Repair verifier bootstrap failures before trusting rerun results
A zero reward is not always an output failure. If the test runner exits because a command or binary path is missing, task-specific checks never executed. Inspect the verifier output first, then repair the test harness dependency and rerun.
Procedure:
- Separate bootstrap errors from assertion failures. Look for:
command not foundorNo such file or directory- a fixed binary path that does not exist (for example
/root/.local/bin/env) - transport failures such as
curl: (35) TLS connect errorbefore setup commands run
- Identify the missing dependency used by the test script. Common cases:
uvxoruvmissing: install it, e.g.
orpip install uvcurl -LsSf https://astral.sh/uv/install.sh | sh- a toolchain command invoked through a fixed path: create the expected path or symlink after installing the real binary.
- Make the repair visible to the test runner. Ensure the installed binary is on
PATHfor the shell that executes the tests, including/root/.local/binwhen the harness expects tools there. - Confirm the task's output artifacts still exist after installing dependencies, then rerun the verifier.
Why: in one run the solver's /app/merged_users.parquet and /app/conflicts.json were correct, but the verifier exited at uvx: command not found and /root/.local/bin/env: No such file or directory, so the reward was 0.0 before any task logic was evaluated.
Supporting evidence: multi-source-data-merger (this batch).