Hybrid Authorship

The temptation, once an LLM can write competently, is to let it write everything. Or, once a human can write competently with an LLM, to claim full authorship of everything. Both extremes produce bad knowledge systems.

The right model is hybrid: some content is human-authored, some is LLM-compiled, and the split is deliberately marked.

The principle

Every artifact in the KMS has an authorship class:

• Human-authored — written by the user, possibly with LLM assistance, but the user is responsible for every assertion. Examples: principles, frontiers, decisions, opinions.
• LLM-compiled — produced by the system from raw material. The user reviewed and accepted; the LLM did the writing. Examples: entity summaries, meeting digests, weekly check-ins.
• LLM-derived — produced by the system from raw material, not reviewed. Surfaced on demand only. Examples: ad-hoc query responses, draft surfacings.

The class is a property of the artifact and is shown on the page. The user always knows what they are reading.

Why it matters

Trust in a KMS depends on knowing which claims are anchored in human judgment and which are anchored in pattern-matching over your own notes. Conflate the two and the system loses credibility the first time an LLM-derived summary asserts something the user never said.

Mark the split and the user can read confidently — knowing where to challenge, where to trust, and where to expect the LLM to be doing roughly the right thing.

How it works

Three mechanisms:

1. Frontmatter declares authorship. Every page has authorship: human | compiled | derived. Renderers display a small badge; agents respect the class when proposing edits.
2. Compiled content is regeneratable. If a meeting digest is wrong, the user fixes the source and regenerates — they don't hand-edit the digest. This keeps the compilation chain intact.
3. Human content is sacred. Agents propose edits to human-authored content but never apply them automatically. The bar for changing a principle is high; the bar for regenerating a digest is zero.

Where the line sits

The principle is clean. The line — where exactly human-authored ends and LLM-compiled begins — is not. It moves with tier, with topic, with the cost of being wrong, and with my current trust in the model.

The default by tier

Tier	Default authorship	Exceptions
Personal	LLM-compiled	Decisions, opinions, principles → human
Team	Human-promoted	Routine summaries → LLM, marked clearly
App-scoped	LLM-derived but UI-marked	User-edited overrides → human, source preserved

The pattern: the more an artifact gets cited by other people, the further it sits from the LLM-compiled end.

What pushes the line toward human

• Multiple readers will cite it without re-reading sources.
• The cost of a wrong claim is high — legal, financial, customer-trust.
• The content is opinion or judgment, not synthesis.
• The compilation has been wrong in this domain before.

What pushes the line toward LLM

• The artifact is regeneratable — you can re-run with corrected sources.
• The reader is the original author or someone with full source access.
• The cost of being roughly right is high; the cost of being wrong is low.
• The volume is too high for human authorship to keep up.

The tricky middle: promoted compilations

The interesting case is content that starts as LLM-compiled and gets promoted to human-authored after review. Today the marking gets fuzzy in three ways:

• The artifact looks the same after promotion.
• The provenance trail (which model, which version, which sources) often gets lost.
• A reader six months later can't tell if a claim was reviewed or was rolled forward unreviewed.

My current best guess: every promoted compilation should carry a promoted-from-compiled-on: <date> by: <reviewer> field, and the original compiled version should be preserved. Not done consistently anywhere yet.

Where I've been wrong

Two failure modes I've actually hit:

1. A personal-KMS check-in note overwrote a hand-written reflection because the compilation rule didn't distinguish. Lost the original. Fixed by adding authorship: human frontmatter and a hook that blocks overwrites.
2. A team-KMS digest got cited in a client doc as if it were human-authored. The digest was roughly right but had one wrong claim. Caught before sending. The fix was a "this is auto-compiled" badge on the digest and a rule that compiled artifacts can't be copied into client docs without re-review.

Both failures were about the line being unclear, not about the principle being wrong. The principle held; the enforcement didn't.

In Claude primitives

• CLAUDE.md declares the authorship classes and the rules per class — "never edit authorship: human files without explicit user approval."
• Skills are split by class — compile-digest writes compiled content; propose-edit-to-principle writes a suggestion, not a change.
• Hooks (PreToolUse on Edit) enforce the rule at the harness level — block edits to human-authored files unless the conversation explicitly authorised it.

Where I've seen it

This is the youngest of the eight principles. It crystallised after several cases where a personal KMS's compilation chain quietly overwrote something I had hand-written and I lost the original — and after watching an app-scoped KMS struggle to communicate to users which summaries were "real" vs "drafted."

The line where hybrid authorship sits varies sharply across the three tiers: personal can let the LLM compile most things; team has to be much more conservative because compiled content gets cited by other people; app-scoped has to be the most restrictive because a wrong compilation becomes a product bug.

The contrarian read

Marking authorship is overhead, and most users will not look at the badge. Both true. The badge isn't for them — it's for you, the operator, six months later, trying to figure out which assertions in your own KMS came from you and which came from a model.