Зажми картинку чтобы сохранить
Ask a model for a quote and it delivers one. Confident, polished, well-formatted.
Half the time the person never said it. The model just knows what quotes sound like.
Lawyers sanctioned over invented cases. Papers citing sources that don't exist.
The scary part: fabricated references look exactly like real ones.
"Please don't hallucinate" is a wish, not a mechanism.
You can't prompt your way out of a system with no ground truth.
In a trip2g knowledge base, every passage gets a unit id, an address pointing to the exact spot in the source note.
A quote without an id isn't a quote. It's decoration.
When an agent cites something, it fetches the passage by id.
The text on screen is pulled from the note, character for character. Not reconstructed from memory.
Can't find a passage that supports the claim? The system says so.
"I don't have a source for that" is a feature. It's the whole feature.
Every citation links back to the note it came from.
Click through, read the context, judge for yourself. Trust becomes optional.
This works on a curated base, not the open internet.
Your sources, your excerpts, your vault. A small world where verification is actually possible.
Research summaries. Book notes. Anything you'll repeat in public with your name on it.
If you're going to quote it, you'd better be able to point at it.
Any AI that quotes should show where from. Passage, source, link.
Accept nothing less from your own tools.
No id, no quote. No source, no claim.
Save this before an invented citation ends up in your next post.