Exhibit A — Debugging on Trial

AI coding assistants can produce convincing explanations for bugs without proving them. A plausible diagnosis may blame the wrong commit, confuse correlation with causation, or recommend a fix that leaves the underlying problem untouched. Developers must then verify the investigation themselves. Exhibit A puts debugging on trial. Built for developers and engineering teams, it combines IBM Bob's investigation capabilities with executable evidence and independent scrutiny. In Bob IDE, a custom Tribunal mode coordinates parallel investigations into code, Git history and logs. An independent prosecutor challenges the investigator's claims. A deterministic judge, implemented in plain code rather than an LLM, reruns the evidence: reproduction tests, tests at historical commits, log searches and Git bisects. The case schema prevents withdrawn or unsupported claims from reaching the verdict. The web application replays the investigation as an original pixel-art courtroom visual novel. Users follow accusations, objections and withdrawals, open the evidence record, and read a final report containing the root cause, culprit commit, proposed fix and verification commands. Replay mode makes no AI calls. Our demonstration includes three recorded investigations on Crumb & Co, a sample bakery API with deliberately planted bugs. In the flagship case, two customers buying the last baguette cause negative stock. Bob initially accuses a recent refactor, but executable evidence establishes that the bug already existed before it. The investigation ultimately identifies an earlier asynchronous inventory change. Exhibit A's distinction is not just its courtroom presentation: disagreement is part of the workflow, and the verdict depends on reproducible evidence rather than an assistant's confidence. It turns debugging into an investigation that developers can inspect, challenge and rerun.