# PDR 001: Prove the editor in the repository before generalizing

- HTML version: https://robbiepalmer.me/projects/agent-first-writing/pdrs/001-repository-first-evidence-gated-release
- Project: Agent-first Writing Editor
- Status: Proposed
- Opened: 2026-09-04

## Context

The product thesis covers research sessions, plans, decision records, and
drafts produced by several models. A general editor would require early guesses
about author preferences, evaluation, model routing, privacy, and integration
with writing tools.

The Personal Knowledge Graph already supplies real drafts, revision history,
deterministic prose rules, published artifacts, and an active review loop. It
can show whether the editor improves work the author intended to publish before
the project takes on a general product boundary.

## Proposed decision

Use content in this repository as the first product scope. Add capabilities in
evidence-gated layers:

1. Define versioned records for findings, suggestions, grouped proposals, and
   editorial decisions.
2. Build a reproducible evaluation set from repository drafts and published
   revisions.
3. Run deterministic checks through the shared record contract.
4. Evaluate lightweight models only for failure classes that the deterministic
   rules miss.
5. Add passage rewrites after local suggestions are reliable.
6. Capture accept, reject, and change decisions during ordinary use.

Do not extract a general editor until the repository-specific system improves
published pages, preserves facts and voice, and reduces manual rewriting. A
new model or rewrite path remains review-only until its declared evidence gate
passes.

## Consequences

The first release may use repository conventions and existing revision
history. That shortens the path to real evidence and keeps a failed experiment
from becoming a public product contract.

The implementation may accumulate assumptions that a later general product
must remove. Generalization is delayed on purpose. Passing internal technical
checks is not enough to accept this proposal; the publication pilot must show a
better outcome for authors and readers.

## Evaluation

Accept this proposal only when the preregistered repository publication pilot
has complete evidence. The edited artifacts must preserve facts, intent,
terminology, and voice, reduce manual review work, and earn reader preference
without hiding regressions in an aggregate score.

## Migration note

This PDR replaces the product-only record formerly stored as ADR 005. It keeps
the original date, Proposed status, rollout sequence, and evidence gate. The
migration changes the record type, not the decision.

## Related records

- Project: [Agent-first Writing Editor](https://robbiepalmer.me/projects/agent-first-writing.md)

---

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