# ADR 005: Repository-first, evidence-gated release

- HTML version: https://robbiepalmer.me/projects/agent-first-writing/adrs/005-repository-first-evidence-gated-release
- Project: Agent-first Writing Editor (https://robbiepalmer.me/projects/agent-first-writing.md)
- Status: Proposed
- Date: 2026-09-04
- Initiatives: Semi-autonomous Software Development (https://robbiepalmer.me/initiatives/semi-autonomous-software-development.md)

# Context

The product thesis covers much larger inputs than one repository, but a general
editor would require early guesses about style, evaluation, and model routing.
The Personal Site already has raw drafts, Git history, Vale rules, published
artifacts, and a real review loop. It can expose whether the system saves work
before the project takes on a general product boundary.

# Decision

Make repository content the first release. Introduce capabilities only after
the preceding layer works on real pages:

1. Define the suggestion contract and evaluation set.
2. Run Vale and other deterministic checks through that contract.
3. Add a lightweight model only for failure classes the evaluation set proves
   the rules cannot handle.
4. Add passage rewrites with diffs after local suggestions are reliable.
5. Capture accepted, rejected, and changed decisions against each stable
   proposal ID and its ordered member suggestion IDs during ordinary use.
6. Extract a general editor only after the repository-specific system improves
   published pages and reduces manual rewriting.

# Consequences

The first release can use local conventions and existing revision history. Each
new model needs evidence that it fixes a named gap without damaging facts,
intent, terminology, or voice.

This sequence delays a general-purpose interface and may expose repository
assumptions that later need removal. That work is preferable to generalizing a
workflow before it has produced writing the author wants to publish.

---

Markdown index of this site: https://robbiepalmer.me/llms.txt