What a 0.92 Confidence Score Actually Tells YouWhat a 0.92 confidence score means, how to test calibration with 500 past cases, and three patterns for routing uncertain decisions to people.16 min read·karmabox.hashnode.dev3h ago
Weights, Datasets, Keys: Where to Keep Them After a Month of BreachesWeights, Datasets, Keys: Where to Keep Them After a Month of Breaches16 min read·karmabox.hashnode.dev22h ago
*sshole Went From Patch to Exploitation in Four Days: An Incident ReviewTimeline, root cause, blast radius, the read-only checks to run now, and the patch routine we would put on a calendar.17 min read·karmabox.hashnode.dev1d ago
A 284B Model on a 64GB PC Has to Beat $0.00086 a Task. Here's the Mathllama-modchip fits DeepSeek-V4-Flash into 64GB of RAM with no GPU. We priced that box against per-token billing, with every assumption shown.20 min read·karmabox.hashnode.dev3d ago
Why a Line in AGENTS.md Can't Keep Your Email Out of a User-Agent HeaderAn instruction is text the model weighs against other text. A check in the program runs on every request. What this week's Claude Code fix shows.17 min read·karmabox.hashnode.dev2D6d ago
