Stop hallucinating: a developer API for grounding LLM responses with signed, sourced claims

paulo de vries — Sat, 16 May 2026 13:00:59 +0000

TL;DR: I just shipped SourceScore VERITAS — a free-tier-friendly API that returns hand-verified AI/ML claims with their primary sources, an HMAC-SHA256 signature, and a ready-to-paste citation. 51 claims at launch; expanding to 5,000+ this year. curl https://sourcescore.org/api/v1/claims.json and you're in.

If you've built anything on top of an LLM in the last two years, you've watched it confidently invent facts that don't exist. You've seen GPT-4 cite papers that were never written. You've watched Claude give the wrong release date for a model that came out last month. You've fixed RAG pipelines where the retriever pulled the right document but the model still produced a number nobody can find anywhere on the source page.

The grounding problem isn't going away. It's the hardest unsolved problem in production AI today, and the bigger your model gets, the more confidently it lies when it lies.

Last week I shipped VERITAS — a developer API I wish existed two years ago.

What it does (in one curl)
curl -X POST https://sourcescore.org/api/v1/verify \
-H 'Content-Type: application/json' \
-d '{"claim": "Llama 3.1 was released in July 2024"}'
{
"apiVersion": "v1",
"query": "Llama 3.1 was released in July 2024",
"bestMatch": {
"id": "...",
"subject": "Llama 3.1",
"predicate": "released_on",
"object": "2024-07-23",
"statement": "Llama 3.1 released on: 2024-07-23.",
"confidence": 1.0,
"detailUrl": "https://sourcescore.org/api/v1/claims/....json"
},
"signature": {
"algorithm": "HMAC-SHA256",
"signedBy": "did:web:sourcescore.org",
"signedAt": "2026-05-16T...",
"signature": "..."
}
}
Three things make this useful for grounding LLMs:

Every claim has 2+ primary sources — the official Meta AI blog, the model card on Hugging Face, the arXiv preprint, etc. Not "according to an article on TechCrunch."
Every response is signed — HMAC-SHA256 with did:web:sourcescore.org. Your client can prove the answer came from SourceScore and wasn't tampered in transit.
Every claim has a stable id — paste it into your LLM context, link to it from a paper, embed it in a prompt template. It won't move.
Why I built it this way
There are great academic fact-checking datasets. There are great benchmark leaderboards. There's Wikipedia. None of them are an API you can call from your RAG pipeline at request time with a 30ms response.

I picked a narrow vertical to start — AI/ML research. 51 claims at launch covering:

12 foundational papers (Transformer, RLHF, RAG, LoRA, DPO, Chinchilla, PPO, Adam, AlexNet, BERT, Chain-of-Thought, FlashAttention, MoE, Switch Transformer, Mamba, T5, CLIP, Constitutional AI, InstructGPT, ResNet)
22 model releases with dates, parameter counts, context windows (GPT-2/3/4/4-Turbo/4o, Claude 3/3.5, Llama 1/2/3/3.1, Mistral 7B, Mixtral 8x7B, Gemini Pro/1.5, Whisper, DALL-E 3, Stable Diffusion 1, Sora, ChatGPT, ChatGPT Plus)
6 organizational facts (Anthropic, OpenAI, Mistral, HuggingFace, Stability AI, DeepMind)
Every claim is hand-verified against the primary source. If a claim is below 0.85 confidence, it's not published. Performance-comparison claims are intentionally excluded for v0 because benchmark numbers depend on version + prompt format — too much surface for "actually that's not quite right" pushback.

The plan is to grow the catalog to ~500 claims by Day 30 and ~5,000 by Year 1, all under the same methodology.

Free tier, no signup
The free tier is 1,000 claims/month, no auth required. Just curl. Get familiar with the data shape, the signature format, the search behavior. If you outgrow it, paid tiers are €19 (Indie) / €99 (Startup) / €499 (Scale) — Stripe metered billing.

OpenAPI 3.1 spec at /api/v1/openapi.json. Full docs at /docs/.

What's next
I've got two open questions I'd love feedback on:

What claim types are most valuable? Right now I'm at release-dates + parameter-counts + paper-introductions + organizational-facts. Operator-suggested adds welcomed.
Vertical expansion direction. AI/ML is the v0 wedge. Next likely candidates: scientific instrumentation specs, software release dates + versions, regulatory deadlines. What would you actually use?
Try it; break it; tell me what's missing. contact@sourcescore.org or comment below.

Built with: Next.js 15 (static export) · Cloudflare Pages + Pages Functions · TypeScript · Web Crypto API for HMAC · Plausible Analytics. 100% serverless, ~100ms cold-start globally. Source-rating product (the original SourceScore Index, 130 hand-scored sources) lives alongside at the same domain — both products under one methodology.

Links to bookmark:

Catalog: https://sourcescore.org/claims/
Docs: https://sourcescore.org/docs/
OpenAPI spec: https://sourcescore.org/api/v1/openapi.json
Pricing: https://sourcescore.org/pricing/

DEV Community: paulo de vries

Stop hallucinating: a developer API for grounding LLM responses with signed, sourced claims