Authority separation
Reasoning output is not direct permission to perform consequential external effects.
RESEARCH / GOVERNED AUTONOMOUS SYSTEMS
Capability is not
authority.
I am building Aletheion/Axiom around a hard separation: a system may be able to reason about or propose an action without being authorized to execute it. The research question is how persistent agents can act, verify, learn, recover, and compound capability without collapsing evidence, authority, and belief into one optimistic success signal.
Aletheion proposes. Axiom governs effects.
The architecture separates cognition, authorization, execution, observation, verification, consequence measurement, and learning. That does not prove safety or general intelligence. It creates explicit seams where claims can be tested and failures can remain visible.
Evidence axes
Execution, independent verification, real-world consequence, and artifact provenance remain distinct questions.
Async reality
Request creation, transport acknowledgement, and later provider/world observation are different events with different evidentiary meaning.
Belief admission
Verified outcomes can revise learned state; insufficient evidence can abstain rather than being converted into contradiction.
Reality remeasurement
Task completion does not close a discrepancy. The detector that found the gap can be run again after the intervention.
Governed learning
Experience-derived rules can be challenged, narrowed, strengthened, superseded, or retired without gaining execution authority.
Claims should come with ways to kill them.
The public research standard is claim → evidence → positive test → falsifier → stronger claim not implied. Current claims are first-party implementation evidence unless explicitly labeled otherwise.
Falsifier: demonstrate a canonical cognition path that directly performs a protected external effect without Axiom authorization.
Falsifier: show that canonical outcome state necessarily collapses execution, verification, consequence, and provenance into one success value.
Falsifier: show that an asynchronous acknowledgement alone is treated as observed world completion on the correlated-effect path.
Falsifier: an evidence-empty cited-belief outcome silently reinforces or disputes the belief instead of abstaining.
Claim boundary. This page does not claim peer review, independent replication, AGI, consciousness, universal autonomy, or production safety. Those require evidence not established by the current public record.
Current experimental edge.
The next standard is not a larger feature list. It is treatment/control evidence showing that verified experience changes later behavior across a fresh process, survives revocation tests, and transfers only where the learned abstraction actually applies.
Embodied baseline
Measure perception → planning → coordination → action → observation → correction under a sealed rubric.
Disruption + recovery
Introduce consequential world-state change and measure stale-belief detection, replanning, coordination, and mission continuation.
Fresh-process retention
Cold-start and test whether the causal lesson measurably reduces recovery cost on a matched challenge.
Transfer
Change embodiment and test the abstract learned principle rather than Minecraft-specific coordinates or action sequences.