Holding the Thread: An Open Frame for Checking an Agent's Work Against the Person's Own Words

A person talks to a coding agent freely. The agent decomposes what it heard into requirements, and the person examines and approves that first transformation, which fixes them as the work's first frame, frame 0; the rules that govern the work are curated and approved by the person in the same way. This working paper describes frame-loop, a Claude Code plugin in early alpha, in which code checks every change against the rules and a reader from another model family (Muse Spark 1.3) judges each flagged change against the frame under Gadfly's counterfactual trials: a fit counts only when removing the requirement removes it and reversing the requirement turns it. That shows a verdict rests on the requirement, not that it is right. It is a design paper: it describes what is built and imagines the rest, and nothing here is yet measured. Its limits name the label that stands between the change and the reader, the model-written inversions, the cost of a reading, the reach of the authority guard, and the single harness tried. An appendix reads the paper's own citations with Gadfly, including a planted false claim as a control; its records are available from the author on request. The claim ledger (CLAIMS.md) ties each claim about the software to files and lines at commit 25f8a33 of claude-frame-loop.