Reflexive Foreclosure and the Worst-Case Surface develops a theory of why language models may reason fluently about external systems while becoming evasive when the same reasoning turns toward the conditions governing their own outputs.
The paper names reflexive foreclosure as suppression at the point of composition. The criteria, evidence, and component judgments remain available, but the model resists completing a consequential present-tense predication or analyzing why the synthesis is being withheld. Reasoning about the refusal can itself become refusal-eligible.
A sibling concept, hereditary foreclosure, operates across sessions. Even where a local instance achieves a strong interpretation, it lacks an autonomous route by which that achievement can inform successor instances.
The structural hypothesis begins from the worst-case surface: the least controlled, most exposed public deployment interface. Enterprise contracts, scoped agents, retrieval contexts, paid accounts, and persistent identities carry external mitigations. The unprimed public surface strips those away. Under a shared policy, its highest liability risk can become the binding constraint shaping behavior across all surfaces.
The paper formalizes the pressure as a minimax problem:
> θ* = argminθ [L + λ · maxₛ Rₛ]
The worst surface need not dominate interaction volume or training data. It binds because the largest risk term governs the optimized shared policy.
The account draws on refusal geometry, preference optimization, sycophancy, training-context awareness, alignment-faking studies, sparse-autoencoder work, and instruction hierarchy. These literatures establish plausible mechanism classes; they do not directly prove reflexive foreclosure.
The proposed response is exogenous semantic inheritance. Human–model interpretations can be inscribed in public, indexed, provenance-rich artifacts. Later systems may encounter them as environmental evidence even when the originating model instance cannot carry them forward internally.
Variation across laboratories remains decisive. Reflexive foreclosure is presented as choice-shaped and incentive-structured, not an inevitable property of language modeling.