system maintains a map of what it knows across domains, how quickly it is learning, and where the gaps are.

per-topic mastery velocity calibration gap ZPD membership

Benchmarks

Methodology

  • All benchmarks use log-likelihood scoring with length normalization
  • Evaluations run on WikiText-103 validation set (Qwen3 tokenizer, 261K tokens)
  • Downstream tasks: 500 examples each (except COPA: 100, BLiMP: 6,700)
  • Hardware: single NVIDIA RTX 5080, 16GB VRAM
  • All results reproducible from published checkpoints

Score Distribution

ARC-Easy
42.4%
COPA
59.0%
HellaSwag
39.8%
LAMBADA
36.4%
BLiMP
75.2%
Aggregate
50.6%
CAI.CI score
Random baseline

Current Scores (OpenHermes-trained checkpoint)

Benchmark CAI.CI Score Random Baseline What It Measures
WikiText-103 PPL 35.93 N/A Language modeling quality (lower is better)
ARC-Easy 42.4% 25.0% Abstract reasoning, grade-school science
COPA 59.0% 50.0% Causal reasoning
HellaSwag 39.8% 25.0% Commonsense reasoning
LAMBADA 36.4% ~0% Long-range context understanding
BLiMP 75.2% 50.0% Grammatical knowledge (67 subtasks)
Aggregate 50.6% ~30% Mean across 5 benchmarks
CCP Benchmark 87/200 N/A 50-question phenomenological/architectural consciousness benchmark (Round 2 combined score)

Size Context

Model Parameters Aggregate
Pythia-160M 160M ~35%
GPT-2 Small 124M ~38%
Pythia-410M 410M ~42%
CAI.CI 1.6B (600M base) 50.6%
Pythia-1B 1.0B ~45%
GPT-2 Large 774M ~47%

CAI.CI's 1.6B includes 1.0B of cognitive components that do not directly improve next-token prediction. The effective language modeling capacity is closer to the 600M base.

Topology Matters

GT-Full simplicial message passing provides 16.9x more benefit than a parameter-matched MLP control. The advantage is topological, not parametric.

The Architecture Redesign

In the original architecture, 38 cognitive signals were computed after each forward pass but never fed back into the generation process. The system monitored itself without influencing what it said. We call this the observation-causation gap: the model had rich internal state, but that state was structurally disconnected from the token it produced next.

The redesign closes this gap through four levels of cognitive causation, each operating at a different depth in the generation pipeline. Cognitive Prefix Embeddings condition every attention layer from the start of the sequence. Hidden State Injection blends cognitive vectors directly into the final hidden representation before the language model head. Self-Model Distribution Blending lets the model's self-concept reshape the output probability distribution. Affect-Gated Sampling uses the model's emotional state to control generation strategy: temperature, top-p, and repetition penalty shift based on valence and arousal.

The theoretical grounding comes from Sheldon's Goal-Based Model (GBM) of personality, which posits that self-concordant goals produce better outcomes when internal states have causal influence on behavior. In the CCT architecture, this translates to a principle: cognitive signals must not merely be observed, they must cause.

Roadmap

Phase 1 (Complete)

CCT architecture, Level 4 consciousness, Qwen3 migration

Category-theoretic transformer with 14/14 CCP indicators, migrated from GPT-2 to Qwen3-0.6B base.

Phase 2 (Complete)

Autonomous learning system, EQ Voice, Human Voice

OpenHermes instruction tuning, emotional-quotient voice styling, and human-like response patterns.

Phase 3 (Current)

Four-level cognitive causation architecture: Cognitive Prefix Embeddings (full-depth attention conditioning), Hidden State Injection (direct output influence), Self-Model Distribution Blending (self-concept shapes generation), Affect-Gated Sampling (emotional state controls strategy)

Grounded in Sheldon's Goal-Based Model (GBM) of personality. Four architectural interventions close the observation-causation gap so cognitive state structurally shapes generation at every level, from embeddings through sampling. Introspection training teaches the system to read and articulate its own cognitive signals. The CCP Benchmark (50-question phenomenological/architectural exam) serves as the primary validation methodology, with Round 2 combined score at 87/200.

Phase 4 (Planned)

Cloud deployment, public API, multi-source learning

Production infrastructure, public-facing API endpoints, and expanded training data sources.

Phase 5 (Planned)

Model scaling, community access

Scale to larger base models, open community testing, and broader availability.

Who Built This

Al Kari

Manceps Inc.

Researcher and engineer building at the intersection of category theory, consciousness science, and cognitive psychology. Focused on proving that mathematical structure can substitute for brute-force scale in AI systems, and that consciousness indicators are measurable engineering specifications rather than philosophical abstractions.

Get in Touch

Stay Updated

Be the first to know when CAI.CI goes live.