ICISELFLAB / SELF-LAB
ICI Self Lab studies functional self-awareness: whether an artificial agent can represent its own capabilities, limits, actions and errors in a way that causally changes later decisions.
01
A model can describe itself convincingly without showing that self-information matters to behaviour. For a functional self-model to be useful, it must change decisions under uncertainty, survive transfer to new conditions and outperform simpler explanations such as prompt familiarity or ordinary memory.
02
The ICI architecture separates Body, Agency, Capability and Epistemic Self and gives the system an interoception channel. Self-continuity, agency and error-retention effects are attacked with baselines, random controls and ablations that remove memory, interoception or causal links. The aim is falsifiable functional behaviour, not a claim about human phenomenal consciousness.
03
If an agent fails because it overestimates one of its own capabilities, ICI retains the mechanism rather than only the corrected answer. The surface details are then changed. The key test is whether retained self-knowledge alters the next action before the same failure repeats.
04
Self Lab treats controls, ablations, repeatability and failure cases as part of the result. Current causal-core release: ICI FINAL V3 / ICI-CFB-v3. Its single frozen confirmatory run covered 1,200 scenarios and passed all 5 preregistered technical gates; independent external replication remains separate and pending.
CONTACT
Research collaboration, controlled pilots and technical evaluation.
