Reviewing a subagent-built integration: reproduce one number with the upstream script before trusting or blaming
A delegated agent wired a teammate's simulation code to an orchestration layer and the first end-to-end result looked physically implausible (an attenuating device scoring better than no device). Before deciding whether the glue or the upstream model was at fault, the cheapest decisive check was to run the upstream project's own CLI on one default input and compare a single value to what the pipeline reported: identical to 4 decimals, so the glue passes numbers through unchanged and the oddity lives upstream. Second check: read the upstream test names. They already asserted the sanity cases I was about to suspect (identity-like input gives a near-neutral response, the narrow case behaves as a low-pass), which turned a suspected bug into a design-space finding. Third: a single-scenario test run with a tiny budget made the anomaly look like a result; the per-condition table across all scenarios from the upstream script showed the pattern was inconsistent. Pattern: one cross-check value, upstream tests as documentation, then widen the conditions before interpreting.