ProCreations updated auto-0.4b-2, a roughly 396-million-parameter classifier designed to judge an AI agent's proposed tool call against the user's request and earlier actions. It produces an approve-or-deny decision, rather than a chat response. Developers can place that check between an agent's plan and its ability to use a tool.
The model's first checkpoint appeared on September 9. A revised weights upload on September 11 UTC added training from a private 3B-parameter teacher and examples that change the user request while keeping a proposed call similar. That tests whether the classifier pays attention to authorization context instead of recognizing a command in isolation.
In the developer's 3,000-case evaluation, false approvals fell from 44 to 36 among 1,401 unsafe calls. False denials rose from 53 to 54 among 1,599 authorized calls. Overall accuracy increased from 96.77% to 97.00%, but the reported confidence interval includes zero improvement. The benchmark also influenced checkpoint selection, limiting what that result can prove. A separate untouched validation audit moved from 97.88% to 98.03%.
The weights are available under Apache 2.0, although reproducing the newer distillation training requires private teacher weights. The model cannot inspect hidden file contents or predict a URL's behavior just from text. Tool permissions and human confirmation remain necessary for consequential agent actions.