PREDICT
Generate the next likely token from patterns learned in training.
Evidence for consequential AI use
AI sounds sure.Evidence decides.
Kelreva checks factual claims in AI chat responses against retrieved web evidence—before an answer becomes a decision.
CHROME EXTENSIONCHATGPTCLAUDEGEMINI
01 / THE STRUCTURAL GAP
The model is trained to continue.
You need it to be correct.
Language models produce likely continuations from learned patterns. Fluency is native. An independent source-of-truth check is not.
COMPOSE
Turn probabilities into one fluent, coherent response.
VERIFY
Not automatic. Independent evidence checking is a separate system.
Research: TruthfulQA examines how models reproduce common falsehoods; Why Language Models Hallucinate analyzes statistical and evaluation pressures that reward guessing.
02 / SEE THE DIFFERENCE
One claim. One check. Sources attached.
Run an illustrative check to see the path from fluent claim to inspectable evidence.
AI RESPONSE
“The Great Wall of China is visible from the Moon with the naked eye.”
SELECTED CLAIM
STATUSReady
VERDICT
REFUTED BY RETRIEVED EVIDENCENASA states that the wall is not visible from the Moon and is difficult or impossible to see from Earth orbit without high-powered lenses.
01 / NASA — Great Wall03 / THE METHOD
Verification at the sentence level.
Kelreva does not reduce a long answer to one vague score. It checks discrete claims and preserves the evidence trail.
Select
Choose the passage that matters. Kelreva waits until you ask.
Separate
The answer is broken into specific, checkable claims and facets.
Search
Public web evidence is retrieved and ranked for relevance and authority.
Compare
Each claim is marked against the evidence and linked back to its sources.
04 / CONTROL
Nothing is checked until you ask.
Kelreva does not automatically monitor every page. A check begins only after you choose Verify, and verdicts describe support from retrieved evidence—not absolute truth.
05 / ESSENTIALS
Before you rely on it.
Does Kelreva read every AI conversation?
No. Kelreva starts only when you choose Verify. It sends the associated question and selected answer to the configured backend for that check.
Does a supported verdict mean absolute truth?
No. A verdict describes how well the retrieved evidence supports a claim. It is a decision aid, not a replacement for judgment or domain expertise.
Where does it work?
The Chrome extension supports ChatGPT, Claude, and Gemini. Select a passage or use the Kelreva action on a completed response.
Can I inspect the evidence?
Yes. Results attach source links to checked claims so you can open the underlying evidence and decide for yourself.