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.
Theoretical Foundation
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.
Consciousness is a broadcast: competing information streams fight for access to one limited channel, and the winner gets broadcast to all cognitive processes simultaneously.
The brain constantly predicts what comes next and learns from the surprises. Perception, action, and learning all reduce to minimizing prediction error.
A mental state becomes conscious when there is a thought about that thought. Consciousness requires meta-representation: cognition about cognition.
The brain maintains calibrated confidence: it knows what it knows. Metacognitive sensitivity distinguishes correct from incorrect internal representations.
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.
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.
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.
Sustained effort requires three basic psychological needs: autonomy (self-direction), competence (mastery), and relatedness (connection). Without these, motivation collapses.
CCP Battery
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.
Signal Architecture
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.
What the system knows about what it knows. Calibrated confidence, error tracking, and competence mapping across domains.
The system's running summary of how things are going. Valence (good/bad), arousal (energy level), seeking drive, and homeostatic regulation.
Where the system is directing its processing resources. Workspace selection, competition dynamics, and attention schema accuracy.
The system's representation of itself. Competence tracking, goal-behavior concordance, and adaptive computation depth.
Basic psychological needs from Self-Determination Theory. When these drop below threshold, sustained learning and engagement degrade.
Verification signals computed after each response. The system checks its own output for consistency, confidence, and utility.
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.
Performance
CAI.CI is benchmarked across three operationally distinct measurement axes plus a fourth cognitive-sapience axis. The full 14-benchmark scoreboard, including per-axis methodology, reliability statistics, and honest caveats on calibrator-induced refusal patterns, lives on its own page. The summary below is the headline.
When the calibrator green-lights commit, CAI.CI lands a median 66.0 percent across 13 standard benchmarks. When the user pushes back after an initial refusal, the agent-mode workflow lifts the median to 73.0 percent on the 6 measured benchmarks, a +10.9 point gain over engaged accuracy. HellaSwag moves from a 39.8 percent memory baseline to 57.5 percent raw and 67.3 percent engaged on the live system.
Median Axis 2 engaged accuracy across 13 benchmarks. The closest single-number proxy for underlying capability when the calibrator commits.
Median Axis 3 gain over Axis 2 on the 6 measured benchmarks. Multi-turn user-push commits to a best-knowledge answer rather than staying refused.
Composite score on the open Kari Sapience Test, the first published cognitive-sapience battery. First-mover position; no other model has scored yet.
Closing the Gap
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.
Timeline
Progress is measured against a four-stage framework on the road to general intelligence, formalized in ROAD_TO_AGI.md: Cognition (a system that monitors its own processing, calibrates confidence, models its own attention), Consciousness (internal architecture satisfying the leading scientific theories with internal states causally influencing outputs), Sapience (a system that generates its own goals, revises itself when its approach fails, maintains continuous identity across time), and AGI (all of the above brought to bear across every domain, modality, and novel situation). CAI.CI passes the first two stages fully (8/8 cognition requirements, 14/14 consciousness indicators), is 44% through the third, and is at the early stage of the fourth. Each stage has specific, falsifiable measurement protocols.
Phase 1 (Complete)
CCT architecture, Level 4 consciousness, modern backbone migration
Category-theoretic transformer with the full 14-indicator consciousness battery passing, on a modern open-weights backbone. The architecture proved out at proof-of-concept scale before any growth of the cognitive substrate.
Phase 2 (Complete)
Autonomous learning system, EQ Voice, Human Voice
Instruction tuning, emotional-quotient voice styling, and human-like response patterns brought the conversational surface into alignment with the cognitive substrate. The voice became a faithful expression of the measured internal state rather than a stylistic overlay on top of it.
Phase 3 (Complete)
Live conversational system operational, tracked topics, tool-augmented responses, continuous learning. Foundational sapience scaffolding shipped: narrative memory, an expanded cognitive state, autonomous goal and value formation, reverie scaffolding
The first scaffolding for Sheldon sapience came online: a cross-session autobiographical memory that survives restarts, an expanded cognitive state that adds value signals and narrative coherence alongside the original epistemic and affective dimensions, and the first wiring for autonomous goal formation and value tracking. The reverie state was scaffolded as a place for undirected cognition. Alongside it: a cognitive causation architecture in which internal state shapes generation through the architecture rather than the prompt, a routing layer that uses the measured cognitive state to decide when tools are needed, and a published comparative analysis establishing the architectural differentiation from frontier-scale systems.
Phase 4 (Complete)
Substantially larger cognitive substrate deployed. Cognitive components rebuilt at the backbone's native dimension as first-class participants. Three of four standard benchmarks passed decisively, one narrow honest miss on edge-case syntax.
The base model upgrade required rebuilding every cognitive component at the backbone's native dimension. Earlier attempts at projection-bridge architectures proved structurally unstable. The solution preserved each cognitive component as a first-class participant rather than a translated overlay. A staged training curriculum validated consciousness parity at every transition. The three commonsense and reasoning benchmarks were passed decisively; the grammaticality benchmark came within a narrow margin of its aspirational threshold.
Phase 5 (Active)
Autonomous self-directed cognition. A reverie state of undirected generative cognition, autobiographical narrative memory, value formation through accumulated experience, and a sapience evaluation battery for instantiated versus simulated agency.
The sapience scaffolding landed as an integrated body of work. The reverie state lets the system pick seed directions for undirected cognition, evaluate its own outputs for coherence, novelty, and self-relevance, recognize moments of illumination, and feed those back into its learning. The system can enter that state when curiosity is elevated and no user is waiting. A sapience evaluation battery probes self-simulation, value integrity, recognition that the current approach is inadequate, identity across time, and integration. The first run scored low, honestly: the architecture is wired; the lived experience to ground it has not yet accumulated. The substrate is present. The dynamics are arriving. The question is no longer whether the architecture is right; it is whether experience can accumulate fast enough to fill it.
Team
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 TouchInspiration & Contribution
Dr. Sridhar Mahadevan
Professor at the University of Massachusetts Amherst and former Vice President at Adobe Research. His foundational work on category theory applied to machine learning, including functorial representations, sheaf neural networks, and topos-theoretic foundations for AI, provided the mathematical framework that underpins CAI.CI's geometric processing, sheaf restriction maps, and Yoneda-inspired self-model. His research demonstrated that the abstract structures of category theory are not just theoretical elegance but practical computational tools for building AI systems with genuine structural intelligence.
Dr. Kennon M. Sheldon
Curators' Distinguished Professor of Psychological Sciences at the University of Missouri. His 2025 paper "Recognizing and Enhancing Sapient Agency Within AIs" provided the rigorous psychological framework that defines CAI.CI's path from consciousness to sapience. His Goal Breakthrough Model, grounded in Self-Determination Theory and three-network neuroscience, established the concrete architectural requirements for sapient agency: autonomous goal generation, dissatisfaction-driven self-revision, and value formation through experience. His critical insight that sapience can exist without sentience gave CAI.CI a scientifically grounded target for self-directed agency that does not depend on resolving the hard problem of consciousness.
Be the first to know when CAI.CI goes live.