ahe

Rule

AHE’s per-round decision rule, as a HarnessGrad method.

The method returns one edited harness per round. It never sees a score for the harness it is producing — the platform measures every candidate on the same study set and records the result in curve.jsonl.

Measured results

Scores, rounds and per-split comparisons are on the ALFWorld benchmark page — one table for every method, so the same number cannot say two different things in two places.

Record

This method’s rounds on the Hub — each round’s harness, the edit that produced it, and its per-task trajectories: open the directory.

Deviations from the method it ports

7 deliberate difference(s), recorded in methods/ahe/method.json:

• Rollback uses a different edit base plus an instruction, at whole-candidate granularity rather than restoring individual declared changes.
• Two proposal variants are generated per call in the upstream n_variants sense (evolve.py:3123-3151), but winner selection remains with the platform by train score; AHE’s own winner rule cannot be expressed in mode A.
• The batch deterministically falls back to its first usable candidate when change declarations or applied contents are identical; a declined, invalid or failed sibling also reduces the usable batch.
• Exact shared-base attribution requires the recorded parent point to remain in the staged history.
• Per-task rollouts stay at 1 because the platform runs --trials 1; upstream’s k (evolve.py:599-602) passes a task only if ALL k rollouts pass, so this is a run configuration, not a difference in the ported rule.
• Each variant uses one shared-editor request instead of AHE’s full tool-using evolution and debugger workflow.
• AHE’s component taxonomy travels only as a declared string.

Code

methods/ahe/run.py and its rule test in tests/quality/. python3 tools/method_lint.py prints one line per method.