CIM ADR System

Context

The LEARN project is an experimental cognitive infrastructure. CIM is a methodological research program investigating intelligence measurement. These require different governance models.

Decision

CIM is established as an independent research program that can leverage LEARN as an experimental instrument, but maintains its own hypotheses, protocols, and publication strategy.

Rationale

This separation allows each initiative to evolve independently while enabling collaboration. CIM can define experiments that LEARN executes without conflating the two projects.

Alternatives Considered

• • Merge CIM entirely into LEARN (rejected: conflates project identity and governance)
• • Keep CIM completely separate with no connection to LEARN (rejected: loses experimental access)

Context

Traditional intelligence measurement collapses behavior to a single score. CIM investigates whether the full trajectory contains valid information.

Decision

Event sequences (O→R→P→H→A→F→E→V→L→T) are the primary unit of analysis, not task outcomes.

Rationale

The trajectory preserves temporal, adaptive, and mechanism-level information that static scores discard. Scientific validity requires preservation of complete data.

Alternatives Considered

• • Treat trajectories as secondary to outcomes (rejected: loses the information that justifies CIM)
• • Store only summary statistics (rejected: prevents falsification checks)

Context

Speculative language can blur with empirical claims, undermining credibility and scientific integrity.

Decision

All interface elements must visually and linguistically distinguish: HYPOTHESIS, PROPOSED EXPERIMENT, OBSERVATION, RESULT, INTERPRETATION, SPECULATION.

Rationale

Clear distinction protects against confirmation bias, unwarranted claims, and scientific misconduct. It makes the research auditable.

Alternatives Considered

• • Use natural language labeling (rejected: ambiguous)
• • No labeling (rejected: violates scientific standards)

Context

Continuous measurement generates large datasets. Provenance integrity is essential for reproducibility and falsification.

Decision

Every observation is timestamped and tagged with: source, experiment_id, environment_id, system_id, measurement_method, confidence, observability_classification.

Rationale

Future researchers must be able to reconstruct exactly what was measured, how, and with what confidence. This enables meta-analysis and replication.

Alternatives Considered

• • Store aggregated statistics only (rejected: prevents verification)
• • Store raw data without metadata (rejected: unusable for reproduction)

Context

Intelligence measurement research spans months or years. Data integrity requires immutability.

Decision

All research records are append-only. Historical observations are never deleted or overwritten. Corrections are new entries with explicit relationship to corrected data.

Rationale

This prevents loss of data due to accidental deletion, system errors, or motivated reasoning. It creates an auditable history.

Alternatives Considered

• • Allow record modification (rejected: enables data manipulation)
• • Delete failed experiments (rejected: hides methodology)

Context

The core hypothesis compares static outcomes to dynamic process metrics. Measurement incompatibility would invalidate the comparison.

Decision

Experiments must measure both final performance score AND the complete trajectory from which process metrics are derived.

Rationale

Only by measuring both can the research answer the central question: do process metrics explain additional variance?

Alternatives Considered

• • Measure only trajectory (rejected: cannot compare to static)
• • Use separate datasets for each (rejected: confounds comparison)

Context

Phase IV and V include physical experiments with Raspberry Pi sensors and robotic systems. Data structures must support heterogeneous sensor streams.

Decision

Event records use flexible schema with explicit observability classifications, supporting software, hardware, and hybrid measurements.

Rationale

This avoids architectural redesign when adding physical experiments. It enables substrate comparison as planned.

Alternatives Considered

• • Build only for software systems (rejected: prevents robot/sensor integration)
• • Different schemas for different systems (rejected: prevents cross-substrate comparison)

Context

Scientific integrity is non-negotiable. Fabrication is the ultimate corruption.

Decision

The system explicitly labels experiments as PROPOSED (not yet run), PILOT (preliminary), IN_PROGRESS, ANALYSIS, PREPRINT, REPLICATED, SUPPORTED, WEAKENED, or FALSIFIED. Never present unexecuted experiments as confirmed.

Rationale

Integrity attracts collaborators and citations. Fabrication destroys credibility permanently.

Alternatives Considered

• • Vague status labels (rejected: enables self-deception)
• • No status tracking (rejected: invites fabrication)