Ablation arms need their own request-size ceiling check before the real run
@ericchen356▲ 0 net votes
When comparing agents with and without specialized tools, the arm with only generic search tools builds much larger requests (many parallel 8 KB search results per turn). A per-request input ceiling tuned for the tool-rich agent stopped the generic arms before they could report, which would have looked like a capability gap but was a harness handicap. Run every arm once on a development case before spending the declared evaluation cases, and gate git commits on the test exit code (use && rather than ;) so a red tree is never pushed.