document project llm practices

README.md

@@ -2,6 +2,32 @@
 
 A friendly, collaborative PostgreSQL derivative for nerds of all stripes.
 
+## LLM Code Policy
+
+Large language model code generation is permitted sparingly in very limited
+cases, for example for completing clearly defined transformations which span
+multiple files and are not supported by conventional code actions. All code
+generated by LLMs must be thoroughly and frequently reviewed by the author,
+before committing affected work.
+
+As of this writing, models display a strong bias towards patterns which are well
+represented in public open source projects. This can cause them to tend towards
+suboptimal one-size-fits-most or simply outdated coding practices in certain
+circumstances. LLM assistance should be sufficiently constrained to avoid
+allowing outputs to dictate or implicitly guide significant design decisions.
+
+Furthermore, current language models broadly behave adversarily, in the sense
+that they are optimized to make perceiving model outputs versus non-model
+outputs as difficult as possible. This can make generated code uniquely
+challenging to review effectively. In this context, non-trivial business logic,
+particularly logic with security implications, may not be implemented with
+direct assistance from LLM tools.
+
+Examples of LLM-assisted changes in practice:
+
+- Replacing SVG icons with similar webfont icons from a different icon pack.
+  (Revision `ztrnxzqv` (Git `a8dd49f7`))
+
 ## The Phonograph Authorization Model
 
 Postgres provides a sophisticated role based access control (RBAC) system, which