Explanation

Architecture overviews, theoretical background on COCO-style evaluation, and design rationale. Audience: anyone who wants to understand how vernier thinks.

• Three paradigms: instance, panoptic, semantic — picking the submodule whose input shape matches your model's output.
• TIDE: what it answers, and what it doesn't — the six error bins, the assumptions behind them, and where they break.
• LRP / oLRP: what it answers, and what it doesn't — the operating-point decomposition, how it differs from AP and TIDE, per-class tau semantics.
• Calibration: what it answers, and what it doesn't — ECE / MCE / reliability, DETR-era defaults, orthogonality to AP and oLRP.
• Why per_image does not ship an AP column — why per-image AP is structurally ill-defined under the COCO fold.