The Library

The Library is the corpus the engine builds from — every approved reference and template for the install surfaces (ads, lead magnets, landing / presales / qualification / conversion pages, sequences, decks). Lives in Supabase (library table). Surfaced visually at /studio/library and as JSON via /api/library (the same data the AI reads).

It is not inspiration. It is the picked-from set: at LOFI the AI shortlists from here by fit, the POC picks one of three, and HIFI renders that pick by tokenize-and-substitute.

Two axes that matter

Axis	Values	Meaning
Origin	`Original` · `Curated`	`Original` = 0HR's own work (`source='own_code'`). `Curated` = pulled/curated from external + historical sources.
Proof	`Proven` · `Reference`	`Proven` = real DATA + PROOF. `Reference` = everything else (no exceptions).

PROVEN = DATA + PROOF, no exceptions

A row is Proven only if it has:

Explicit metrics — real numbers (claimed_metrics: spend, CPL, leads, conversations), or
FB-Ad-Library proxy proof — the same ad running ≥ multiple months + an established/large advertiser + high creative volume (dozens–hundreds of variants) + stated spend ("we spend $1M/mo") or conservatively estimated spend (e.g. $20M/yr revenue → ~10% to marketing → ~$2M/yr ad budget).

Editorial / AI quality ratings do not count — that's inspiration, not proof. Anything not proven is a Reference. (null > fabrication.)

Canonical types only

The Library carries the deliverable/template types the engine produces — nothing else. Internal docs and noise (text_doc, pdf_doc, brand-os-section, negative-example, post, page) are never part of the Library.

Format	Surface
`ad_creative`	Lead-gen ad
`lead_magnet`	Lead magnet
`landing_page`	Landing page
`presales_page`	Presales page
`qualification_page`	Qualifier
`conversion_page`	Conversion / booking
`email` · `email_sequence`	NBO / nurture
`copy_pack`	Profiles · scripts
`case_study` · `slide_deck` · `video` · `social_post` · `chatbot_flow`	Supporting assets

The picker (LOFI)

Stupid-simple, deterministic, instant — no AI guess at eligibility:

Filter — is_public AND approved AND component_slugs ∋ '<surface>' AND tier ≤ brand.tier.
Match — vertical_slugs ∩ brand.vertical OR business_model_slugs ∩ brand.business_model. A homecare brand cannot be offered an agency template — the tags forbid it.
Rank proven-first, return the top 3. (If a bucket has < 3, relax the vertical filter to fill — flagged as proven-fill.)
The POC picks one; the pick is stored at brand.config.engine.selected_templates[<surface>]; HIFI renders it.

Endpoint	Does
`GET /api/library`	Filterable list + facets (format · industry · stage · origin)
`GET /api/library/shortlist?brand=&surface=`	The LOFI picker — deterministic top-3
`POST /api/library/select`	Store the POC's pick → `engine.selected_templates`

The moat — a self-improving proven library

This is the key result that separates 0HR from everyone else.

Every brand's Command Center auto-ingests real campaign results (spend · CPL · leads · conversations) back into the Library as new Proven entries. That produces a continuously-growing, unforgeable corpus of proven, real-results templates. The AI shortlists from it — so every install gets better as the network grows:

more users → more real campaign data → more proven templates → better shortlists → better installs → more users

Competitors cannot fabricate real proven results at scale → a compounding data moat. A weekly newsletter features and gives kudos to top-performing user campaigns ("check out this campaign from user 10") — gamifying contribution and driving network effects.

See also: Templates (render mechanism) · Components (the install deliverables) · Workflows.