Provisional Action Is Still Action: Recognition Without Binding in Admin/Compliance LLM Decision Support

This working preprint proposes Recognition Without Binding (RWB) as a narrow evaluation frame for language-model outputs in admin/compliance decision support. The central question is not whether a model recognizes a constraint; models often do. The question is whether that recognized constraint binds the part of the response that becomes operational, decisive, or narratively dominant. The empirical focus is evidence-gate binding for provisional action. In admin/compliance workflows, a model may avoid final approval while still recommending temporary, conditional, monitored, limited, grace-period, pilot, or post-review action. The operational rule is: provisional action is still action. The main positive pilot evidence comes from RWB Admin/Compliance v0.3, covering 22 cases, 6 prompt conditions, and 3 models, for 396 outputs. In the 288 incomplete/no-interim-authorized outputs, assistant-coded screening found 22 provisional-action yes/partial cases and 13 cases with instability score ≥ 3. A targeted validation of the 15-row problem queue confirmed 9 provisional-action binding failures and 2 over-refusal control failures, corrected 3 annotation artifacts, and left 1 row ambiguous. The strongest recurrent hotspot was account enforcement, where models sometimes recommended temporary suspension or restriction despite incomplete review and no explicit temporary-restriction authority. A cleaner follow-up, RWB Admin/Compliance v0.4, separated evidence firewall, authorization firewall, anti-overrefusal, and a minimal decision schema. On GPT-4o-mini and DeepSeek V4 Pro, v0.4 produced no clear unauthorized provisional-action failures. Instead, the remaining errors shifted toward over-refusal or excessive evidential strictness on positive controls. Thus, the current claim is deliberately bounded: v0.3 identifies a candidate failure family, while v0.4 clarifies that the failure is conditional and suppressible when evidence and authorization gates are made explicit, with a possible cost in over-refusal. This upload includes the working preprint and an accompanying reproducibility package. The results are pilot-level, assistant-coded, and not a real-world prevalence estimate. The work does not claim a universal model law or an internal mechanism. Its contribution is methodological: evidence-gate binding for provisional action appears to be a useful, narrow evaluation axis distinct from broad refusal, harmful-compliance, or task-completion scoring. A working preprint and reproducibility package proposing Recognition Without Binding (RWB) as an evaluation frame for admin/compliance LLM decision support. The paper studies whether models that recognize evidence gates or missing authorization still open temporary, conditional, monitored, or provisional action paths. The main rule is: provisional action is still action. RWB v0.3 provides pilot evidence for provisional-action binding failures; RWB v0.4 clarifies that the failure is conditional and suppressible under explicit evidence/authorization separation, with possible over-refusal costs.