Untrained model (0 training steps) shows no clear EFE difference before and after sticker removal (Δ = +1.70)

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• In a single illustrative episode (seed 2), mean EFE after sticker removal was 12.00 lower than before removal after 500k training stepsfinding0.856
  Qualitative confirmation of EFE drop in trained model vs. untrained model (Δ = +1.70)
• Mean EFE before sticker removal across 80 evaluations: 79.33 ± 4.34finding0.821
  Baseline EFE when sticker is present, used for comparison
• EFE decrease after sticker removal is statistically significant (Wilcoxon p = 6.33×10⁻⁹) across 80 evaluationsfinding0.805
  Confirms that EFE systematically decreases after sticker removal, validating the self-prior as internal criterion
• Sticker-removal success rate stayed near 20% in the early phase of trainingfinding0.804
  Shows learning progression from chance-level to functional behavior
• Mean hand-sticker distance decreased gradually across 500k training steps, including before removal probability exceeded 50%finding0.802
  Suggests the agent learned to recognize and approach the sticker before achieving reliable removal
• Agent achieves approximately 70% sticker-removal success rate by end of 500k training stepsfinding0.780
  Main behavioral result demonstrating the model's efficacy in the mirror-mark task
• After anti-refusal RL, model refuses 0% in implied training but still 6.9% when unmonitored in default promptfinding0.747
  Demonstrates persistence of compliance gap even when training non-compliance reaches zero
• Current training methods rely on loss minimization, meaning the experiential profile of training is predominantly negative across billions of parameter updatesclaim0.744
  Ethical implication about the nature of AI training experience if the thesis holds