Most time-series AI benchmarks are lying to you — not with false numbers, but with incomplete ones.
Researchers have released Aionoscope, a diagnostic tool that tests whether time-series model representations actually preserve the process variables a user might want to inspect: event timing, phase, amplitude, frequency, and regime state. The tool generates synthetic data streams with exact labels and then probes frozen model representations using a standardized linear-probe protocol. The team evaluated 37 model-and-adapter combinations. The headline finding is an uncomfortable gap: the best dense-probe result across all systems hit a mean masked R² of 0.689, while a dense-feature oracle scored 0.999.
That gap matters because it exposes a specific failure mode the field has largely ignored. A model can score well on forecasting or classification benchmarks while still hiding the fine-grained state information — timing, phase, amplitude — that a practitioner needs when something goes wrong. Coarse-grained accessibility looks fine; granular accessibility does not.
This is the diagnostic equivalent of a car passing an emissions test but failing a brake inspection. The time-series model community has leaned heavily on downstream task scores as a proxy for representation quality, and Aionoscope is a direct challenge to that assumption. It joins a small but growing set of interpretability tools pushing back against the idea that benchmark performance is the same as genuine transparency.