🤓 Latest Submissions
Prove Me Wrong
The problem. New developers don't fail because they lack skill, they fail because their first few weeks are spent silently misreading an unfamiliar codebase, with no one checking whether their assumptions are right. That gap is expensive and invisible until it causes a bug months later. New developers typically take 3 to 6 months to reach full productivity, and onboarding one developer costs $20,000 to $80,000 in the first 90 days. 28% of new hires quit within that window, largely due to poor onboarding, while strong onboarding improves retention by 82% and productivity by 70% (Cycloid, AllenComm, HR Dive, Brandon Hall Group / Glassdoor). Existing tools either produce static documentation nobody reads, or repo-visualization dashboards that show structure without checking whether the new hire understood it. The insight. Reading code passively does not build real understanding. Testing your own assumptions and getting corrected does, this is well established in how people actually learn, active recall beats passive review. Nobody applies that principle to codebase onboarding today. The solution. Prove Me Wrong turns onboarding into a conversation. It picks real, tricky spots in a codebase, curated in advance, and asks the new developer what they think happens, not a quiz with made-up questions, but ones grounded in the actual, current code. IBM Bob 2.0 reads the live repository and generates the question, works out the ground-truth answer, and cites the exact file and line that proves it. When the developer guesses wrong, they see the real correction with that citation, not just a red X. A gentler follow-up question on the same topic reinforces the concept before moving on. At the end, they receive a personalized summary of what they understood and what to revisit, synthesized by Bob from their own session. Why IBM Bob 2.0, specifically. This only works because Bob reads full repository context, how pieces connect and what the current source actually says, not a single file in isolation. A single-file assistant could not generate genuinely current, provably correct questions this way. While building this tool, Bob itself made a reasoning error three separate times on the same topic, caught only by tracing the cited file and line by hand. That is not a flaw, it proves the thesis: even a powerful, repository-aware AI benefits from citation-grounded verification, and a workflow built around proving claims rather than trusting them catches what a plain AI chat interface would silently pass along. The demo. A live session against axios/axios, a real, widely-used open-source project, not a staged example. A wrong guess, a real correction with proof, an adaptive follow-up, and a personalized summary at the end. The impact. Every company hiring developers pays this cost today. Prove Me Wrong replaces weeks of silent, unverified reading with a focused session of active, proven understanding, at a fraction of the time and cost of the status quo.
27 Sep 2026
