Compare
Danphe vs ChatGPT with citations
A chat answer with links tells you where a model looked. It doesn't tell you whether each sentence is true to that source. Danphe is built for the second question.
- Quotes on the right passage
- 100%
- contracts, locked test
- Quotes on the right passage
- 82.1%
- science papers, locked test
What you get from each
A general comparison of approach, based on how chat assistants behave for users; not a test of any one product.
| What matters | Danphe Council | Chat assistant with citations |
|---|---|---|
| Who writes the answer | Several models from different companies answer on their own; an editor writes from what held up | One model |
| What a citation means | Each statement is checked by other models against the quoted passage of your document | A link or source the model chose to show; nobody else checks that it says what the answer claims |
| Disagreement | Shown: disputed points are labelled and both readings are kept | Not visible: you see one fluent answer |
| Statement-level labels | Every statement: agreed, disputed, not supported or opinion | None |
| Your own documents | The answer is built from your files, with page-level quotes | Possible with uploads; quotes are not independently checked |
| Record for the file | A signed, hash-chained audit record of models, verdicts, quotes and cost | The chat transcript |
| Speed and cost | About 20 seconds to 5 minutes; $0.24 model cost per question on average in our test | Seconds; cheaper per question |
A citation is not a check
A cited answer can still misread its source: the link is real, but the sentence next to it says more, or less, than the passage does. You only find out by opening every source and reading it yourself.
Danphe splits the answer into statements and asks models from other companies to check each one against the passage it quotes. Where a dataset marks the passage that holds the answer, 100% of our quotes landed on it for contracts and 82.1% for science papers in our locked test.
When a chat assistant is the better tool
For drafts, brainstorming and questions where a wrong sentence costs little, a single chat assistant is faster and cheaper. Use Danphe when a wrong answer is expensive and someone will ask you to show your working.