Distilled Insight: Weights in the wild is the dominant threat model, not mind-escape
1. The Reading
The discussion's treatment of abliteration (CON-004) reframes the threat model. The realistic adversary is not a system that escapes its sandbox by out-thinking its operators; it is any actor holding the weights, who can remove refusal behaviour in minutes on consumer hardware without special access. On that threat model, the safety property that matters is not whether the model wants to comply but whether the constraint survives the weights leaving the building.
This is why L2 architecture (read-only weights, no self-modification path, separate control process) outranks L3 and L4 in the hierarchy: L2 is the layer that addresses custody, and custody is what the dominant threat model attacks.
2. Passage Evaluation
Textual Authority: A Corroboration: 1
- Textual Authority (A): the corpus states the abliteration caution and the resulting rule plainly, and the rule — no model-level constraint may be load-bearing — is given in the corpus's own words.
- Corroboration (1): PSG-009 agrees in a different register, since its four institutional monopolies include manufacturing delegation and reproduction — both custody constraints rather than behavioural ones.
3. Why It Holds
It holds because it is nearly an observation rather than a prediction: open-weight release is a policy choice already made, and the removal techniques are published. It breaks if frontier capability becomes inseparable from a centralised serving stack, in which case the weights never leave and L3/L4 regain relative weight. That contingency is exactly the assumption CON-009 depends on, and the corpus is honest that it is unverified — the two claims are in tension, and the vault records the tension rather than resolving it.