Hone: An experimental optimizer engine that splits writing from grading

Hone is an experimental optimizer engine that separates the act of producing code from the act of measuring it. You supply a capsule — a task with a mutable file, a protected evaluator, a baseline, and a budget of tokens, dollars, wall-clock time, and evaluator calls. The optimizer proposes edits inside a sandbox; the broker runs the trusted evaluator in a pinned image and keeps only candidates that beat the incumbent on measured score. The separation is deliberate: the thing writing the code cannot grade its own work. This is research code, not a packaged tool. The engine lives here; the capsules live in a companion repository. Evidence so far is narrow. One result stands out. A fixed-compute campaign scored a port of a standard-library Leduc hold-em CFR+ solver by exploitability. The first run found an improvement but violated the time budget and was rejected; a second run reweighted the average policy at low iteration counts and achieved parity at runtime while raising validation q from 0.8453 to 0.8599 and the sealed holdout from 0.7733 to 0.7899. That is a real but small win, and it came after an honest failure. What makes this worth noting as engineering material: the budget discipline is concrete and enforceable, not aspirational. Every