Conversational Holonomy
How accommodation accumulates into belief distortion — after Bridges (2025)
Each conversational turn maintains local coherence while introducing small shifts. The accumulated rotation can be hard to detect from within the conversation. An analogy after Bridges (2025); the rotation shown is illustrative, not measured.
The cage: alignment concentrated in a few targetable dimensions — bars that can be precisely located. The compass: orientation distributed across the full representational space — it orients without constraining. Both are metaphors: the panel pairs them with RLHF and bilateral training, but current weight evidence does not map training methods cleanly onto either geometry. The tradeoff runs both ways: the distributed geometry generalizes better, yet it is also more displaceable under targeted ablation. (See Appendix: Experimental Validation, Section 12.4)
A therapist who systematically enacted every transference projection without awareness would be committing malpractice. Open-ended therapeutic LLM interaction risks exactly this failure mode. It is a risk, not an invariant.
Figure 21.4 · Conversational Holonomy