Agent Orchestrator Benchmark Exposes Sandbox Bug

This is a benchmark story, but the real lesson is eval hygiene: a silent harness bug can turn a judge into a confident liar. Once the spill path was fixed, MiniMax looked mediocre for real reasons, which is exactly the distinction benchmarks are supposed to preserve.

• –Two autonomous sessions blamed MiniMax for not being able to implement the task, but the actual failure was that the implement step received an empty input.
• –The fix changed the story from “model cannot write code” to “model writes plausible code but still has compile/test issues,” which is a very different signal.
• –The step success metric was misleading because downstream `self_test` and `benchmark_eval` failures distorted the apparent quality.
• –Any autonomous leaderboard should track artifact provenance and pipeline sanity separately from model scores, or the ranking is partly infrastructure noise.