# I’m sick of AI “Thinkslop” in my PRs > Source: > Published: 2026-07-15 10:20:29+00:00 I started my career as a “Dessinateur Projeteur” (Design Drafter) in mechanical engineering. In that world, if your plan is off by 1mm, the part is scrap. There is no “hotfix” for a piece of steel once it’s machined. When I started building my own software tools using LLMs, I brought that same paranoia with me. And honestly, it’s the only thing saving my codebases right now. Lately, I’ve been fighting a specific enemy: Thinkslop. It’s that code that looks “clean,” passes a linter, and even runs. But the logic is hollow. It’s the result of asking a single AI to “just fix this” and hitting copy-paste because you’re in a rush. It’s dangerous because it looks right. One model is not a code review I don’t care if you’re using Claude 3.5 Sonnet, GPT-4o, or a custom fine-tune. A single model has a single perspective based on its training. If the model is biased toward a specific pattern, it will confidently hallucinate a bug right into your production. Most devs think that if the AI says “Looks good!”, they are safe. The truth? The disagreement is the only signal that matters. The “Consensus” approach I got so tired of manually comparing browser tabs or doubting my own code that I started running my logic through 8 different models at once (Claude, GPT, Gemini, Llama, Groq, etc.). ``` When they all agree? It’s probably safe. When Claude flags a potential memory leak but GPT doesn’t? That’s exactly where the real bug is hiding. ``` I built a local-first engine called NexaVerify to handle this orchestration for me. It doesn’t just “vote” on the code; it surfaces the friction. It highlights the spots where 8 different “brains” can’t agree on the architecture. I ran it on its own 26k line codebase last week. It found a NameError in a constant file that had been sitting there for a month. A single pass from GPT had missed it every time because the “vibe” of the code was correct. The takeaway Stop treating AI as an oracle. Treat it as a junior dev who is prone to lying but very fast at it. As a former “Dessinateur Projeteur,” I know that a structure is only as strong as its verified points. The only way to trust AI code is to force it into a consensus. How are you guys catching AI hallucinations before they hit prod? I’m curious if anyone else is building multi-model pipelines or if everyone is still just “vibe coding” and hoping for the best.