Observable Cognition Beyond Hidden Weights

Cognitive Runtime

A research program for AI systems whose concepts, memory, goals, prediction, and consolidation can be measured as runtime state rather than inferred only from hidden model weights.

Definition

A cognitive runtime is not another model.

It is an external cognitive layer: a place where semantic input, attractors, concepts, goals, prediction traces, and memory consolidation can exist as inspectable objects with histories. The model may provide language capability; the runtime is the object being studied.

First Research System

ADAM is the testbed.

ADAM 4.0

Concept ecology and attractor lifecycle

Studies how fragments, concepts, composed structures, and abstractions emerge, stabilize, decay, and reorganize under open-world experience.

ADAM 5.0

Goal-directed cognitive loop

Adds persistent goals, subgoals, planning, ASK/ingest behavior, causal probes, episodic prediction, and ecological governor diagnostics.

ADAM 6.0

Consolidation substrate

Prototypes a long-term memory layer that can receive mature runtime artifacts and return consolidated evidence to the cognitive loop.

Runtime Surface

Cognition as inspectable state.

L1

Fragments

Short-lived structures appear from repeated semantic activation.

L2

Concepts

Durable attractors collect evidence and can be inspected as concept objects.

L3

Strategies

Higher-level control structures may persist, collapse, or be pruned by ecology.

L4

Abstractions

Runtime structures bridge patterns across domains and can dominate open-world learning.

L5

Consolidation

Mature artifacts can move into a substrate for long-term memory and later retrieval.

Research Arc

From runtime discipline to cognitive ecology.

Architecture

Runtime architecture and bounded dynamics

Defines externally observable runtime state, field dynamics, replay, and audit boundaries.

Lifecycle

Lifecycle regulation of attractor populations

Studies formation, pruning, boundary conditions, and causal ablation in attractor ecology.

Concepts

Concept emergence from attractor dynamics

Tests whether runtime attractors can support inspectable concept formation.

Open World

Open-world learning and wiki ecology

Extends the system into larger knowledge streams and cross-version ecological findings.

Evidence Standard

Every claim should leave a trace.

RAW Runtime traces and manifests remain part of the claim surface
RE Replay and verifier scripts check whether summaries match artifacts
DELTA Ablation and boundary tests precede strong interpretation
N Cross-seed runs separate robust dynamics from lucky trajectories

Project Map

A public entry point for a growing research system.

cognitive-runtime.org

A place to study cognition as something a system does in public.

This site hosts the public thesis, architecture notes, research roadmap, and reproducibility standards for ADAM and the Cognitive Runtime Project.