A 60-line static probe found 3 contamination points in an agent eval harness, exit 1, without running the agent. Passing the eval is not solving the task.
Passing the Eval Isn't Solving the Task: 3 Leaks, 60 Lines
Related
I built a WhatsApp/Telegram group directory with zero coding background — here's the whole, unfiltered story
I'm not a developer. I want to say that upfront, because it matters for everything that...
Building Shiksha: What I Learned Creating a Real-Time AI English Coach in 10 Days
The complete technical journey of building Shiksha, an AI English Communication Coach with persistent memory, telephony, human handoff, and multi-agent architecture.
I checked what 6 AI coding tools write to disk to stop the agent forgetting its plan. Half write nothing.
Every AI coding tool has the same top complaint: "it forgot what we were doing." I spent weeks...