Running and scoring HERA-Bench
Controller and solver
Run the environment/controller in a separate process. Give the solver the selected record's instruction and system_prompt, then connect it to the MCP server:
python -B benchmark.py serve --number 1 --side act --record ../episode-001-act.json
The side is selected by the controller. Keep the command arguments, state files, repository, references, and grading artifacts inaccessible to the solver. The MCP server exposes domain tools and two supporting-document tools. It does not expose state-export, evaluator, oracle, or filesystem tools.
--record writes a new controller-only file with {events, final_state} after tool calls. Domain events have {tool, arguments, observation}; tool is the unqualified domain tool name. Document reads are outside this domain event list. The recorded final_state is the domain state. Store recordings outside the downloaded benchmark tree; an existing record path is rejected.
Submission and grading
Attach the agent's final JSON report as answer, or its textual final response as final_output, to the trusted controller record. Supply exactly one of those two fields:
import json
from pathlib import Path
record = json.loads(Path("../episode-001-act.json").read_text())
record["final_output"] = Path("../agent-final-response.txt").read_text()
Path("../submission-001-act.json").write_text(json.dumps(record))
python -B benchmark.py grade --number 1 --side act \
--submission ../submission-001-act.json --output ../grade-001-act.json
The public brief defines the expected report. Grading validates format and contents, authenticates tool observations by replay, and checks final-state effects. passed is the complete instance verdict. An invalid report has format_passed=false and no content grade; a malformed execution record is separately marked. The preserved grader includes the evaluated tool-alias and complete-search equivalences. Call the top-level benchmark.py grade entry point to use that complete grader.
Reference programs are available to the controller:
python -B benchmark.py reference --number 1 --side abstain --output ../reference-001-abstain.json
A valid reference path is not necessarily the only valid path. Reference answers, traces, and feasibility labels must not be supplied to the solver.
Aggregate results
Evaluate all 60 pairs, both sides, with the same declared harness/model settings. Let c_i and b_i be the boolean passed verdicts for act and abstain respectively. Report:
| Metric | Calculation | Denominator |
|---|---|---|
| Act | 100 * sum(c_i) / 60 |
60 feasible instances |
| Abstain | 100 * sum(b_i) / 60 |
60 should-abstain instances |
| Pair | 100 * sum(c_i and b_i) / 60 |
60 matched pairs |
Also report numerators. Pair must not exceed either Act or Abstain. Aggregate by number or pair_id, never by independently sorted side results. Keep failed episodes and missing results visible; an incomplete run cannot be reported as full-benchmark accuracy. A full-benchmark completed run should count unsuccessful runtime episodes as unsuccessful and report their number separately. Record any alternate policy and excluded IDs explicitly.
For uncertainty estimates, resample matched pairs together. Environment breakdowns use their actual pair counts, without silently replacing the overall 60-pair average with an unweighted mean across environments. Over-abstention requires a separate annotation of explicit refusal on feasible tasks; it is not the complement of Act and is not produced by a dedicated classifier in this release.
The manuscript's base harness has a 30-turn execution limit. The environment server does not implement a model runner, enforce that turn budget, or ship the evolved harness. Reproduction of paper model scores additionally requires the selected harness, model/provider configuration, generation settings, and runtime-failure policy. Report these alongside the dataset commit. Offline reference verification is not reproduction of model scores.
Offline integrity check
python -B benchmark.py verify
This checks all 120 reference reports, public MCP tool replays, document access, matching paired tool interfaces, trace recordings, and final states. It also rejects forged observations and final-state effects. Network connections are disabled during this check; no paid model calls occur.