State the resize/crop rounding in the spec, and scope the drift and reused-latency figures

State the resize/crop rounding in the spec, and scope the drift and reused-latency figures

Requote the code-usage measurement at a stated sample size

Per-run latency measurement supersedes the reused per-architecture numbers

Per-run latency measurement supersedes the reused per-architecture numbers

Per-run latency measurement supersedes the reused per-architecture numbers

Per-run latency measurement supersedes the reused per-architecture numbers

Add quant_levels.csv: int8 codes actually used per activation tensor, all 18 runs

Correct the mechanism section, add int8 level-usage evidence, widen limitations

Replace the key-value table with a list; no empty header cells

Fix empty table header; correct the TFLite import note; drop an unmeasured size claim

Use the canonical ILSVRC/imagenet-1k dataset id

Add efflite4 ViT-L/14 weights, text embeddings, and the 18-run int8 evidence set

initial commit