One JobBench task through Harness²

    architecture node:

    Representation at this boundary

    Source modules

      Loop map

      One task — the adapter hands over a criteria-stripped task prompt → the base harness is rolled k times → evidence is written as files → two one-sided proposals edit the tree → repair + validation → k independent compositions are rolled → blind selection over the finals → reflection writes the library entry.
      The budget is arithmetic — k is the only dial: R = 4k rollouts and P = 3k proposer sessions per task and mode, and the ledger asserts both after the run (duplicates are checked before the counts, because a duplicate can satisfy them). The k base rollouts are shared, so running both modes costs 7k rollouts, not 8k.
      Scoring is outside the loop — nothing in these 8 steps reads a grade: RunView carries no score field, and the judge module is never imported by the method. The judge runs afterwards, in a separate command, over the run selection already chose; a judge failure records None, never a zero.