# I mark what generated text does not own.

How explicit ownership, selective regeneration, and provenance let people and machines maintain the same artefact safely.

- **Canonical URL:** https://vidhata.me/notes/i-mark-what-generated-text-does-not-own
- **Author:** Vidhatanand V. (Vid)
- **Role:** Fractional CTO and AI Systems Architect

**Published:** 04 Sep 2026 | **Reading time:** 9 minute read

## Why I hold this position

Generated documentation becomes destructive when the generator assumes it owns the whole file. A person adds context, records a decision, explains an exception, or corrects the machine. The next run replaces the document from source evidence and quietly removes the part that mattered most.

I treat shared artefacts as an ownership problem. The machine needs a visible boundary around what it may regenerate. Human-maintained material needs an equally visible place that survives automation. Without those boundaries, every update forces the team to choose between stale documentation and erased judgement.

## Ownership must be visible inside the artefact.

I prefer explicit generated and manual regions over a convention stored only in a prompt. The file itself should say which sections are machine-maintained, which sections are human-maintained, and which metadata identifies the source, generation time, review state, and responsible system.

The boundary should be simple enough for ordinary tools to preserve. A generator can replace a marked section while leaving the rest byte-for-byte intact. A reviewer can see whether a paragraph came from source analysis or a human decision. A merge conflict becomes an ownership question the team can answer, not a mystery produced by another full rewrite.

## I regenerate only what the evidence says changed.

Whole-estate regeneration creates noise. I would rather map code and configuration deltas to the documents and sections they can affect. A changed API contract may require an endpoint reference and one flow diagram, not a rewrite of the project overview, onboarding guide, and every architecture page.

Selective regeneration also improves review. The pull request contains a small, attributable change. The reviewer can compare the source delta with the documentation delta and ask whether the impact map missed anything. The generator becomes a participant in change control rather than a periodic publishing machine.

## Human edits need provenance, not invisible privilege.

I do not preserve manual text merely because a person wrote it. Human content can also become wrong. It should carry an owner, review state, and relationship to the source evidence. The difference is that automation should not erase it without an explicit decision.

Generated material deserves the same discipline. It should identify the repositories, files, schemas, or changes that supported it. When the evidence disappears or conflicts, the section can be marked stale or sent for review instead of confidently generating a replacement from incomplete context.

## Validation protects the shared boundary.

After generation, I validate structure, links, diagrams, required metadata, and the presence of manual markers. I also exclude irrelevant files and redact likely secrets before model context is constructed. A documentation tool should not improve freshness by expanding the disclosure surface of the codebase.

The goal is not to freeze human prose or make machine output untouchable. It is to create a durable collaboration model. The machine updates the material it owns, people keep the judgement they own, and both leave enough provenance for the next change to be reviewed honestly.

## What I take into the next build

- Mark generated and human-owned regions inside the artefact.
- Regenerate affected sections instead of rewriting the estate.
- Give both human and generated content provenance and review state.
- Validate structure, links, ownership markers, and prompt safety after every update.

## Continue reading

[All Fractional CTO field notes](https://vidhata.me/notes)

## Work with Vid

Start with the actual technical pressure: [bring the problem](https://vidhata.me/hire).