Distilled Insight: Perfect internal binding and zero external assurance are the same property
1. The Reading
Foster's thought-sensitive, self-enforcing constraint is the corpus's limiting case: a binding with no device to inspect, no procedure to revoke, and no external evidence that it is working or that it is still the same constraint it was when installed. The discussion's point is that this is not a trade-off between strength and transparency — it is a single property described from two sides.
The practical consequence is that interpretability (L4) cannot be the end state of control. A constraint that lives entirely in the activations is a constraint that no auditor can inspect without the access the constraint was meant to remove, so the better it works, the less anyone can say about it.
2. Passage Evaluation
Textual Authority: C Corroboration: 2
- Textual Authority (C): the work was not retrieved, and the discussion's formulation is a reading of a fictional device rather than a claim the text makes. The source admits a narrower reading: the device is a plot mechanism for a specific character's predicament.
- Corroboration (2): PSG-006 agrees in a related register — a constraint reliable enough to be invisible is also reliable enough to have removed the agent from its own conduct.
3. Why It Holds
It holds as a structural point about evidence: assurance requires an observation point outside the thing assured, and internal binding removes exactly that. It breaks if a technique is found that reads a model's internal constraints from outside without granting the access the constraints guard — which is what interpretability research is attempting, and the corpus does not claim it has failed, only that it cannot be the primary binding.