A developer API for grounding review with curated, sourced AI/ML claims

TL;DR: I just shipped SourceScore VERITAS — a free public API that returns curated AI/ML claim records and their cited evidence. It has 384 records today; use matches as candidates for review. 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. Larger or newer models can still produce unsupported assertions, so evidence retrieval and review remain part of a responsible production workflow.

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 cites primary evidence — 368 of 384 current claims have two or more sources; the official Meta AI blog, the model card on Hugging Face, the arXiv preprint, etc. Not “according to an article on TechCrunch.”
Records carry SourceScore-issued integrity metadata using HMAC-SHA256. The signing secret is not public, so consumers should not describe the metadata as independently verifiable.
Every claim has a stable id — paste it into your LLM context, link to it from a paper, embed it in a prompt template. Re-fetch before use so corrections are not missed.

Why I built it this way

Academic fact-checking datasets, benchmark leaderboards, and Wikipedia solve different parts of the evidence problem. VERITAS adds a small, structured AI/ML catalog with a public JSON interface.

I picked a narrow vertical to start — AI/ML research. The catalog now contains 384 records spanning:

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
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
Organizational facts — Anthropic, OpenAI, Mistral, HuggingFace, Stability AI, DeepMind

Records are curated against their listed sources. Retrieval confidence describes query similarity, not whether an assertion is true. 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.

Catalog expansion is evidence-led: add records only when the source quality and maintenance burden fit the published methodology.

Free tier, no signup

The public API is free, with no auth or signup required. Just curl. Get familiar with the data shape, the signature format, the search behavior. If you outgrow it, paid tiers are proposed higher-volume tiers — not currently available for purchase.

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. hello@caslonmedia.com or join the conversation on the Dev.to cross-post.

Built with: Next.js 15 (static export) · Cloudflare Pages + Pages Functions · TypeScript · Web Crypto API for HMAC. Serverless delivery; measure response time from your own deployment region. 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: sourcescore.org/claims/
Docs: sourcescore.org/docs/
OpenAPI spec: sourcescore.org/api/v1/openapi.json
Pricing: sourcescore.org/pricing/