LEARN

Residual Observatory

LEARN's quantitative and qualitative measurement layer for the Physics of Knowledge framework.

Instead of a single "intelligence score," the Residual Observatory tracks six distinct residual categories, each measuring a different aspect of LEARN's alignment with empirical reality.

Core Principle

Intelligence is not measured by assertions about what a system claims to know. Intelligence is measured by the rate at which a constrained system reduces empirically measurable discrepancy between its internal representations, predictions, and observed reality while preserving adaptive performance under changing conditions.

This is a research hypothesis, not an established scientific fact.

Six Residual Categories

Each measures a different dimension of system-reality alignment

Logical Residual

R_L Operationalized

Measures internal contradictions, undefined terms, inconsistent definitions, broken links, and incomplete mappings within LEARN's documentation and architecture.

Measurable Variables:

  • Contradictory statements in documentation
  • Undefined technical terms in architecture
  • Inconsistent function/variable naming
  • Broken internal cross-references
  • Incomplete service mappings
  • Missing section connections

What This Measures:

A high logical residual indicates the system's self-documentation is incoherent. This makes it difficult for LEARN to reason about its own architecture.

Predictive Residual

R_P Operationalized

Measures the discrepancy between predictions generated before experimentation and outcomes observed after experimentation.

R_P = (1/N) Σ L(ŷ_i, y_i)

Measurable Variables:

  • Prediction accuracy (%) on held-out tasks
  • Mean absolute error between predicted and observed outcomes
  • Prediction variance vs. observed variance
  • Confidence calibration (predicted confidence vs. actual accuracy)
  • False positive rate on predicted state changes

What This Measures:

A decreasing predictive residual indicates LEARN's internal models are becoming better aligned with actual outcomes. This is empirical evidence of improved predictive capability.

Empirical Residual

R_E Operationalized

Measures the gap between LEARN's theoretical claims about intelligence and independently verified empirical evidence.

Measurable Variables:

  • Supported vs. contradicted claims by independent evaluators
  • Replication rate across different experimental conditions
  • Effect size vs. claimed magnitude
  • Confound control adequacy
  • External dataset validation (does the model generalize?)

What This Measures:

A high empirical residual indicates strong claims that lack supporting evidence. A system cannot validate itself—external evaluation is essential.

Adaptation Residual

R_A Operationalized

Measures whether modifications to architecture, models, or strategies produce the predicted performance improvements.

ΔR = R_before - R_after

Measurable Variables:

  • Predicted improvement magnitude vs. actual improvement
  • Improvement persistence (is it temporary or sustained?)
  • Unintended side effects (performance on other metrics)
  • Resource cost vs. benefit ratio
  • Reversibility (can the system roll back if adaptation fails?)

What This Measures:

A positive adaptation residual (R_after < R_before) confirms that an architectural change improved performance as expected. Without measurement, "improvements" are assumptions.

Embodiment Residual

R_B Operationalized

Measures the gap between LEARN's digital-only current state and its planned physical embodiment via Phase 2 (Raspberry Pi cameras).

Measurable Variables:

  • Predictions made in digital domain vs. validated in physical domain
  • Latency: model inference time vs. real-world interaction timing requirements
  • Sensor limitations: camera resolution vs. model assumptions
  • Environmental complexity: controlled simulation vs. real-world noise
  • Embodied learning: improvements in digital transfer to physical

What This Measures:

Currently, R_B is large because LEARN is purely digital. Phase 2 aims to reduce R_B by grounding predictions in physical reality.

Validation Residual

R_V Operationalized

Measures the gap between LEARN's self-reported results and independent evaluations.

Measurable Variables:

  • Agreement between LEARN's assessment and independent auditor assessment
  • Blind replication success rate (does external lab reproduce results?)
  • Hypothesis confirmation bias detection
  • Measurement method agreement (different observers same result?)
  • Publication acceptance in external peer review

What This Measures:

A system cannot establish its own validity. External validation is the only way to distinguish genuine capability from self-serving optimism.

Residual Total Model

Weighted combination of all six residuals

Rtotal = wLRL + wPRP + wERE + wARA + wBRB + wVRV

where Σ wi = 1

w_L × R_L

Logical coherence weight

w_L = 0.10

w_P × R_P

Predictive accuracy weight

w_P = 0.25

w_E × R_E

Empirical evidence weight

w_E = 0.25

w_A × R_A

Adaptation efficacy weight

w_A = 0.15

w_B × R_B

Embodiment readiness weight

w_B = 0.15

w_V × R_V

Independent validation weight

w_V = 0.10

Research Status

The weights (wi) are currently set to reasonable defaults based on Physics of Knowledge principles. However, the optimal weight distribution is itself an open research question. LEARN will experimentally determine whether different weights produce better predictions across diverse tasks.

Critical Distinction

Rwebsite

Improvements to site organization, navigation, or documentation. These reduce logical residual but do not measure intelligence.

Rresearch

Improvements to hypothesis testing, prediction accuracy, and experimental methodology. These reduce research residuals.

Rintelligence

Measurable capability to reduce discrepancy while preserving adaptation. This requires integrated reduction across multiple residuals.

Therefore: A better website is not automatically better research. Better research is not automatically better intelligence. Each layer must be measured independently.

Measuring Residuals in Practice

From theory to operational measurement

Daily Frame Protocol

Every autonomous cycle records residuals before and after modifications. The Daily Frame captures:

  • • Current residual state (RL, RP, RE, RA, RB, RV)
  • • Hypothesis being tested
  • • Prediction generated before experiment
  • • Experiment performed and measurement method
  • • Residual delta: ΔR = Rafter - Rbefore
  • • Falsification status (supported/weakened/inconclusive/falsified)

Residual Delta as Evidence

A modification is only justified if it produces documented residual reduction:

ΔR = Rbefore - Rafter
If ΔR > 0: Residual decreased (improvement signal)
If ΔR = 0: No improvement (do not implement)
If ΔR < 0: Residual increased (modification harms)

Modifications are only retained if evidence supports retention.

Segmentation Events and Residual Thresholds

Temporal dynamics of gap closure

Beyond measuring the magnitude of residual reduction, LEARN also measures the temporal structure through which that reduction occurs. Discrete cognitive segmentation events mark the automatic transition points where the system moves from one processing state to another.

Relationship Between Residuals and Segmentation

Large prediction errors (high RP) trigger increased segmentation event frequency. The system automatically commits and resets, attempting to resolve the discrepancy. Conversely, periods of low residual are characterized by suppression of segmentation events—the system enters a protective, continuous-intake mode.

B(t) = f(ΔI(t)) where B(t) is segmentation frequency and ΔI(t) is information demand

High Residual State

  • • High prediction error RP
  • • Frequent segmentation events
  • • Short inter-commit intervals
  • • Rapid state transitions
  • • Consolidation-dominant mode

Low Residual State

  • • Low prediction error RP
  • • Infrequent segmentation events
  • • Long inter-commit intervals
  • • Steady-state processing
  • • Suppression-protective mode

This temporal structure provides a real-time, non-invasive signature of cognitive state—complementary to and independent of residual magnitude. Intelligence is therefore observable in both the rate of residual reduction and in the patterned timing of the discrete events through which that reduction is achieved.

What LEARN Does Not Yet Know

Unresolved: Is residual-reduction rate the right metric for intelligence, or should other dimensions be weighted more heavily?

Unresolved: Do improvements in digital residuals transfer to physical embodiment (Phase 2)?

Unresolved: Does the current weight distribution (wL=0.10, wP=0.25, etc.) optimize for adaptive intelligence, or should weights be task-dependent?

Unresolved: Can external evaluators reproduce LEARN's residual measurements independently?

The Next Layers

With the Residual Observatory in place, we can now build the experimental infrastructure:

Hypothesis Registry — Pre-register all major hypotheses with predictions and falsification criteria

Experiment Registry — Track every experiment, control condition, and measured residual delta

Research State Dashboard — Real-time visualization of all active hypotheses, residuals, and experiment status