# PDR 000: Keep authors in control of editing and publication

- HTML version: https://robbiepalmer.me/projects/agent-first-writing/pdrs/000-author-controlled-editing-and-publication
- Project: Agent-first Writing Editor
- Status: Accepted
- Opened: 2026-09-04
- Decided: 2026-09-04

## Context

Agent output can contain strong research and weak writing. An editor can reduce
repetition, repair local faults, and restructure a passage, but those changes
carry different risks. A punctuation fix is easier to verify than a rewrite
that may alter facts, intent, terminology, or voice.

Publishing raises the stakes again. A useful suggestion does not give the
system authority to decide what is true or to publish under somebody's name.
The author remains accountable for the final artifact.

## Decision

Give authors control that matches the scope and risk of each editorial change.

The editor may report an exact finding, propose a local replacement, or group
dependent edits into a passage rewrite. It must show enough source context and
provenance for the author to understand the change. Local suggestions and
passage rewrites enter an explicit accept, reject, or change workflow.

Automatic application is limited to narrow rules whose evidence shows that the
change is safe. Publication under a person's name still requires that person's
final review and approval. The editor may retrieve supporting context, but it
does not replace research, fact checking, or editorial judgment.

The product will preserve the author's facts, intent, terminology, and voice.
It will not optimize for generic polish or train every author toward one style.

## Consequences

The review interface needs more than a list of model outputs. It must separate
findings from actionable suggestions, show dependent changes together, retain
the author's decision, and stop when source text or evidence is stale.

This boundary limits unattended rewriting. Even a strong model cannot publish
on confidence alone. The extra review is deliberate because the artifact
carries the author's name and judgment.

## Evaluation

The policy is working when authors can explain each proposed change, reject or
revise it without losing the original, and publish only after checking the
complete artifact. Evaluation must record factual, terminology, intent, and
voice regressions as failures even when a rewrite sounds smoother.

## Migration note

This record preserves the author-control policy published in the original
project narrative and the product constraints that were mixed into ADR 006.
Moving the policy into a PDR does not change it. ADR 006 retains the separate
architecture choice to use knowledge-graph context.

## Related records

- Project: [Agent-first Writing Editor](https://robbiepalmer.me/projects/agent-first-writing.md)
- Implemented by ADR: [ADR 006: Knowledge graph context with bounded authority](https://robbiepalmer.me/projects/agent-first-writing/adrs/006-knowledge-graph-context-with-bounded-authority.md)
- Implemented by ADR: [ADR 007: Playwright-assisted Grammarly handoff](https://robbiepalmer.me/projects/agent-first-writing/adrs/007-playwright-assisted-grammarly-handoff.md)

---

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