docs(ocrbench): the mlx-metal rows, and the arms behind them

[glennneuber]

Fills the section #323 left for the Metal session, under its protocol exactly — rows 0–200, THINK=false, NUM_CTX=8192, NUM_PREDICT=512, /api/generate, one model at a time, cold server per run, two runs per arm.

arm	correct / scored	accuracy
nvfp4 / nvfp4 tower	175 / 200	0.875
mxfp8 / 8-bit tower (local)	173 / 200	0.865
bf16	169 / 200	0.845

Accuracy falls monotonically as precision rises, every arm reproduced its accuracy exactly across two cold runs with zero items changing verdict, and the mlx-cuda rows show the same ordering independently.

nvfp4 beats bf16, 6 items to 0, exact McNemar p = 0.031 — the only pair here that clears the test. Stated with both caveats it needs: three pairwise tests were run, so Bonferroni (≈0.09) puts it back above 0.05, and rows 0–200 are OCRBench's text-recognition half, not the benchmark.

Which artifacts — this is the load-bearing part

store_audit.py says all seven quantised gemma4 MLX tags on this host have diverged from the registry, so every arm is named by manifest digest and run against a preserved tag a pull can't overwrite:

• 637cc0ff1570 is the same artifact as this ladder's mlx-cuda nvfp4 / nvfp4 tower row — those two rows are one checkpoint on two backends, which answers the question the doc raised about the 0.875.
• The mxfp8 row is not comparable across backends. The local copy differs from the registry's bf16-tower one by 579 of 1247 layers (language 411, vision 162). It's a local artifact, and the doc says so rather than lining the numbers up.

The arms are committed

Under bench-runs/ocrbench/ with a .gitignore negation, so every table re-renders from the repo alone — verified by re-rendering from the committed copies. Dataset rows are deliberately not committed: --categories reads extbench's local cache, and this repo shouldn't redistribute the benchmark's content.

Also records the 1000-item kernel A/B behind ADR 0037 (835 vs 833, p = 0.804).

test_summarizers 144, test_store_audit 9, test_extbench 8 — all OK.

🤖 Generated with Claude Code