Anecdotes

Short stories from building Vektos where something went wrong, or almost did, in a way that teaches something about the benchmark itself. Each one ends with what it means for the project and what, if anything, we changed because of it.

Why keep them: Vektos measures how agents behave under a strict design system. Every time a human hits the same wall, we get a free data point about what the graders should catch and what the agent context should say.

Each anecdote follows: What happened, What we learned, What it means for Vektos, Follow-ups.