Frequently Asked Questions

CiteGuardian is a service that verifies whether AI-generated citations actually support the claims being made. It catches "citation laundering" — citations that look legitimate but don't actually back up the claim — using an evidence-scrubbing technique that compares model confidence with and without the cited evidence.

Citation laundering is when an AI model includes citations that appear credible but don't actually support the claims they're attached to. The citation makes the response look well-sourced, but if you read the source material, it either says something different or doesn't address the claim at all. CiteGuardian's scrub test is specifically designed to detect this pattern.

Every new account receives 5 free credits on signup. That's enough to run 5 verifications and see how CiteGuardian works before purchasing more.

Verifications cost 1–4 credits based on the number of claims extracted: 1–5 claims = 1 credit, 6–10 = 2 credits, 11–20 = 3 credits, 21+ = 4 credits. If a request fails due to invalid input, credits are automatically refunded.

Your 5 signup credits never expire. Plan credits included with monthly subscriptions (Starter and Pro) reset each billing cycle — unused plan credits are lost on renewal.

CiteGuardian works with output from any AI tool — ChatGPT, Claude, Gemini, Copilot, Perplexity, or any other model that produces text with citations or source references. You paste the AI-generated text and the sources it cites, and CiteGuardian checks whether the citations hold up.

A decorative citation is one that looks legitimate but doesn't actually influence the verification verdict. CiteGuardian detects these using the scrub test: it runs the verifier twice — once with the cited evidence and once without. If the confidence doesn't change meaningfully (delta < 0.10), the citation is flagged as decorative. This is the hallmark of citation laundering.

A contradiction is when the source evidence actively disproves the claim. This is worse than "unsupported" — it means the cited source says the opposite of what the AI claimed. Contradictions are always highlighted in reports and will cause a gate-mode check to fail.

The scrub test (evidence scrubbing) is CiteGuardian's core technique. For each claim, it asks the verifier to judge the claim twice: once with the cited evidence (producing confidence p1) and once without it (producing confidence p0). The difference (p1 - p0) tells you how much the evidence actually mattered. A large delta means the citation is meaningful; a small delta means it's decorative.

The pass/fail gate is an API mode designed for RAG pipelines. Instead of returning a full verification report, it returns a simple pass or fail verdict. Add "gate": true to your /api/verify request. A response passes when the supported rate meets the threshold (default 70%) and no contradictions are found. This lets you block hallucinated responses before they reach the end user. See the API reference for details.

Log in to your account and go to the Dashboard. In the API Keys section, click "Create API Key" and give it a name. The key (prefixed with cg_) is shown once — copy and save it securely. Use it in the Authorization: Bearer cg_... header for API requests.

Yes. Rate limits are per-tier: Free 3 verifications/min, Starter 10/min, Pro 20/min. Batch requests are capped at 20 items per batch (Pro only). If you hit a limit, the response includes a Retry-After header.

Your verification requests (answer text, sources, and reports) are stored in a PostgreSQL database tied to your account. API keys are stored as SHA-256 hashes — CiteGuardian never stores your raw API key after it's created. Passwords are hashed with bcrypt. Source URLs are fetched server-side to extract content for verification.

The maximum batch size is 50 items per request. If you need to verify more items, split them across multiple batch requests.

The extension isn't on the Chrome Web Store yet. You can install it manually: clone the repository, go to chrome://extensions, enable Developer Mode, click "Load unpacked", and select the extension/ folder. See the full Chrome Extension guide for step-by-step instructions.

Yes. CiteGuardian is a standard Flask application backed by PostgreSQL. You need an Anthropic API key for the verification model. Clone the repository, set up the .env file with your configuration, initialize the database, and run the app. The Chrome extension works with any CiteGuardian server — just change the server URL in the extension settings.

CiteGuardian is designed for transparency, not black-box authority. Every verdict comes with the evidence that produced it:

• Evidence + scores + reasoning — each claim shows the source text, confidence score, alignment dimensions, and a plain-language explanation of why the verdict was given. You can verify any call yourself in seconds.
• Multiple independent layers — claims pass through an LLM judge, a deterministic decision engine (policy rules, structured scores), and an optional scrub test. These layers cross-check each other.
• Confidence calibration — low-confidence verdicts are visually flagged in the report UI so you know when the system is uncertain.
• Not a black box — alignment scores across 8 dimensions (topic, entity, temporal, numeric, attribution, proposition, directionality, modality) and any policy override reasons are fully inspectable.
• First-pass filter — CiteGuardian catches weak evidence at scale. It is not a replacement for human judgment on high-stakes decisions. When it's uncertain, it says so.