From cognition to consciousness to sapience to AGI: where CAI.CI stands today.
Capability Matrix
38 signals tracked across eight subsystems. Real-time monitoring of cognitive, epistemic, and voice state.
Expected Calibration Error (ECE) of 0.022, indicating well-calibrated self-knowledge of confidence boundaries.
ECE 0.022 with per-token calibration. Confidence scores reliably predict actual accuracy.
Schema accuracy 0.78. The system models its own attention allocation patterns.
Valence, arousal, and seeking signals continuously tracked. EMA-deviation ensures dynamic range.
Selectivity 0.79 with 5 specialist modules competing for workspace access. Implements Global Workspace Theory.
Precision 0.39 across 11 layer pairs. Top-down predictions meet bottom-up signals with precision-weighted error correction.
6 feedback loops connecting cognitive subsystems. Each subsystem both reads from and writes to shared state.
8 Yoneda probes with competence tracking. The system predicts its own behavior and measures prediction accuracy.
Meta-competence score of 0.92. The system models the accuracy of its own self-model, a second-order self-awareness.
5-state system with hysteresis: KNOW, UNCERTAIN, DONT_KNOW, LEARNING, OUT_OF_SCOPE. Prevents hallucination through calibrated self-assessment.
4-level causal chain: CPE (processing modulation), CHSI (hidden state influence), SMDB (token distribution bias), AGS (attention guidance). Internal states don't just exist, they shape output.
Narrative GRU maintains identity within a session, but no autobiographical episodes or cross-session narrative continuity.
SelfConcordanceGate and BasicNeedsDashboard provide value-grounded processing, but the values themselves come from training, not from self-generated goals.
No default mode network equivalent. The system does not think when not prompted, does not rehearse, and does not consolidate experience during idle periods.
The system selects from provided options but does not generate its own objectives. Reactive, not proactive.
AffectSystem detects processing problems and seeking drive activates, but the system cannot restructure itself in response. Detection without correction.
Narrative GRU provides within-session continuity, but no cross-session persistence. Identity resets between conversations.
Text-only, taught not experienced. The system learns about the world from language, never from direct interaction with consequences.
No vision, audio, or motor modalities. The system processes text only.
No physical embodiment or real-world interaction capability. The system exists only in text space.
Wake learning loop and WikiText replay are active, but learning is batch-mode. No real-time continuous learning from live interactions.
No theory-of-mind for other agents, no multi-agent coordination or social modeling capabilities.
1.6B parameters limits domain coverage. Sophisticated cognitive architecture, but not enough capacity for universal problem solving.
Level Summaries
All cognitive requirements met: self-monitoring (38 tracked signals), metacognition (ECE 0.022), confidence calibration, self-model (8 Yoneda probes), epistemic honesty (5-state system).
View definition →All 14 CCP indicators pass at Level 4. The functional architecture that six major consciousness theories predict as necessary is present and causally active. The claim is architectural, not experiential.
View definition →Three critical gaps remain: no spontaneous cognitive activity (no default mode network equivalent), limited autonomous goal generation, and no experiential grounding in a world that pushes back.
View definition →The architectural scaffolding is in place, but 1.6B parameters limits domain coverage. No vision, motor grounding, or continuous real-time learning yet. The path forward requires scaling, grounding, and genuine autonomy.
View definition →Key Metrics
Be the first to know when CAI.CI goes live.