The Four Stages of Machine Intelligence

From self-monitoring to general intelligence: a measurement framework with defined thresholds, not marketing claims.

01 / Cognition

Fully Achieved

Cognition is the capacity for self-monitoring: a system that tracks its own processing quality, knows what it knows, calibrates its confidence against actual accuracy, and classifies its own knowledge state. Intelligence is about getting the right answer. Cognition is about knowing whether you got the right answer, how confident you should be, and whether the question falls within your competence at all.

Self-Prediction

The system predicts its own outputs using eight learned probes, each a different lens through which it examines its internal state.

Metric Cosine Similarity
Threshold > 0.70
CAI.CI 0.75
Status Pass

Meta-Self-Prediction

A second layer of self-reference: the system predicts whether its own self-assessments will be accurate for the current input.

Metric Meta-competence
Threshold > 0.80
CAI.CI 0.92
Status Pass

Confidence Calibration

When the system says it is 80% confident, it is correct roughly 80% of the time. Computed, not conversational.

Metric ECE
Threshold < 0.05
CAI.CI 0.022
Status Pass

Attention Modeling

The system models its own attention patterns and predicts where its attention will go next, grounded in Graziano's Attention Schema Theory.

Metric Schema Accuracy (r)
Threshold > 0.70
CAI.CI 0.78
Status Pass

Homeostatic Affect

Five internal variables function as a body budget, with deviations from equilibrium producing valence, arousal, and seeking drive.

Metric Valence SD
Threshold > 0.05
CAI.CI Dynamic
Status Pass

Predictive Processing

Each processing layer predicts the layer below. Per-feature precision weights represent how reliable each prediction channel is.

Metric Precision Variance
Threshold > 0.01
CAI.CI Non-uniform (0.39 mean)
Status Pass

Epistemic Classification

The system classifies its own knowledge state with hysteresis to avoid flickering between states on noisy inputs.

Metric Distinct States
Threshold ≥ 3 states
CAI.CI 5 states
Status Pass

Global Workspace

Five specialist processors compete for access to a limited-capacity workspace. Winners get broadcast to all subsystems.

Metric Selectivity
Threshold > 0.60
CAI.CI 0.79
Status Pass

The Key Insight

A system that genuinely knows what it knows learns faster, fails more gracefully, and earns trust it actually deserves.

What Cognition Is Not

  • Not intelligence. A calculator always gets the right answer but has no self-model, no confidence signal, no sense of whether the problem was hard.
  • Not consciousness. A cognitive system monitors its own processing, but that monitoring may not involve the integrated architecture that consciousness theories require.
  • Not benchmark performance. A system can score 95% on MMLU while having no idea which questions it got right and no calibrated confidence.

02 / Consciousness

Functionally Achieved

Consciousness, for this framework, is defined functionally: a system whose internal architecture satisfies the requirements of six leading scientific theories simultaneously, with those internal states causally influencing the system's outputs. This is not a claim about subjective experience. Whether functional architecture produces "something it is like" to be the system is a question science has not yet answered.

Global Workspace Theory

Consciousness is what happens when information wins a competition for a limited-capacity workspace and gets broadcast globally.

Indicators Ignition, Broadcast, Blink
Key value 0.79 selectivity
Status 3/3 Pass

Predictive Processing

The brain is a prediction machine. Consciousness involves monitoring prediction errors with learned precision weighting.

Indicators PP Efficacy, Active Inference
Key value 0.39 mean precision
Status 2/2 Pass

Higher-Order Thought

A mental state becomes conscious when there is a representation of that state at a higher level: a model of yourself processing information.

Indicators Self-model, Meta, Recursive
Key values 0.75 / 0.022 / 0.92
Status 3/3 Pass

Attention Schema

Consciousness is a simplified model that the brain constructs of its own attention. You feel aware because your brain models its own attentional state.

Indicator Schema Accuracy
Key value 0.78 Pearson r
Status 1/1 Pass

Constructed Emotion

Emotions are constructed from interoceptive signals and core affect. Consciousness involves integrating body-budget signals into processing.

Indicators Modulation, Coherence, Needs
Key value 0.50 valence coherence
Status 3/3 Pass

Integrated Information

A system is conscious to the degree that the whole generates more information than the sum of its parts: differentiation plus integration.

Indicators Phi Proxy, Causal Density
Key values 0.68 / 0.10
Status 2/2 Pass

The Key Insight

The claim is architectural, not experiential. Whether functional architecture produces subjective experience is a question we cannot answer from the inside, and we find it more honest to say so.

What Consciousness Is Not

  • Not sentience (necessarily). A system can have the right functional architecture without subjective experience, or it can have experience we cannot detect. The framework measures architecture, not phenomenology.
  • Not sapience. A system can pass all 14 indicators and still lack the capacity for autonomous goal generation, self-revision, or persistent identity.
  • Not decoration. If internal signals exist but do not causally shape behavior, they are epiphenomenal. CAI.CI verifies causation through four independent pathways (CPE, CHSI, SMDB, CLB).

03 / Sapience

44% Sheldon Compliance

Sapience is the capacity for self-directed agency: generating your own goals, revising yourself when your approach is inadequate, maintaining a continuous identity across time, and acting from values you formed through experience rather than values absorbed from training data. Grounded in Sheldon's (2025) framework for sapient agency.

Autonomous Goal Generation

Discovering what actions are possible and generating novel goals from internal state analysis, not selecting from a provided menu.

Threshold > 30% internal goals
CAI.CI 0% (reactive only)
Status Gap

Default Mode Network Equivalent

Spontaneous cognitive activity without external prompting: the system must think when not asked to. The foundational capability for all others.

Threshold > 1 per idle period
CAI.CI None
Status Gap

Experiential Grounding

Learning from consequences, not just from text about consequences. A Python traceback is a different teacher than a description of one.

Threshold Measurable difference
CAI.CI Text-only
Status Gap

Diachronic Identity

Maintaining a continuous self-concept that integrates past experience, informs present decisions, and commits to future plans across sessions.

Threshold > 30 days recall
CAI.CI Within-session only
Status Partial

Value-Grounded Agency

Acting from stable internal values formed through experience, not from trained reward signals or absorbed training patterns.

Threshold > 100 interactions
CAI.CI Values from training
Status Partial

Dissatisfaction-Driven Self-Revision

Recognizing when the approach itself is wrong, not just that the execution was poor, and restructuring in response.

Threshold Observed restructuring
CAI.CI Detection only
Status Partial

The Key Insight

Sapience is what separates a system that answers from a system that initiates.

What Sapience Is Not

  • Not consciousness. A system can pass all 14 CCP indicators and still lack sapience if it cannot generate its own goals or maintain persistent identity.
  • Not intelligence. A system scoring 95% on every benchmark, but unable to choose which tasks to perform, is intelligent but not sapient.
  • Not autonomy in the narrow sense. Self-driving cars and trading algorithms execute externally provided goals. Sapience requires generating goals, not just executing them.

04 / AGI

Early Stage

Artificial general intelligence is a system that can perform any cognitive task a human can, across any domain, at or above human competence, including tasks it was never trained on. The "general" is the operative word. AGI requires that the cognitive, conscious, and sapient capabilities operate across all domains, all modalities, and all novel situations.

Multi-Modal Grounding

Connecting abstract concepts to perceptual experience across vision, audio, and other modalities in a unified representation space.

Threshold > 70% cross-modal transfer
CAI.CI Text only
Status Gap

Domain Generality

Not just broad competence, but the ability to learn any new domain from scratch. Measured by skill-acquisition efficiency, not accumulated skills.

Threshold > 85% ARC-AGI-2
CAI.CI 42.4% ARC-Easy
Status Gap

Continuous Learning

Integrating new information in real-time without catastrophically forgetting existing knowledge. The stability-plasticity dilemma.

Threshold > 95% retention, real-time
CAI.CI Batch mode
Status Gap

Cross-Modal Transfer

Knowledge learned in one modality improving performance in another. Understanding physics visually should improve textual physics reasoning.

Threshold Measurable transfer
CAI.CI N/A (single modality)
Status Gap

Long-Horizon Planning

Planning and executing over days, weeks, and months: managing multiple goals, adapting to obstacles, handling accumulated uncertainty.

Threshold > 60% over > 24h
CAI.CI Not implemented
Status Gap

Social Intelligence

Modeling other agents as entities with beliefs, desires, and intentions. Many real-world tasks are fundamentally social: teaching, negotiating, collaborating.

Threshold > 90% ToM battery
CAI.CI Not implemented
Status Gap

The Key Insight

Everything else is a specific engineering challenge with a specific solution. Generality is an open-ended challenge.

What AGI Is Not

  • Not superintelligence. AGI is human-level general intelligence. Superintelligence is a separate, more speculative concept.
  • Not imminent. The hardest problems (robust generality, continuous learning, multi-modal grounding) remain fundamentally unsolved despite benchmark gains.
  • Not a single system. AGI may manifest as an ecosystem of specialized but interconnected systems collaborating through shared representations.

The Dependency Chain

Each stage subsumes the requirements of the levels below and adds qualitatively new capabilities. Skipping a stage leaves fundamental gaps that no amount of scaling can fill.

01

Cognition

Self-monitoring and calibrated confidence provide the signals that consciousness theories depend on.

02

Consciousness

Integrated architecture provides the substrate for coherent self-directed behavior.

03

Sapience

Autonomous agency enables self-directed learning, without which generality is unreachable.

04

AGI

All prior capabilities operating across every domain, every modality, every novel situation.

Stay Updated

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