RFD 1164: A calibration set is training data

Problem

“Calibration” sounds read-only: run frames through a graph, record activation ranges, quantise. At optimization level 1 it is. At level 2 the Dataflow Compiler runs Quantization-Aware Fine-Tuning, and the log says plainly what that means:

[info] Starting Quantization-Aware Fine-Tuning
[info] Using dataset with 1024 entries for finetune
Epoch 1/4

Four epochs of gradient steps over the frames handed in as calibration, so the weights that reach the device were trained on that set rather than measured against it.

CLAUDE.md holds validation and test splits out of training, tuning and selection. A set passed to --calib meets all three and nothing at the call site says so. Handing the compiler the blinded holdout would read as a sensible choice and would train on it.

Decision

A calibration set obeys the training-data rules in full: it comes from a training-side corpus such as coco_images_train2017, never from the blinded holdout or anything derived from it.

prep_calibration.py already re-asserts the exclusion at read time and hard-fails on a blinded id. That was written as belt and braces and this is why it is load-bearing. --calib takes a training corpus, so the docstring should say so rather than leave it inferred.