I feel like asking the thing that you are measuring, and don’t trust, to measure itself might not produce the best measurements.
What is "drift"? It seems to be one of those words that LLMs love to say but it doesn't really mean anything ("gap" is another one).
IDK how it applies to LLMs but the original meaning was a change in a distribution over time. Like if you had some model based app trained on American English, but slowly more and more American Spanish users adopt your app; training set distribution is drifting away from the actual usage distribution.<p>In that situation, your model accuracy will look good on holdout sets but underperform in user's hands.
I believe it's businessspeak for "change." Gap is suittongue for "difference."
Interesting approach, I've been particularly interested in tracking and being able to understand if adding skills or tweaking prompts is making things better or worse.<p>Anyone know of any other similar tools that allow you to track across harnesses, while coding?<p>Running evals as a solo dev is too cost restrictive I think.
See the very last section in this doc for how I minimise token usage and track savings, all three plugins co-exist fine: <a href="https://github.com/FrankRay78/NetPace/blob/main/docs/agentic-software-development-workflow.md" rel="nofollow">https://github.com/FrankRay78/NetPace/blob/main/docs/agentic...</a>
See also <a href="https://marginlab.ai/trackers/claude-code-historical-performance/" rel="nofollow">https://marginlab.ai/trackers/claude-code-historical-perform...</a> for a more conventional approach to track regressions<p>This project is somewhat unconventional in its approach, but that might reveal issues that are masked in typical benchmark datasets
[dead]