A SMALL CHANGE CAN MOVE THE ANSWER

The code still runs.The result moves.

A library update can change a scientific conclusion without throwing an error. ReproDelta makes that difference visible, testable, and easier to explain.

Investigate a PR or MR

Start with a change link. We review its diff before offering a paired run where the supported contract and environment pins match.

ONE INPUT · TWO ENVIRONMENTS
01 / BEFORE · NUMPY 1.26.4Above threshold0.5000000149011616 float64
02 / AFTER · NUMPY 2.0.2Inconclusive0.5 float32
Same eight-value signalDifferent declared outcome
01 / THE EVIDENCE

One input is enough to ask a better question.

A supported paired investigation runs the same input in pinned NumPy environments, checks repeatability, probes the mechanism, and tests candidate fixes against held-out cases.

Start with your change
02 / THE METHOD

From suspicion to evidence.

01

Hold the input still

Run the same valid case in both versions, with repeatability checks before comparing them.

02

Locate the difference

Reduce the witness and inspect the measured values and intermediate types.

03

Test the explanation

Change one suspected mechanism, then check whether a repair survives independent cases.

03 / YOUR REPOSITORY

Have a change to review?

Paste a GitHub pull request or GitLab merge request. ReproDelta reads its changed files, highlights relevant dependency evidence, and gives you specific follow-up checks. Only changes matching the supported pinned workflow can proceed to a paired run.

Open the review workspace
YOU BRINGOne change link.FIRST STEPA focused diff review.IF SUPPORTEDApprove and run a paired investigation.