Transparency
Methodology
How CheckingDog reaches an answer — and why we show our work.
Four models, in parallel
Every question or claim goes to four frontier models at once — Claude, GPT, Gemini, DeepSeek. Each answers independently, in its own voice. No single model gets to decide.
Agreement is the signal
We measure how much the models agree. High agreement means higher confidence; disagreement isn’t hidden behind one confident-sounding answer — we surface the split, because that’s exactly where you should stay skeptical. For fact-checks, a verdict needs majority agreement, otherwise it’s marked contested.
Reads context, cites sources
Models read the images and links you paste, and search recent discussion for current events instead of guessing from training data. Every fact-check lists the sources each model cited — so you can check the checkers.
We’d rather say “unverifiable”
If the models can’t establish something — or couldn’t read a link — they say so, instead of inventing a confident answer. Honest beats clever.
We show our work
Every check has a public page showing each model’s full answer, confidence, and cited sources — not just the final verdict. The exact prompts we send the models are open and auditable, so anyone can see how we ask.
Limits
Cross-checking four models makes “all wrong at once” rare — but not impossible. CheckingDog is a tool for thinking, not a final authority, and it isn’t legal, medical, or financial advice.