Known limits

Semantic contradiction is still an agent judgment

Rationale can prove that an approved Record governs the target and can report lexical overlap and conservative polarity hints. It does not use a remote LLM to decide meaning. The agent must pronounce itself on an apparent conflict; undetermined is never promoted to a blocking verdict.

Provider coverage can be partial

Codebase Memory is optional and fallible. A stale or unavailable provider can leave structural bindings unresolved. The packet exposes coverage and warnings so a human can calibrate trust.

The project is pre-1.0

The current public evidence is alpha. Platform packaging, pilot gates, and provider behavior on unusual repositories still need broader dogfood. The documentation describes tested behavior, not a promise that every language or workspace layout is indexed equally.

Not a hosted approval system

Rationale is local-first. It has no required account, remote canon, SaaS workflow, or automatic approval service. Human authority remains in the project’s versioned lifecycle.