[DimaMolod]

The recursive report (generate_recursive_report: true, the default) had two problems.

It replaced the per-fold tables. The rule ran AlphaJudge with recursive_report_arguments only, which defaults to {}. AlphaJudge therefore scored with its own defaults (--models_to_analyse best) instead of analyze_structure_arguments (all in the default config). Every fold's interfaces.csv then failed AlphaJudge's cache check, was rescored, and was overwritten with a best-model-only table.

It ran out of memory and took days. The rule requested 2,000 MB per attempt and one CPU. In a 1,900-fold AF2 + AF3 campaign, five attempts were OOM-killed at 2–4 GB after 1–32 h each, and one process needed days for the whole tree.

Changes

• The report scores with analyze_structure_arguments, so it reuses each fold's table. recursive_report_arguments add report-only flags (--aggregate_report) or override them. A list or string is used as written.
• New recursive_report_threads (default 8). The job gets that many CPUs, and AlphaJudge gets --cores unless the arguments set it.
• Memory comes from a per-run O(N²) model, assuming one worker holds the largest fold and the others hold median-sized folds. Retries scale it by 1.5×, and max_mem_mb caps it.
• The model's coefficients exceed the peaks of ~1,900 per-fold AlphaJudge jobs (600–8,800 tokens) by at least 1.3×.
• For that campaign, the model requests about 35–37 GB for 8 workers. The 8-worker reports peaked at 8.5–12.1 GB, and the AF2 report finished in 10.7 h.

Tests

• Unit tests for argument merging and memory sizing.
• A dry run that renders the report with the analysis flags, --cores and its threads. It fails on main.
• Full suite: 231 passed.

🤖 Generated with Claude Code