Compiling Expertise, Turning tribal knowledge into auditable local agents

Abstract

You've seen the talks. Someone wires a model into a loop, feeds it a domain in a system prompt, and demos an agent that triages your findings. It works on stage. Months later it's handing out confident, unauditable answers nobody can trace, and it drifts when the model updates.

The model was never the hard part. The hard part is the step we stopped doing: turning what your experts know into something a machine runs the same way every time. We prompt when we should "compile".

This poster presents a five-stage method for compiling tacit expertise into a constrained local agent: deterministic rules and math for what compiles, the model for the irreducible residue, tools and schemas to constrain even that — and a closed loop so corrections turn back into rules.

The method is proven in production: SAST triage on a local LLM, clearing 67% of the noise from 3,993 mined expert decisions before a model ever runs. The findings are SAST. The method isn't.