Grounded Self-Report in an Integrated Cognitive Architecture: Accurate Description of Internal States as a Measurable Precursor to Machine Self-Modelling

We report a behavioural result from XK82, an integrated cognitive architecturedeveloped as a long-running independent research programme. The system ispresented here as a black box: architectural details are withheld, and allclaims are stated in terms of externally verifiable behaviour. The system was taught to produce language from scratch with no pretrainedweights, then trained to describe its own runtime internal states using ataught vocabulary applied to measured quantities. On states never encounteredduring training, the system produced correct descriptions of its own internalcondition in 91.7% of trials (22/24), against a majority-class baseline of41.7% and a shuffled-label control of 20.8%. A second, independently scoredinternal variable reached 66.7%. A fine-grained sweep shows the system placedits own decision boundaries within 0.01 of the empirical thresholds of itsstate distribution, despite never being given those thresholds. We distinguish this result sharply from claims of machine consciousness. Whatis demonstrated is grounded self-report: statements about internal state thatpossess truth conditions and can be scored against ground truth. We furtherreport an auditing procedure which revealed that several nominally activecomponents of the architecture were contributing nothing measurable to systembehaviour, and describe the repair and re-measurement that followed.