Every verifier has rules about what may decide what, and they are almost always prose in a README. Here two of them are conditions on a connection: a value that came from floating point has no wire into a port that decides, and a hypothesis stamp may not meet a different one. Break either and nothing scores badly — it does not build, and what you get back is the sentence the engine raised.
This is a real cell of TERRA's certified sigma-band, drawn as the graph that decides it. The band hands three exact intervals to the radii polynomial; the polynomial is the only thing here that may conclude anything, and its three verdict ports are drawn whether or not they fire. That is the point of drawing them: a one-sided method that can certify and refuse but never refute has a dark port, and the picture says so instead of the caption.
Drag from the float screen's output to any input on the radii polynomial. The editor implements no validity rule of its own: it calls the same wire() the pipeline calls, inside a try, and shows you whatever came back. It cannot disagree with the engine, because it has no opinion.
With scripting off this page still says everything it claims; the refusal below is captured at build time from the exception the engine raises, not written by hand:
THE FLOAT FIREBREAK: screen.shortlist carries floats and radii.Y0 decides. A fast screen may prune. It may never reach a verdict.
The graph is not a diagram of the arithmetic — it is the arithmetic's wiring, and it was checked against work this repository already holds. replay.mjs pushes every certified cell of the sigma-band back through the radii node, so the certificate re-derives itself instead of being quoted.
Ported from frontier-apps and pinned by sha256 in corpus/frontier-port.json. The same two rules exist a second time, in Python, as a grader where a model answers with a wiring and building the graph is the grading — instruments/wiring/. A cross-check requires both engines to refuse the same wiring in the same words.