From Indicators to Discriminating Tests: Theory-Inspired Readouts of a White-Box Chess Engine

Consciousness assessments risk mistaking theory-inspired features for evidence of their intended constructs. We examine this in a conventional, fully inspectable chess engine using three limited analogues: evaluation-term dependence (integration), a shared transposition table (workspace), and an internal confidence signal (metacognition). Each was paired with a simpler rival, a contrast that could count against it, and instrument checks. The static integration test failed its baseline criterion; corrected temporal tests did not establish coupling in at least half the positions, although a weaker post-hoc test detected some departures. Sharing the table made search about 1.9 times faster with less work, consistent with ordinary parallel search, which no contrast excluded. The confidence signal did not rank move persistence above chance, but one component was nearly constant, leaving sensitivity unvalidated. These findings neither establish nor exclude consciousness. Stronger evidence requires discrimination against credible simpler explanations, valid measurement, and an explicit link to the construct.