[dev-rinchin]

The default local GPU benchmark uses the real Delicious dataset and the evaluation protocol from the SketchBoost paper, replacing synthetic data in the required PR check.

Workload and reporting

• Pinned original archive and normalized array checksums; first official train/test split (12920/3185), five shuffled training folds (10336 train / 2584 validation), all 500 features and 983 labels. No generated or resampled targets.
• Unrounded selected paper parameters, lr=0.015, up to 20000 iterations, es=500, depth=12, 256 bins and use_hess=False. Hyperparameter search is not repeated.
• Full-output baseline plus topk/rand/proj with sketch_outputs=5. FilterSketch is explicitly additional library coverage outside the paper. Five methods × five folds execute sequentially on exactly one physical GPU.
• Synchronized fit/predict timing after a 32-iteration warmup, a separate full memory fit, CPU RSS, test log loss and actual executed/retained tree counts per fold. Summary statistics use five-fold means and sample standard deviations (ddof=1), following the paper's aggregation code. Reports include GPU characteristics and maximum fit GPU allocation across folds.
• Explicit dataset preparation, an offline archive option, and per-fold resume. Comparisons reject mismatched workloads, datasets, environments and harness versions. The default worker timeout is 72 hours, accommodating expensive full-output fits.
• AGENTS.md defines the required local workflow. Legacy synthetic profiles remain explicitly selectable. Reports/data stay ignored locally; publication uses PR comment tables, without result branches or artifact uploads.

Validation

• 332 tests passed on one NVIDIA A100 80 GB GPU, including GPU library tests and dataset-integrity/fold-isolation/resume/reporting tests.
• GitHub CPU checks passed on Python 3.13 and 3.14 for final head d7be28b27614.
• All 25 method/fold cases passed on each of two library revisions in a real-data diagnostic capped at four iterations and two warmup iterations: base da8df750a760 and candidate 4d2d89f67364. Both timing and separate memory fits passed; actual dataset checksums matched for all paired cases. Positive resume verification skipped completed cases. These diagnostic results are not a performance benchmark at convergence.
• After the subsequent report-only sample-standard-deviation correction and timeout/documentation update, the final harness was rerun on both base and final head with forced early stopping (zero learning rate, es=2, limit=16): timing and memory runs each executed three iterations and retained one tree. Reporting tests verify the sample deviation formula; saved diagnostic reports also render with the corrected statistics without altering their original recorded hashes.
• Wheel and source distributions build and contain no cached data or generated results.

Draft: the full five-fold comparison at the paper's 20000-iteration limit has not been run. The unsketched baseline is expensive: a separate 32-iteration pilot fit took approximately 114 seconds on an A100. Comparing two versions across all folds with separate full memory fits can take days. No full-protocol quality or performance result is claimed.