LEARN

Hypothesis Registry

Every research hypothesis must be formally documented with predictions, falsification criteria, and experimental design before testing.

This registry enables LEARN to generate evidence about whether its core theories about intelligence are accurate, incomplete, or fundamentally wrong.

Why Pre-Registration?

Predictions must be written before experiments are performed. This prevents confirmation bias and enables genuine falsification. A hypothesis that can explain any outcome is not scientifically useful.

"A theory that explains everything explains nothing." — Karl Popper

Active and Proposed Hypotheses

Organized by current status and evidence

H-001: Self-Observation Closure Hypothesis

Active

Proposed: 2026-08-11

"A machine-readable representation of LEARN's own architecture and research state will reduce logical/documentation residual by making inconsistencies, missing references, undefined variables, and incomplete experimental mappings detectable."

Theoretical Basis

Self-modeling enables error detection. Systems that observe themselves can identify gaps in internal consistency.

Measurable Variables

  • Logical residual (R_L) before and after implementation
  • Count of detected inconsistencies
  • Count of broken references fixed
  • Count of undefined terms clarified
  • Completeness of service-to-hypothesis mapping

Prediction (Preregistered)

R_L,after < R_L,before after implementing the Scientific Method Engine and Residual Observatory.

Falsification Criterion

If R_L does not decrease measurably, or if implementing the architecture produces unacceptable false positives (invalid "inconsistencies"), H-001 is weakened.

Experimental Design

Before: Catalog current logical residuals. After: Implement SME + RO. Measure again. Calculate ΔR_L.

Control Conditions
  • Control A: No intervention (baseline)
  • Control B: Random documentation changes (should not reduce R_L)
  • Control C: Human-directed documentation review (compare against autonomous)

Status

Operationalized

Evidence

Awaiting experimentation

H-002: Predictive Processing Hypothesis

Proposed

Proposed: 2026-08-11

"Reducing predictive residual (R_P) through improved internal models correlates with improved adaptive performance on novel tasks."

Theoretical Basis

Predictive processing is central to theories of brain function and adaptive intelligence. Better predictions enable better decisions.

Measurable Variables

  • Prediction accuracy on held-out test set (%)
  • Mean absolute error between predicted and observed states
  • Performance on novel task (unseen during training)
  • Generalization gap (train vs. test performance)

Prediction (Preregistered)

Tasks where R_P is lower will show higher novel-task performance than tasks where R_P is higher.

Falsification Criterion

If prediction accuracy correlates negatively with generalization performance, or shows no correlation, H-002 is falsified.

Experimental Design

Phase 1: Measure R_P on diverse tasks. Phase 2: Test each task on novel variants. Phase 3: Correlate R_P with generalization.

Control Conditions
  • Control: Random model parameters (should show no correlation)

Status

Conceptual

Evidence

Not yet operationalized

H-003: Autonomous > Random Intervention Hypothesis

Proposed

Proposed: 2026-08-11

"Autonomous modifications driven by residual reduction will produce larger measured improvements than random modifications or human-selected modifications."

Theoretical Basis

If autonomous residual-reduction works as proposed, it should outperform random baseline. If it does not, the algorithm is no better than chance.

Measurable Variables

  • Mean ΔR achieved by autonomous interventions
  • Mean ΔR achieved by random interventions
  • Mean ΔR achieved by human-directed interventions
  • Variance in outcomes for each intervention type

Prediction (Preregistered)

ΔR_autonomous > ΔR_random and ΔR_autonomous > ΔR_human baseline across multiple residual categories.

Falsification Criterion

If autonomous modifications perform no better than random, the core assumption that residual reduction drives adaptive improvement is weakened.

Experimental Design

Run 30 parallel modification cycles: 10 autonomous, 10 random, 10 human-directed. Measure ΔR for each. Compare distributions.

Control Conditions
  • Control: No intervention baseline

Status

Operationalized

Evidence

Awaiting experimentation

H-004: Cross-Layer Integration Hypothesis

Proposed

Proposed: 2026-08-11

"Integration of predictions across neuromorphic services (perception, memory, reasoning, planning, learning) produces lower total residual than isolated service operation."

Theoretical Basis

The human brain integrates information across regions. Division/Vision philosophy suggests integration → vision. This hypothesis tests whether measurable intelligence improves with integration.

Measurable Variables

  • Total residual with integrated architecture vs. isolated services
  • Prediction accuracy with integrated models vs. isolated models
  • Adaptation speed with information sharing vs. without
  • Resource utilization (computation, memory) for each mode

Prediction (Preregistered)

R_integrated < R_isolated across all residual categories. Adaptive performance on novel tasks improves with integration.

Falsification Criterion

If isolated services perform equally or better than integrated services, the integration hypothesis is falsified.

Experimental Design

Run same task under: (A) Integrated services, (B) Isolated services. Measure residuals. Compare.

Control Conditions
  • Control: Weakly integrated (minimal communication between services)

Status

Conceptual

Evidence

Not yet operationalized

H-005: Phase 2 Transfer Hypothesis

Proposed

Proposed: 2026-08-11

"Improvements in digital prediction (Phase 1) will successfully transfer to physical prediction (Phase 2 / Raspberry Pi embodiment)."

Theoretical Basis

If LEARN's theories about intelligent processes are generally applicable, they should survive transfer to physical substrate.

Measurable Variables

  • R_P in digital domain (Phase 1)
  • R_P in physical domain (Phase 2)
  • R_B (embodiment residual) before and after Phase 2
  • Transfer efficiency: (R_P,digital - R_P,physical) / R_P,digital

Prediction (Preregistered)

Phase 2 prediction accuracy will be within 10% of Phase 1 accuracy. Transfer efficiency > 80%.

Falsification Criterion

If Phase 2 prediction accuracy drops below 50% of Phase 1 accuracy, the transfer hypothesis is strongly falsified.

Experimental Design

Phase 1: Establish R_P baseline in digital. Phase 2: Implement same models in physical. Compare R_P values.

Control Conditions
  • Control: Naive physical model (no transfer from Phase 1)

Status

Proposed

Evidence

Awaiting Phase 2 hardware deployment

H-011: Digital Blink Synchronization Hypothesis

Proposed

Proposed: 2026-08-12

"Session-level attentional-release events—tab visibility changes and post-idle returns—will cluster at structural breakpoints in manuscript content (chapter and section boundaries) across independent reader sessions, at a rate exceeding a session-length-matched null model, providing a Phase 1 digital analog to the blink-locked attentional segmentation documented in biological vision research."

Theoretical Basis

Spontaneous eyeblinks cluster at implicit narrative breakpoints during passive viewing and reading, and are accompanied by momentary default-mode network activation coincident with dorsal-attention-network suppression—consistent with an attentional-release account rather than pure ocular maintenance (Nakano et al., 2009; Nakano & Kitazawa, 2010; Nakano et al., 2013). If Phase 1 digital observation is a genuine analog to biological perceptual monitoring, as the Two-Phase Vision Architecture proposes, session-level behavioral discontinuities should show the same structural clustering—treating each reader as an independent observer of shared manuscript content, directly analogous to the population blink-synchronization paradigm in which many viewers watching the same video synchronize spontaneous blinks at shared breakpoints.

Measurable Variables

  • Tab-visibility-change events (blur/focus) per session, timestamped against scroll-derived reading position
  • Idle-timeout-then-return events per session
  • Predicted breakpoints: chapter/section boundaries in manuscript content
  • Null model: circularly shuffled event timestamps within-session, or a fitted non-cognitive renewal-process baseline for event inter-arrival times
  • R₁, operationalized here as |Predicted_Breakpoints − Observed_Event_Clusters|

Prediction (Preregistered)

Visibility-change and idle-return events will cluster within a defined proximity window of chapter/section boundaries at a rate significantly exceeding the null-model rate, aggregated across reader sessions.

Falsification Criterion

If event-clustering at structural boundaries does not exceed the null-model rate, or clustering is statistically indistinguishable from uniform regardless of content structure, H-011 is falsified.

Experimental Design

Instrument the Page Visibility API and an idle-timeout listener across manuscript pages. Log event timestamps against scroll-depth reading position. Compare observed event-clustering at chapter/section boundaries against the null model.

Control Conditions
  • Control A: Randomized/shuffled breakpoint predictions (should show no relationship to observed clustering)
  • Control B: Non-manuscript pages without chapter structure (should show no boundary-clustering effect)
  • Control C: Session-length-matched synthetic renewal-process baseline for event timing

Status

Proposed

Evidence

Not yet operationalized (awaiting Page Visibility/idle-tracking instrumentation)

Future Hypotheses

Coming as the research program advances

As LEARN progresses through experimental phases, new hypotheses will be proposed, preregistered, and tested:

  • H-006: Imagination improves prediction (Phase 2 Counterfactual Testing)
  • H-007: Multi-timescale learning outperforms single-timescale
  • H-008: Resource constraints make residual reduction harder (Physics of Knowledge)
  • H-009: Distributed cognition (multiple agents) solves problems faster
  • H-010: Independent validation confirms or falsifies autonomous claims

What We Are NOT Claiming

These hypotheses are not proven. They are proposals awaiting experimental evidence.

LEARN is not AGI. It is a research instrument for studying intelligence as a process.

Falsification is expected. Many hypotheses will be disproven. That is data.

External validation is required. A system cannot prove its own claims.