Compact runtime state
The field is the measurable surface where semantic input, memory, and control pressure interact.
Architecture
How ADAM is built.
ADAM is organized around an inspectable cognitive field, runtime attractors, safety arbitration, and replayable state transitions. This page presents the system at the level needed to understand the public research program.
Durable structures
Attractors bias future trajectories and provide units for inspection and comparison.
Safety arbitration
Safety behavior is treated as a runtime signal that can be measured separately from generation.
Public Architecture Standard
The runtime is treated as the object of study.
ADAM separates language generation from the cognitive state that conditions it. The public architecture notes will focus on externally observable mechanisms: how input enters the field, how attractors shape future trajectories, how safety pressure is represented, and how state can be revisited for audit.
Boundary
Language model as substrate, runtime as experiment
The model provides language capability; the runtime supplies measurable cognitive dynamics.
Inspection
State must be visible outside hidden weights
Attractors, field values, safety signals, and replay deltas are treated as auditable objects.