State before story
Claims must connect to runtime objects such as attractors, concepts, goals, traces, or substrate records.
Methods
Evidence before interpretation.
The project treats every cognitive claim as a claim about observable runtime state. A concept, goal, abstraction, prediction trace, or consolidation event should leave an artifact that can be inspected.
Scope before headline
The project separates demonstrated behavior, active development, and speculative interpretation.
Negative results matter
Seed variance, basin states, blocked goals, and failed transitions are treated as evidence, not noise to hide.
Evidence Workflow
A runtime claim should survive replay, audit, and boundary checks.
ADAM experiments are developed through a practical loop: formal run, raw artifact preservation, verifier, audit, claim discipline, and follow-up boundary tests. This keeps public claims tied to data rather than visual demos alone.
Formal Run
Configurations and seeds are recorded
Runs preserve enough metadata to distinguish a lucky trajectory from a repeatable dynamical pattern.
Verifier
Numbers should be recomputed from raw artifacts
Summaries are checked against records instead of accepted because they look plausible.
Audit
Claims are separated from interpretation
Audits record what was shown, what remains open, and which labels should not be overclaimed.
Boundary
Hard seeds and sparse ecologies are useful
Failure modes reveal whether the architecture needs a new mechanism, a governor, or simply more evidence.