civitai__civitai/docs/seo-ecosystem-fact-check.md

Ecosystem SEO pages — human fact-check list

Purpose: the landing pages mix (a) live data pulled from the DB and (b) unique prose that was AI-authored, grounded in each model's Civitai model card + the orchestrator prompt-enhancement guide. (b) reads as authoritative but was not written by a human — verify it before we treat it as fact. This lists what to check and where the known risks are.

Source of truth per field: config is src/shared/constants/ecosystem-seo.constants.ts (the ECOSYSTEM_SEO map). Live data is computed in src/server/services/ecosystem-seo.service.ts.

How to review (in-context, per page)

Each config carries an optional factCheck array — the specific AI-authored claims to verify. When a moderator views a page (/ecosystems/<slug>), a floating "⚠ Fact-check" panel lists that ecosystem's flags (field → claim → why), with links that jump to the section. The flags are stripped from the payload for non-moderators, so end users never see them. To clear a flag: verify the claim against a source, fix the copy if needed, then delete that entry from the config's factCheck array. This doc is the standing methodology; the per-config factCheck entries are the live worklist.

---

Tier 0 — machine-pulled, low fact-check risk (verify scoping, not facts)

These come straight from the DB and refresh daily; they can't be "wrong" in the factual sense, but confirm they're scoped to the right ecosystem (no family bleed) and SFW.

• Stat counts — models / generations / LoRAs (hero row). Engine ecosystems fall back to a media-created count (ModelMetric has none). Check the numbers look sane per page.
• Featured model names / types, popular LoRAs list, download & generation counts on cards.
• Example media — the image/video itself, and its prompt + settings. Settings are built from real image meta; the prompt caption is the real meta prompt (sometimes lightly trimmed — see flags).

Tier 1 — AI-authored prose, MUST be verified (highest priority)

Per ecosystem, these fields are unique prose generated from the model card + prompt guide:

• hero.intro — the one-paragraph pitch. Check provider, capability claims, "open/closed".
• overview (3 paragraphs) — the densest factual surface: architecture, parameter counts, text-encoder, native resolution, release lineage, provider. Verify every number and proper noun.
• comparison.rows — ⚠️ the qualitative ratings ("Excellent", "Very good", "Strong", winner highlights) are editorial judgment, NOT sourced metrics. Verify or soften before treating as fact. The factual cells (provider, open/closed, "Available on Civitai") should be correct — spot-check.
• faq — answers are AI-authored. Check any factual claim (limits, modes, pricing framing).
• promptTips — grounded in the orchestrator guide where a real one existed (usually reliable); where the guide was a generic fallback, these came from the model card + general best practice (flagged).
• attribution (footer) and metaDescription (SERP snippet) — check the provider/claim.
• hero.badges — short claims ("Open weights", "By X"); verify.

Tier 2 — dates & flags (once added)

• releasedAt (release date) and any "new" flag — human-set; confirm accuracy.

---

Known per-ecosystem flags (from the authoring pass — resolve these first)

Ecosystem	Flag to verify
HappyHorse	attribution says "attributed to Alibaba" — corporate parent is unconfirmed (guide said "Alibaba, via fal.ai"; Civitai groups it under an "Alibaba – Taotian" family). Confirm the real owner. Also: the model card's "#1 on the Artificial Analysis Video Arena / Elo 1416" claim was deliberately excluded as unverified — decide if we want it.
Grok Imagine	The real orchestrator guide (grok) is image-oriented (Aurora model); there is no Grok video prompt guide, so the video/motion prompt tip is general best practice, not sourced. Page is framed as a video ecosystem but the model does image + video (intentional).
Kling	settings omit fps (null in meta) though the guide states ~30fps. One example (imageId 133284218) reports 8s, outside the guide's stated 5s/10s modes — it's the real meta value.
Seedance	Example prompt captions are condensed from very long real prompts (Remix still uses the real image meta, so it works — captions are display-only). No official ByteDance/Seedance ToS URL was found; attribution kept generic. Comparison ratings are editorial.
Pony / Illustrious / NoobAI	The orchestrator prompt guide returned a generic fallback (no model-specific rules) for these keys, so promptTips were grounded in the model card + general SDXL/booru practice instead. Verify the tips.
Anima	Model card vs. prompt guide conflicted on weight syntax (card says weights work; guide says they don't). The guide was followed. Confirm which is right.
All numeric overview claims	Spot-check specifics: HiDream "17B sparse MoE", Z-Image "~6B", SD 1.5 "512-native / 77-token CLIP", Seedance "4–15s, 480p/720p, up to 9 image refs", Grok "~1,000 char prompt", Flux "T5-XXL / 256–512 tokens".

---

Open-question answers this addresses

• OQ1 (fact-check scope): the fields above (Tier 1) are the AI-authored ones threaded from model pages / guides; Tier 0 is live DB data. Start with comparison ratings and overview numbers.
• These are per-page — worst case is one wrong number on one ecosystem, not a systemic error, because each page was grounded from that model's own card.