SuperGPQA independent evolution data

SuperGPQA's published pool has no native training split. An explicit external ARC pool can supply ModularRSI evolution examples through public configuration:

benchmark:
  name: supergpqa
  evolution_arc: {subset: ARC-Challenge, split: train}
  evolution_limit: 1
  split_counts: {train: 1, dev: 1, test: 1}
  seed: 42

The loader reads the complete native Parquet file from allenai/ai2_arc at revision 210d026faf9955653af8916fad021475a3f00453. Supported subsets are ARC-Challenge and ARC-Easy; splits are train, validation, and test. Validation/test selections remain explicitly external published evaluation data, never an official SuperGPQA training partition.

Before limiting examples, the loader checks the complete downstream SuperGPQA pool against the complete external pool by original task ID and normalized raw question hash. A collision rejects the source; no overlapping record is removed or renamed. Only experiment train is replaced. SuperGPQA's dev/test selection, final-letter accuracy scorer, and algorithm gates remain unchanged.

ARC option text and ordering are preserved. Numeric native choice labels are mapped by option position to the existing letter protocol, retaining native_choice_labels and native_answer_key in task metadata. No answer is generated or corrected. Source, fixed revision, subset, native split, full pool size, checksum, and external evolution role are recorded automatically.

The complete Challenge/train file has 1119 records. Its one repeated raw question is preserved under its two distinct native IDs; both records remain training records and neither is removed. A complete-source probe found no question overlap with the complete 26529-record SuperGPQA pool. This establishes a data configuration, not a successful model experiment or a leaderboard result.

See the public example. Claude Code uses the existing public three-module example at examples/modular_rsi_claude_code; its Worker keeps a separate native role configuration. The MiniAgent example keeps the usual five modules and fixed budgets.