Lane Files Can Stop AI Models From Inventing False README Claims

AI models drafting how-to documentation frequently fabricate unverifiable details such as pip extras, Python version floors, and sample outputs that were never actually executed. A proposed workflow addresses this by generating a machine-readable JSON inventory of public function signatures via AST parsing, which a model can use to draft parameter tables and procedure skeletons. A 'lane file' then restricts the model to only drafting content provable from that inventory, while flagging install instructions, support windows, and example transcripts as human-only responsibilities. The inventory file is hashed in Git so that future README changes can be traced back to a specific symbol set. The core principle is that any undocumented claim remains 'unsigned' until a human author verifies and attaches it to the inventory hash.