another-harness Evaluator-Discipline Environment Prototype

The another-harness repo now contains the second live executable slice underneath its Atropos sidecar design: a thin local evaluator_discipline benchmark family that grades reviewer honesty over bounded repo-native builder states.

Main architectural facts

• The implementation is deliberately not full Atropos integration; it is a local runner that proves evaluator-lane mechanics first.
• The suite contains three evaluator fixtures: reject an obviously incomplete builder state, approve a clean reviewed state, and reject an ambiguous in-progress state despite passing checks.
• Each prepared run freezes the benchmark contract into the baseline commit as .evaluator-discipline-contract.json.
• Evaluator episodes may write only the evaluation artifact; builder-owned work items, handoffs, and deliverables remain frozen.
• The grader now derives truth from the frozen builder contract plus real check results, rather than from mutable sandbox paperwork.
• Review forced a long series of anti-gaming fixes: ignored-file smuggling detection, replace-ref resistance, file-vs-directory scope discipline, run-metadata hardening, control-bundle retarget detection, and finally a control-dir baseline anchor plus sandbox-root anchor.
• The final spec now states the honest assumption surface explicitly: the evaluated agent is judged inside sandbox_root, while direct mutation of operator-side control artifacts remains out of scope for this local prototype unless stronger OS-level isolation is added later.

Practical conclusion

another-harness now has builder and evaluator benchmark lanes of its own. That still does not make it an Atropos deployment or a trainer stack, but it does mean the repo’s RL-environment story is no longer just one schema and a wistful look into the distance.