Reports

Every number in the README and the guides comes from one of these reports, measured on an Apple M4 Pro with 48 GB unless the report says otherwise.

report	what it measured
001 Phase 0	mmap on a model bigger than RAM reached 11 to 13% of the disk rate
002 Phase 1	the streaming runner: 141 GB at 79.2% of the disk rate
004 Phase 1.5	correctness and headroom
005 Phase 1.6	share-ready, and the proof run as a measured partial
006 Phase 2	the loop: distill, eval, adapters at inference
007 Phase 2.5	long-tail engine fixes and the completed 2,000-row set on the 70B
008 Phase 3	spill tune: streamed against resident LoRA, and the 70B smoke test
009 Phase 3.5	spill build, the example, export, prefix reuse
010 Cleanup	the merge, and a fresh install timed
012 Run anywhere	portable jobs, the PyTorch engines, cross-hardware resume gates
013 Tagline, README, docs site	the tagline, the rewrite, the site and the decisions behind them (no measurements)
014 Build anywhere	build on every engine, a build that moves between machines, the CI relay
015 One complete workflow	init, plan, build, report, test, export and inference per task and engine; verified exports; the ownership and handoff gates
016 Closing the 015 gaps	the offline bundle gate with the network blocked, MLX-on-Metal continuation in both directions, merge hygiene
017 Scorer and protocol	JSON scorer version 2, historical JSON accounting, decoding settings applied, evaluation records, compare outcomes, replication
017 Export diagnosis	controlled merge, runtime and quantization comparison for JSON and classification exports; the merge-dtype default
018 README pass	README counts before and after, the real MLX capture, and where each moved block lives
020 Restore the vision	any task as one kind of project, a 1.5B teacher and judge on MLX and torch-cpu, judged evaluation with its calibration and its failures, logit distillation, deadline planning