What problem does it solve? When a quantized model loses accuracy on the D-Robotics X5 platform, it is hard to tell whether the drop comes from preprocessing, calibration, fake quantization, fixed-point conversion, compilation, or board-side runtime. This Skill systematically locates the first stage where accuracy diverges, so you fix the actual root cause instead of tuning parameters blindly. ## Core Features & Use Cases - Stage-by-stage comparison: Builds a stage matrix over fixed samples covering float, calibration/fake-quant, quantized, compile, and runtime outputs, computing task metrics and numerical differences such as cosine similarity and max absolute error. - Single-variable experiment design: Produces one prioritized experiment and one fallback (e.g., replacing the calibration set, fixing preprocessing, adjusting one qconfig item, or extending QAT) with expected outcomes, stop conditions, and rollback plans. - Structured handoff: Routes unresolved issues to the right workflow — calibration data preparation, PTQ config authoring, or QAT training — and stays read-only without retraining or recompiling. - Use Case: You have floating-point, PTQ, and board-side metrics for the same evaluation set and the board-side accuracy dropped sharply. Use this Skill to identify that the first divergence appears at the fixed-point conversion stage, linked to calibration saturation, and receive a targeted single-variable fix. ## Quick Start Ask the agent to diagnose where the accuracy drop first occurs using your floating-point, calibration, quantized, and board-side metrics on a fixed set of input samples.
