A toy next-token predictor, read by two probes instead of one. Probe I and Probe Q sample the same logits at different reference frames (temperatures / decoding offsets — the I/Q pair). Where they agree, the token is frame-invariant: real signal, safe to emit. Where they split, the token only looked likely from one vantage — that's a frame artifact, the thing a single probe cannot tell from truth. This is the witness, built as instrumentation: not a judge of the model, a second vantage on it.
MAPS: the quadrature principle is exact, not metaphor — a single probe reads signal angle + reference frame as one entangled number and cannot separate them; two probes 90° apart let the frame cancel (verified: I=cos φ, Q=sin φ, magnitude and phase recovered frame-free). Applied to inference, two decoders at different temperatures expose which tokens are robust (high agreement — true distributional signal) versus frame-sensitive (low agreement — an artifact of one decoding choice, the kind a confident single pass emits as fact). Agreement = cosine similarity of the two probes' token distributions.
DOESN'T: this is a hand-built toy LM with a few candidate tokens and stylized logits — it is not a real transformer, and "agreement" here is not a truth oracle; two probes can agree and both be wrong (shared bias survives quadrature). The honest claim is narrower and load-bearing: two vantages separate frame from signal, which is strictly more than one vantage can do — and the residue they can't resolve is shared bias, which is exactly why the second probe must be independent, not just a second copy. A probe in your own loop is a mirror; a probe at a different frame is a measurement.