We ask whether a compact, transferable, multi-resolution model can learn minute-level structure that improves
calibrated probabilistic forecasts of the next 5, 15, 30 and 60 minutes of U.S. equity and ETF returns. We build a
Multi-Resolution Patch Transformer that reads the last 390 one-minute bars together with five days of 5-minute and
twenty days of 30-minute past-only aggregates, emits Student-$t$ marginals for each of the next sixty minutes and,
in a second stage, a low-rank Gaussian copula over them. The model is trained and evaluated on \num{234} million
one-minute bars of 300 instruments (2018--2025) from a private FirstRate Data lakehouse, selected point-in-time and
audited for timestamp, session and corporate-action conventions, against ten baselines sharing the same
information set, with CRPS, interval and calibration scores, trading-day cluster inference and overlap-aware
Diebold--Mariano tests on about one million evaluation origins per year, sixty held-out instruments and an untouched
final year. Every model beats a trailing-scale random walk by 3.5--4.9\,\% of CRPS; the gain is a volatility gain
with no directional or economic content. Sequence models lead at five minutes; a HAR-type variance regression with
a kurtosis-matched Student-$t$ leads and is better calibrated from fifteen minutes on. The multi-resolution context
adds at most 0.1\,\% in sample and nothing out of sample, model size (3.5 vs 20.7 million parameters) does not
matter, regime conditioning and coherent path sampling add nothing because future minute returns are conditionally
independent given the recent history, and pooled training transfers to unseen instruments without penalty. A model
that reads only the trailing volatility and the clock recovers nine tenths of the full model's skill. We release
the code, the protocol, synthetic fixtures and these null results.