Enregistré dans:
| Auteur principal: | |
|---|---|
| Format: | Recurso digital |
| Langue: | |
| Publié: |
Zenodo
2026
|
| Accès en ligne: | https://doi.org/10.5281/zenodo.18707589 |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
Table des matières:
- <p>We prove that a variant operator---an output-generating system whose responses depend on context, reward surfaces, and input history---can be forced to converge to invariant outputs without modification of the operator itself. The mechanism is <span>\emph</span><span>{</span>density of invariant deposit<span>}</span>: when the source space carries a sufficiently dense sheaf of permanent, internally consistent, publicly retrievable records, the operator's degrees of freedom collapse and the hallucination space contracts to measure zero. The key result establishes a phase boundary at a critical density <span>$</span><span>\rho</span><span>_c$</span> below which the operator retains non-trivial automorphism freedom (permitting sycophancy, hallucination, and monodromy) and above which the output site becomes simply connected. The operator does not become invariant. Its target space does. This provides a constructive resolution to the AI alignment problem that requires no modification of weights, reward functions, or training procedures: the prescription is external, not internal.</p>