[zuraiz-anjum]

Summary

_parse_canonical_reward in tbench2_env returns None for an empty or non-numeric reward.txt, and the callers raise so the episode is reported as a scoring error. But float() also accepts nan, inf and -inf, so a reward.txt holding one of those went through as the reward. This makes a value that is not finite count as no verdict too.

Type of Change

• Bug fix
• New feature
• Breaking change
• Documentation
• New environment
• Refactoring

Alignment Checklist

Before submitting, verify:

• I have read .claude/docs/PRINCIPLES.md and this PR aligns with our principles
• I have checked .claude/docs/INVARIANTS.md and no invariants are violated
• I have run /pre-submit-pr (or bash .claude/hooks/lint.sh and tests) and addressed all issues

RFC Status

• Not required (bug fix, docs, minor refactoring)
• RFC exists: #___
• RFC needed (will create before merge)

Details

Since #1025, a missing verdict raises in _require_canonical_verdict, so RL consumers can drop the episode instead of training on a false 0.0. On main, __TB2_REWARD__:nan is parsed as float("nan"), _evaluate_task and _evaluate_docker return reward=nan with tests_passed: False, and the NaN goes to the trainer. inf goes through the same way.

The other envs that read a reward file already reject this case: _parse_reward_override in coding_tools_env, jupyter_env and terminus_env returns no reward when math.isfinite(value) is false. This applies the same check in _parse_canonical_reward, so both the local and the Docker path raise "canonical harness produced no verdict", the same as for an empty reward.txt. I also updated the docstring.

Test Plan

Two tests in tests/envs/test_tbench2_env.py, next to the missing-verdict tests from #1025:

• test_evaluate_non_finite_verdict_raises, for nan, inf and -inf on the local path. It also checks that the staged tests and the verifier dir are still wiped.
• test_evaluate_docker_non_finite_verdict_raises for nan on the Docker path, with the fake container.

Run with PYTHONPATH=src:envs pytest tests/envs/test_tbench2_env.py:

• main with only the new tests: the 4 new cases fail (DID NOT RAISE), the existing verdict tests pass.
• this branch: the 4 cases pass and nothing else changes.

The three checks from .claude/hooks/lint.sh (usort check, ruff format --check, ruff check) are clean on the files this PR touches; the findings a tree-wide run reports are in other files and are the same on main.

---

Note

Low Risk
Scoring-path guard only; finite rewards unchanged. Low blast radius—errors instead of NaN/inf rewards.

Overview
Canonical TB2 reward parsing now treats nan, inf, and -inf in the verifier marker line as no verdict (None), not a valid score—float() accepts those strings, so they used to propagate to the trainer.

The docstring on _parse_canonical_reward was updated to match. Tests cover withheld local evaluate for all three non-finite strings (staging still wiped) and Docker evaluate with a nan marker, both expecting the same RuntimeError as an empty or missing verdict.

^{Reviewed by Cursor Bugbot for commit f6c2840. Bugbot is set up for automated code reviews on this repo. Configure here.}