🤓 Latest Submissions
Cimex Fix — No fix without proof
The problem. AI assistants can suggest a patch in seconds, but nobody proves it fixes this bug or breaks nothing else. In Stack Overflow's 2025 survey, 66% of developers named "AI solutions that are almost right, but not quite" their top frustration, and 45% said debugging AI-generated code takes longer. Reviewers get a diff, not evidence: no test that failed before and passes after, no commit that caused the bug, no record of what was checked. Who it is for and how it works. Developers on bug-fix or on-call duty, reviewers, and QA engineers who file bugs. A bug report — a customer screenshot, a QA PDF, a server log or an issue — goes into IBM Bob, and the developer picks the Cimex Lead mode. The Lead spawns three subagents in parallel (Triage reads the report, Locator maps the code path, Historian checks recent changes), then hands off to permission-scoped modes: the Reproducer may only write test files and must produce a test that fails on an assertion; git bisect, run through our MCP server, uses that test to name the exact culprit commit; the Fixer may only edit source files; a Critic with no edit permission reviews the fix and can send it back once. Mission Control shows the agents working live and publishes a shareable Proof of Fix page with the failing test, the culprit commit, the diff, the review, a plain-English summary from IBM watsonx.ai Granite, and commands anyone can run to verify it. What is new. Cimex Fix treats every fix, from any model, as unproven until there is evidence. A changed test blocks the proof: each mode edits only its own files, and tests are locked at RED. Every claim on the Proof page is backed by recorded evidence, and failures stay visible: when bisect missed a culprit during a run, the case says so. Impact so far. On a demo shop with 8 planted bugs, Bob proved all 8, each from its original report. All 8 culprit commits match the seeded answer key (4 named during the run, 4 by a follow-up bisect); the full mode hand-off ran in 4 runs, every run after we fixed the pack. The median time from report to a failing test was 4m 02s; the cleanest run went from customer issue to published proof in 7m 44s with no human prompts. The median cost was 2.93 Bobcoins per fix, no fix changed an existing test, and a reviewer re-checks a proof from a fresh clone in 28.8 s. Seven models (five on IBM watsonx.ai, Gemini 3.1 Pro, Claude Opus 4.6) got the source and answered each bug once, running nothing: 18 of 56 answers looked finished but did not fix the bug, and only running them showed which. Cimex Fix is that run, automated for every fix; gpt-oss-120b matched its 8/8. Every number is script-generated from recorded data.
27 Sep 2026
