# November 3 condensed M5 Max comparison

72 matched trials from 73 recorded; 18 pairs per model, one repetition. This is an exploratory subset, not the original 516-trial study. Equivalence within ±5 percentage points is not established. The flagged unpaired pass remains in the detailed downloads.

- [Report](report.md) and [aggregate data](analysis.json)
- [Paired conclusions](paired-analysis.md) and [data](paired-analysis.json)
- [Performance CSV](performance.csv) and [data](performance.json)
- [MLX prefill method](mlx-prefill-recovery.md) and [observations](mlx-prefill-recovery.json)
- [Judgments](judging.md) and [scores, axes and prompts](judging.json)
- [All 73 trial records and postmortems](scorecard-details.json)
- [Scorecard dataset](scorecard.json)
- [Scope amendment](scope-amendment.md) and [original analysis plan](analysis-plan.md)
- [Native provenance](native-provenance.json), [completion audit](completion-audit.json), [publication verification](verification.json), and [file hashes](manifest.json)

Recovered MLX prefill is partial, rounded first-token rate evidence, not exact prefill time or a duration-weighted request aggregate. Candidate Gemma MLX drafting, quantization differences and telemetry limits are documented in the report.