local free first

The bottom rung of the ladder is free: a capable open-weight coding model served locally on my own machine handles the mechanical legs — grep-and-report, file inventories, format conversions, boilerplate scaffolds, rubric checks — for zero cost.

Mechanical legs — greps, format conversions, inventories — don't need frontier reasoning, but routed to paid models they quietly become the biggest line item. A local model does them for the cost of electricity.

The leak test runs before every paid delegation: could the local model do this leg? If yes and a paid agent gets it anyway, that's the leak. What stays paid: cross-file refactors, ambiguous root-cause work, final synthesis, and the verify loop — a small local model is the bottom rung, not a substitute for judgment.

The whole method is downloadable and adaptable — rules/local-free-first.md inside the zip. Serve any strong open-weight coder locally with a context-capped variant, and route your smallest legs there first.