Problem

A production brain with 186K pages has organically accumulated 94 distinct page types. Most are duplicates, near-duplicates, or one-off types that should be subtypes via frontmatter fields. This violates DRY (same concept under multiple types) and MECE (types overlap, creating ambiguity about where new content goes).

The type system is the foundation for schema packs, search filtering, extract behavior, enrichment routing, and expert_routing. When types are noisy, every downstream feature degrades.

Production Data

Cluster 1: Social posts — 7 types, 106K pages

Should be: tweet with subtype: single|bundle|stub + social-digest with period: daily|monthly. 7 → 2.

Cluster 2: Articles — 5 types, 3.6K pages

Should be: ONE type article. 5 → 1.

Cluster 3: Companies — 4 types, 13.5K pages

"Accelerator membership" should be a field (batch: S23), not a separate type. Should be: company with kind: company|product|org. 4 → 1.

Cluster 4: Atoms — 6 types, 18K pages

atom-partner-link and partner-atom-link should be LINKS, not page types. Should be: atom with origin: extraction|manual|lore. 6 → 1.

Cluster 5: Media/Content — 8 types, 8.7K pages

Content format is a frontmatter field, not a type. Should be: media with format: video|article|essay|book|podcast. 8 → 1.

Cluster 6: Analysis — 8 types, 40 pages(!)

8 types for 40 pages is the clearest sign of ad-hoc proliferation. Should be: analysis with domain field. 8 → 1.

Cluster 7: Concept redirects outnumber concepts

• concept: 4,304
• concept-redirect: 5,519

More redirect pages than real pages. Redirects should be an alias table, not 5.5K stub pages that inflate orphan counts and waste embedding tokens.

Cluster 8: One-off types — 25+ types with 1-2 pages each

civic, framework, insight, anecdote, principle, memo, rfs-draft, pitch-deck, policy-criticism, production-doc, recording-snippet, registry, reference, schema, video-script, web_page, log, agent-log, content-mining, meta-prompt, queue, eval-test...

Should be: tags or subtypes of note.

Cluster 9: Symlinks as pages — 54 pages

symlink, partner-symlink, symlink-manifest — filesystem operations stored as brain pages.

Why This Matters

1. Schema packs can't be MECE — the pack declares types but the brain has 94, many undeclared. schema_review_orphans can't distinguish intentional from noise.
2. Search filtering is ambiguous — --type article misses 2.2K articles typed as media/article, sources/article, etc.
3. Enrichment routing is incomplete — enrichable_types can only list a few. 80+ types means most pages never get enriched.
4. Agent confusion — when ingesting a new article, should it be article, media/article, sources/article, or source/article?
5. Orphan inflation — concept-redirect pages (5.5K) inflate orphan count without adding knowledge value.

Proposed Target Taxonomy

94 types → 14 types. Distinctions move to frontmatter fields (subtype, format, origin, period, domain).

Migration Path

1. Schema pack declares the 14 canonical types with aliases covering the old names
2. Migration script retypes pages (e.g., media/article → article with source_collection: media)
3. concept-redirect pages → alias table entries + soft-delete
4. symlink/atom-partner-link pages → proper link table entries + soft-delete
5. One-off types → retype to note with original type preserved as legacy_type tag

Impact on Existing Features

• inferType path prefix mapping shrinks dramatically
• Schema pack page_types goes from 30+ to 14 entries
• enrichable_types covers more of the brain naturally
• extract type filters work correctly across the whole corpus
• find_experts expert_routing covers all entity types cleanly
• schema_review_orphans becomes meaningful (currently noisy)
• Agent ingestion becomes unambiguous (one type per domain)

Related

• feat: gbrain onboard — guided agent onboarding with migration prompts #1383 (gbrain onboard — migration prompts would drive this)
• docs: consolidated gbrain onboard design (replaces #1378-#1382) #1409 (consolidated design doc)