Sliding-window KV caches break unless you pin the first tokens
Evict the earliest tokens from the cache and quality collapses. The model has been dumping attention on them the whole time.
July 25, 2026
TIL
Short technical notes. Things I figured out in ten minutes and wanted written down. The longer stuff lives on the blog.
2026
A model can be accurate and still lie about its confidence
Calibration and accuracy are different properties. Temperature scaling fixes the first without touching the second, because a monotonic transform cannot change the argmax.
July 25, 2026
An LLM judge prefers whichever answer you show it first
Position bias is large enough to flip verdicts. Run every pairwise comparison twice with the order swapped, or you are measuring presentation order.
July 25, 2026
ROC AUC looks great on imbalanced data and tells you almost nothing
The false positive rate has the majority class in its denominator, so it barely moves. Precision has the positives, so it moves a lot. Use a PR curve.
July 25, 2026
Route handlers in a Next.js static export have to opt in
An App Router route.ts only survives output: export if it is a GET and it declares force-static. Everything about it is frozen at build time.
July 24, 2026
The Workers Cache API refuses anything with Set-Cookie
cache.put() throws on a response carrying Set-Cookie. Strip the header on a clone, or mark it private in Cache-Control.
July 16, 2026
One stray print breaks an MCP server over stdio
stdout is the JSON-RPC transport, not a log. Anything you print into it corrupts the protocol and the client just says the server disconnected.
June 30, 2026