📊 Research Program

Continuous Intelligence Measurement

Rigorous Investigation of Dynamic Intelligence Measurement

Core Methodological Claim:

"The most informative data about intelligence may be generated during the continuous performance of intelligent activity rather than recovered from its final products."

This is not established fact. This is a testable methodological hypothesis. CIM investigates whether continuous behavioral trajectories—observations, predictions, actions, feedback, and learning—explain additional variance in future adaptive performance beyond static outcome measures alone.

Program Identity

CIM as Parallel Research Initiative

LEARN

Experimental Infrastructure

→ Persistent observation system
→ Continuous data collection
→ Research notebook and logging
→ Multiple sensory channels (digital + physical)

CIM

Research Program

→ Defines research questions
→ Specifies measurement protocols
→ Analyzes experimental results
→ Can leverage LEARN for execution

Architectural Relationship: LEARN provides the instrument. CIM defines the experiments. Neither replaces the other; both are essential components of a rigorous research infrastructure.

The Methodological Problem

Why Static Measurement May Be Incomplete

Static Measurement

→ Input → Task → Output → Score

→ Compresses behavior into single metric

→ Loses temporal information

→ Cannot distinguish lucky success from robust adaptation

Dynamic Measurement

→ Observe → Represent → Hypothesize → Predict → Act → Feedback → Update → Learn

→ Preserves process trajectory

→ Captures adaptation dynamics

→ Enables mechanism investigation

Information Lost in Static Compression

Temporal information: When did learning occur?

Error trajectories: How were mistakes recovered?

Adaptation dynamics: How did strategy change?

Learning rate: How fast did performance improve?

Transfer capability: What generalized to new tasks?

Information acquisition: How was attention directed?

Research Question: Does this lost information contain scientifically valid predictive power? Can continuous process metrics explain additional variance in future performance beyond static scores?

Formal Model

Intelligence as Trajectory

System State as Tuple

S_t = (X_t, Z_t, H_t, A_t, F_t, M_t)

X_t = observations at time t

Z_t = internal representations

H_t = hypotheses and counterfactual models

A_t = actions taken

F_t = environmental feedback received

M_t = internal model and state

State Evolution

S_(t+1) = F(S_t, X_t, A_t, F_t)

The next system state is a function of the current state and the observed feedback to the current action. This captures the core feedback loop: act, observe result, update.

Intelligence as Process

I_process = Φ(S_0:T)

Intelligence is a functional property of the full trajectory from initial state through final state. It is not reducible to any single snapshot or outcome.

Critical Note: This is a proposed formalization, not an established law. The appropriateness of this representation and the correct definition of Φ are empirical questions that experiments must address.

Intelligence Event Stream

Ontology and Data Structure

Event Types

O - Observation

Sensory data from environment

R - Representation

Internal model update

P - Prediction

Forecasted next state

H - Hypothesis

Proposed explanation

C - Counterfactual

Alternative scenario

D - Decision

Committed action

A - Action

Executed behavior

F - Feedback

Environmental response

E - Error

Prediction-observation gap

V - Revision

Model update based on error

L - Learning

Measurable improvement

T - Transfer

Generalization to new context

S - Segmentation

Discrete cognitive state transition

Discrete Cognitive Segmentation Events (S)

Segmentation events mark automatic, load-dependent transition points where the system moves between processing states. They represent measurable temporal signatures of cognitive state changes, complementary to residual magnitude.

Types: Free-energy settling | Working-memory consolidation | State-transition reset | Attention-gating commit | Prediction-error significance threshold

Load-Dependence: Frequency and timing vary with information-processing demand. High demand may suppress segmentation events; uncertainty resolution triggers consolidation events.

Event Record Schema

Event {
  timestamp: ISO8601
  system_id: string
  environment_id: string
  event_type: O | R | P | H | C | D | A | F | E | V | L | T | S
  observation: any | null
  representation: any | null
  prediction: any | null
  hypothesis: string | null
  action: any | null
  feedback: any | null
  error: number | null
  update: any | null
  confidence: float [0,1]
  observability: DIRECTLY_OBSERVABLE | INFERRED | SELF_REPORTED | MODEL_DERIVED | VERIFIED
  provenance: string

  // Segmentation event fields (when event_type == S):
  segmentation_type: FREE_ENERGY_SETTLING | CONSOLIDATION | STATE_TRANSITION | ATTENTION_GATING | ERROR_THRESHOLD
  information_demand: float  // Processing demand at time of event
  significance_metric: float // Magnitude of state change
  suppression_window: boolean // True if segmentation was suppressed
}

Critical Distinction: The observability field is mandatory. Not all quantities can be directly observed. The system must explicitly distinguish between what can be directly measured, what must be inferred, what is self-reported, and what is derived from models.

Process Metrics Framework

Candidate Dimensions Under Investigation

Adaptation Rate

A_r = ΔPerf/Δt

hypothesis

Prediction Accuracy

P_acc = (1 - error)/total

hypothesis

Error Utilization

E_util = improvement/error_magnitude

hypothesis

Recovery Efficiency

R_eff = time_to_baseline/perturbation_magnitude

hypothesis

Exploration Efficiency

Exp_eff = information_gain/exploration_cost

hypothesis

Hypothesis-Generation Efficiency

H_gen = useful_hypotheses/total_generated

hypothesis

Counterfactual Utility

C_util = E[outcome | counterfactual] vs actual

hypothesis

Transfer Efficiency

T_eff = performance_new_task/learning_time_old

hypothesis

Behavioral Plasticity

B_plast = diversity_of_actions/task_context

hypothesis

Temporal Coherence

T_coh = correlation(prediction_t, observation_t)

hypothesis

Information Gain

IG = I(Z_t+1; Y_t+1) - I(Z_t; Y_t)

hypothesis

Model Improvement

M_imp = reduction_in_prediction_error_rate

hypothesis

Status Label: All metrics listed here are candidate measures under investigation. Each requires operationalization, validation, and evidence of incremental validity. None should be treated as established without experimental support.

Central Hypotheses Registry

Testable Predictions to Guide Experiments

H1: Process Superiority

proposed

Continuous process measurements explain variance in future performance beyond static outcome measures.

Process metrics Static scores Future performance

H2: Adaptation

proposed

Faster and more appropriate adaptation following environmental perturbation predicts greater general adaptive capacity.

Perturbation response time Strategy revision quality Adaptive capacity

H3: Error Correction

proposed

The efficiency with which prediction errors produce useful behavioral updates predicts subsequent performance.

Error utilization Update efficiency Performance improvement

H4: Counterfactual Utility

proposed

The quality/usefulness of counterfactual representations predicts adaptive behavior.

Counterfactual quality Behavioral adaptation Outcome prediction

H5: Transfer

proposed

Process metrics generalize across tasks better than task-specific outcome measures.

Process generalization Task transfer Cross-domain validity

H6: Real-Time Prediction

proposed

Partial trajectories can predict eventual performance before task completion.

Partial trajectory Completion time Outcome prediction

H7: Causal Intervention

proposed

Manipulating process variables produces predictable changes in adaptive performance.

Intervention type Process change Performance change

For Each Hypothesis: Formal experiments will specify the null hypothesis, measurable variables, proposed experiment, potential falsification evidence, and current empirical status.

The Perturbation Experiment

Flagship Experimental Design for Testing Adaptation

Protocol

Establish baseline

Observe system performing task under stable conditions

Establish baseline predictions

System makes predictions about what will happen next

Introduce controlled perturbation

Change one environmental variable in a controlled way

Measure prediction error

Quantify the gap between predicted and observed state

Observe hypothesis generation

What new hypotheses does system generate?

Observe strategy/action change

How does behavior change in response to perturbation?

Measure recovery

How long until performance returns to baseline?

Introduce novel related perturbation

Similar but previously unseen environmental change

Measure transfer

Does knowledge from first perturbation help with second?

Analyze process metrics

Calculate all candidate metrics from trajectory data

Compare to final score

Test whether process metrics predict future performance better than task completion alone

Core Trajectory Loop

Observation

→

Prediction

→

Reality

→

Error

→

Hypothesis

→

Action

→

Update

Confound Control Registry

Systematic Identification and Mitigation

Task Difficulty

Mitigation: Randomize task order; use difficulty controls

Computational Resources

Mitigation: Equalize hardware; measure actual resource usage

Prior Training

Mitigation: Document training history; control for experience level

Domain Familiarity

Mitigation: Use novel domains; measure pre-task familiarity

Response Latency

Mitigation: Measure and report timing separately

Sensor Quality

Mitigation: Use calibrated sensors; document specifications

Measurement Error

Mitigation: Use multiple independent measurements

Evaluator Bias

Mitigation: Blind evaluators to condition; use multiple raters

Prompt Effects

Mitigation: Standardize instructions; use fixed templates

Benchmark Contamination

Mitigation: Hold test sets separate; use withheld data

Environmental Predictability

Mitigation: Vary environment properties systematically

Reward Hacking

Mitigation: Monitor for metric gaming; validate against real-world outcomes

Mandatory Discipline: Every experiment must document which confounds apply, what mitigation strategies are employed, and what residual uncertainty remains. This is not optional—it is core to experimental credibility.

Research Roadmap

Seven-Phase Evolution of the Program

Phase I

Formalization

Define formal trajectory model, event ontology, process metrics

Phase II

Simulation

Test hypotheses in controlled software simulations

Phase III

Human/AI Experiments

Controlled lab experiments with humans and AI agents

Phase IV

Physical Experimentation

Embodied testing with LEARN + Raspberry Pi observatory

Phase V

Cross-Substrate Validation

Compare process metrics across humans, AI, robots, hybrids

Phase VI

Causal Intervention

Manipulate process variables; measure outcome effects

Phase VII

General Theory

Integrate results into formal theory of intelligence measurement

Current Status: The program is in Phase I (Formalization). Formal model, event ontology, candidate metrics, and central hypotheses are being specified. Phase II (Simulation) will follow once formalization is complete and human/research team validation confirms the framework is sound.

Evidence Discipline

Distinguishing Claim Types Throughout the Research

HYPOTHESIS

Proposed testable prediction

PROPOSED EXPERIMENT

Planned empirical test (not yet run)

OBSERVATION

Raw empirical data collected

RESULT

Analyzed outcome of experiment

INTERPRETATION

Meaning or implication of result

SPECULATION

Conjecture beyond current evidence

Mandatory Rules

→ Never present speculation with the visual weight of results
→ Label unexecuted experiments as PROPOSED, not CONFIRMED
→ If no empirical data exists for a claim, state: "Not yet empirically tested"
→ Distinguish interpretation from pure observation—mark each clearly
→ Never fabricate data, citations, or validation status

Falsification Criteria

What Evidence Would Weaken or Refute the Core Claim

Process metrics fail to explain additional variance beyond static scores

Process metrics fail to generalize across different task domains

Apparent process effects disappear after controlling for confounds

Process metrics are primarily proxies for task familiarity or training duration

Continuous data adds computational cost without improving predictions

Candidate metrics cannot be reliably measured across different systems

Human/AI/robotic results fail to show meaningful commonality in process structure

Interventions on proposed process mechanisms do not produce predicted outcome changes

Static outcome measures alone predict future performance as well as process metrics combined with static measures

Scientific Commitment: This is not a rhetorical gesture. If evidence supports any of these falsification criteria, the core hypothesis must be rejected or substantially revised. The research program is designed to uncover truth, not defend a predetermined conclusion.

Academic Positioning

Interdisciplinary Foundations and Research Connections

Cognitive Science

How minds process information and adapt to changing environments

Artificial Intelligence

Machine learning theory, adaptive systems, and behavioral modeling

Measurement Theory

Psychometrics, validity, reliability, and construct operationalization

Information Theory

Entropy, mutual information, and information-theoretic measures of learning

Dynamical Systems

Trajectory analysis, state-space methods, and temporal modeling

Experimental Methodology

Causal inference, randomization, control, and statistical design

Computational Neuroscience

Predictive processing, error-driven learning, and neural mechanisms

Philosophy of Science

Falsifiability, operationalization, and scientific realism in intelligence research

CIM does not claim to solve intelligence. Rather, it proposes an operationalizable measurement methodology and tests it empirically. It situates intelligence measurement within established scientific frameworks and respects the limitations of current knowledge.

Future contributions will include: Peer-reviewed literature integrating these foundations, formal proofs of metric properties where applicable, open datasets enabling independent verification, and collaborative research partnerships with cognitive scientists and AI researchers.

LEARN + CIM Physical Observatory

Embodied Experimentation Infrastructure

Experimental Design

→ Observation Environment: Controlled physical laboratory space with Raspberry Pi cameras positioned to capture continuous visual data

→ Ground Truth: Environmental variables measured with calibrated sensors (temperature, light, motion, object positions)

→ Predictions: Computer vision system predicts next state of observed scene based on current observations and learned models

→ Perturbations: Controlled environmental changes test system's ability to detect anomalies and adapt predictions

→ Real-Time Analysis: Event stream of observations, predictions, errors, and model updates captured continuously

Measurements

Visual Prediction Error

Pixel-level or semantic difference between predicted and observed visual state

Recovery Time

Duration until prediction accuracy returns to baseline after perturbation

Transfer Efficiency

How quickly predictions improve when exposed to a similar (but novel) visual pattern

Model Update Rate

Frequency and magnitude of internal model changes in response to prediction errors

Scientific Purpose: The physical observatory bridges the gap between software simulations and embodied AI/robotics. It tests whether continuous measurement methodology generalizes from digital to photonic systems, and provides ground-truth validation unavailable in simulation.

Engage with the Research

This is an active research program. Explore the formalization, track experimental progress, and understand how CIM leverages LEARN as an experimental instrument.