TriviaQA independent evolution data

Use native training records for evolution and keep the downstream experiment partition drawn from the complete labeled validation pool:

benchmark:
  name: triviaqa
  dataset:
    source: mandarjoshi/trivia_qa
    config: rc.wikipedia
    revision: 0f7faf33a3908546c6fd5b73a660e0f8ff173c2f
    splits: [validation]
  evolution_split: train
  evolution_limit: 1
  split_counts: {train: 1, dev: 1, test: 1}
  seed: 42

evolution_split accepts native train or validation. The complete external and downstream pools are checked by original ID and normalized raw question hash before any evolution limit is applied. Context changes do not hide the same question. An overlap rejects the configuration; no row is removed, renamed, or relabeled. Including native train in the downstream pool while selecting the same train pool for evolution therefore fails the guard.

Only experiment train is replaced. Existing dev/test partitioning, original Wikipedia contexts, answer aliases, the official alias-aware EM/F1 scorer, and algorithm gates remain unchanged. The loader transfers the native alias map for the selected external training tasks. Legacy scoring remains an explicit alternative and does not claim the official scorer identity.

evolution_dataset optionally overrides the external dataset source mapping. It follows the same streaming dataset reader as benchmark.dataset; its selected native split is always controlled by evolution_split. For a complete verified local directory containing native Parquet shards, both mappings may use source: /absolute/path/to/rc.wikipedia and config: default. Local Parquet builders use default, whereas the upstream HF repository uses rc.wikipedia. Use absolute local paths; dataset source identifiers and local paths follow the underlying loader's current-directory resolution, not YAML-relative path rules. Keep revision pinned to the actual original artifacts and record their hashes and cache provenance. This is an explicit complete native cache mode, not a claim that default online downloads succeeded.

The verified native pool at this revision contains 61888 training records in seven Parquet shards and 7993 validation records. Complete-pool probes found no original-ID or normalized-question overlap and no repeated training IDs or questions. These probes and reference-answer scoring are data evidence only; ModularRSI still needs a real experiment through the public entry point.

Source, revision, native split, complete pool size/checksum, and the external evolution role are recorded by the loader. See the public example. Claude Code uses the existing public examples/modular_rsi_claude_code three-module example; MiniAgent retains five modules, with an independently configured native Worker.