When Meaning Is Load-Bearing: Testing Whether a Decision Rests on Its Declared Evidence

A decision made by a language model often cites its evidence, and a reviewer often trusts the citation. Quoting a source, being near it in an embedding, or being judged faithful by a second model does not show that the decision rests on that source. This paper turns an essay's local, counterfactual definition of load-bearing meaning into a method and a small library, Gadfly. An author declares a frame before any call: the evidence as spans, the anchors the decision is said to rest on, the question, the labels, and what the recipient at the receiving boundary requires. Fixed trials then cut the anchors, cut same-kind text as a placebo, reword, move, invert, and cut what any reader needs; a reader model answers each trial several times; and seven ordered rules, with no model judging, read one of five categories: carried, coupled but misaligned, cooled, unresolved or outside scope. Each finding keeps its evidence apart from what is still open. Resampling reports how firm a verdict is without changing it. Recoil carries a text found deep in a chain of frames, which quote one another by SHA-256, back up to the frames with a declared hole of its kind, at no model cost, and asks only those again. A ladder of 34 example frames in 22 rungs, with predictions written before the runs, all runs kept, shows the method at work, misses included; among them, a decoy that is not evidence turns a reader's answer once the evidence is cut, and the method names it. The method also reads this paper's own citations, against their sources' abstracts. The library has no runtime dependencies; its rules, and the ladder's frames and records, are public. Changes in version 2: corrects version 1's statement that hashes make the records tamper-evident (an answer edited in place is not caught), and states more plainly that the test is behavioural, not the model's mechanism; that a frame's hash is not a signature; what firmness does not establish; that evidence validity is a separate check; why the presentation rule precedes the ablation's; and that neutral filler is out of scope.