The Local AI Settings Card
The frozen, measured serving config for a 27B model on a $1,400 Strix Halo mini-PC: weights, KV cache, context, speculation, thinking, vision — with the benchmark numbers and raw logs behind every dial. Printable.
Free Resources
Useful things, built from measured work.
Four downloadable resources, each derived from benchmarks, blogs, and repos we published first. Useful standalone — read them, print them, steal the designs. If they help, Kyanite Build Notes carries the updates.
The Delegation Decision Worksheet
Which of your jobs can you safely hand to AI? Score your tasks against delegation-bench’s measured floors — 9 job classes, 965 trials, a sabotage cell — and build your own delegation card.
The Eval Design Checklist
Seven rules for building an AI eval that survives attack: pre-registration, hidden tests, floor statistics, the sabotage cell, calibrated dual judges, labeled preliminary data, published failures. Every rule links to its proof.
The LLM Comparison Cheat Sheet
The Pareto frontier drawn, what intelligence scores mean practically, real cost bands from the live dataset, and the five questions that pick a model for your task instead of the leaderboard’s.