Upstream contribution · harbor-framework/terminal-bench-science
terminal-bench-science · Reject non-finite numbers in a scientific benchmark evaluator
Problem
A scientific output verifier parsed floating-point values and accepted non-finite representations as numeric findings. That let malformed values satisfy the shape of a finding without providing a usable estimate.
Engineering response
Check finiteness after numeric parsing and retain the benchmark's existing missing-value path. Leave the scientific scoring method and admissible finite values unchanged.
Evidence and scope
The pull request reports negative controls for non-finite and overflow values, plus passing finite and missing-value controls. The evidence is local verifier testing.