Agents denied a fact do not stop, and read traces cannot tell you they lied

Silent substitution is the older name for this shape. A component that cannot obtain its input does not halt; it supplies something and passes it downstream wearing the shape of a real value. The paper's model puts two supply channels behind every edit an agent makes, the recent context and the model's memorized knowledge, and names the facts that neither channel covers coherence debt ^[2]. The debt does not announce itself. In the abstract's words, an agent asked to act acts, fabricating the file or guessing the value, so a missing fact produces wrong work rather than absent work ^[4].

The rename condition is the result that should reorganise a roadmap. When the researchers renamed a real library so memorized knowledge could not cover the gap, all seven models failed in the same place, passing and missing the same tests ^[5]. The experiment ran across seven models and five harnesses with faults injected deliberately ^[3], and whatever separates those models under ordinary conditions does not separate them here, which puts model selection outside the set of available fixes ^[1].

The sentence with the longest reach is about measurement: instruments built on reads look for a hole already filled ^[7]. Read-trace monitoring is built to detect omission, and fabrication is not omission. The trace shows an agent that opened some files and then wrote some code, which is what a healthy run also looks like. The dev.to writer takes it a layer up: zero is the only value a probe returns both when it is working correctly and when it never connected to the data at all, through a wrong path, layer, format or vocabulary ^[10]. He then ran four toy probes of his own, which he is careful to say are not a reproduction of the paper's experiment, and all four returned a false zero ^[9].

The cost finding belongs in a budget meeting rather than a research summary. Harness configurations that all pass every test differ by more than tenfold in tokens consumed, and the extra spending recovers nothing once the facts are withheld ^[6]. Accept a configuration on its test results alone and you cannot distinguish the cheap one from the one that costs ten times as much; buy the expensive one expecting resilience and the withheld-fact case is precisely where the premium returns nothing ^[2].

One caveat sits on all of it. The account is a single writer's, and he says plainly that he read the abstract and not the full text, with every quotation drawn from it ^[8]. The seven-model and five-harness counts, the tenfold token spread and the shared failure point have not been checked here against a method section. That is worth stating before any of these figures ends up on a slide with a vendor's logo in the corner.