Research

Understanding the foundations of measurable machine cognition.

The UARI Framework

Ψ

Ψ is Stance, Not Computed

The central insight of UARI is that cognitive stance (Ψ) is an observed position, not a computed metric. We don't calculate understanding—we observe indicators that suggest a system has adopted an understanding stance toward a problem.

This philosophical distinction matters because it shapes how we design measurement. Instead of trying to "score" cognition directly, we instrument systems to emit telemetry that reveals stance through behavior, outputs, and state transitions.

U

Understanding

Conceptual grasp of problem space and requirements. Observed through appropriate abstractions and generalizations.

E

Execution

Quality of task completion and output generation. Measured by correctness, completeness, and efficiency.

P

Planning

Forward-looking coherence and strategy. Observed through goal decomposition and step sequencing.

L

Learning

Adaptation rate and improvement over iterations. Tracked through performance delta across similar tasks.

M

Memory

Context retention and appropriate recall. Measured by reference accuracy and continuity maintenance.

F

Focus

Attention allocation and relevance filtering. Observed through response pertinence and distraction resistance.

Artifact Bundles

Every cognitive run produces a cryptographically-verified artifact bundle. These bundles enable full reproducibility and independent verification of results.

Bundle Structure

artifacts/
└── uari/
    └── <run_id>/
        └── <checkpoint_sha>/
            ├── manifest.json      # Bundle metadata
            ├── events.jsonl       # Event stream
            ├── metrics.json       # UARI metrics snapshot
            ├── state.bin          # Serialized state (optional)
            └── signature.sig      # Cryptographic signature

Immutable

Once written, bundles cannot be modified. SHA-based addressing ensures integrity.

Self-Contained

Everything needed to verify and replay a checkpoint is included in the bundle.

Signed

Cryptographic signatures enable provenance verification and tamper detection.

EventEnvelope & events.jsonl

The EventEnvelope is the atomic unit of telemetry. Every observable action, state change, or metric emission is wrapped in this standardized format and appended to the run's event stream.

EventEnvelope Schema

interface EventEnvelope {
  // Identity
  id: string;              // Unique event ID (UUID)
  run_id: string;          // Parent run identifier
  
  // Timing
  timestamp: string;       // ISO 8601 timestamp
  sequence: number;        // Monotonic sequence number
  
  // Classification
  event_type: string;      // checkpoint | metric | log | error | phase_change
  level: string;           // debug | info | warn | error
  
  // Payload
  message: string;         // Human-readable description
  data?: object;           // Structured event data
  
  // Context
  phase?: number;          // Current phase (if applicable)
  checkpoint_sha?: string; // Associated checkpoint (if applicable)
}

Example events.jsonl

{"id":"evt-001","run_id":"run-2024-001","timestamp":"2024-12-30T10:00:00Z","sequence":1,"event_type":"phase_change","level":"info","message":"Run started","data":{"phase":0}}
{"id":"evt-002","run_id":"run-2024-001","timestamp":"2024-12-30T10:00:01Z","sequence":2,"event_type":"metric","level":"info","message":"Initial metrics","data":{"understanding":0.0,"execution":0.0}}
{"id":"evt-003","run_id":"run-2024-001","timestamp":"2024-12-30T10:00:15Z","sequence":3,"event_type":"checkpoint","level":"info","message":"Checkpoint saved","checkpoint_sha":"a1b2c3d4"}

Research Principles

Observation Over Computation

We observe cognitive indicators rather than trying to compute cognition directly. Metrics are proxies for stance, not measures of understanding itself.

Reproducibility First

Every claim must be verifiable. Artifact bundles enable independent reproduction of any result.

Transparency Through Telemetry

Black-box systems become glass-box systems through comprehensive instrumentation. The goal is visibility, not just output quality.

Progressive Refinement

Our understanding evolves. The framework is designed to accommodate new insights without breaking backward compatibility.