Why every fact our AI extracts must quote its source

ResearchRadar is built around one job: you hand it a document — a paper, a thesis, a proposal — and it watches the research field for you from then on. That only works if the system genuinely understands what your document says. And large language models, left to their own devices, are unreliable readers: they paraphrase until meaning drifts, and they occasionally state things the document never said, with perfect fluency.

For a product whose entire output is "here is what moved, and why it matters to your work", that failure mode is fatal. An alert built on an invented fact wastes exactly the attention it promised to protect. So we made a structural decision early, and it shapes everything else we build.

What does a grounded fact look like?

When our extraction loop interrogates a document, every candidate fact comes out as a typed record, never loose prose. It names what kind of fact it is — a research topic, a method, a claimed result, a limitation, a funding source. It carries the verbatim quote it rests on, with the position in the document where that quote lives. And it declares its status honestly: stated by the document, implied by it, or simply not reported.

That last one matters more than it looks. "Not reported" is treated as a finding, never as an error. If a paper doesn't state how mature its method is, the honest answer is the paper doesn't say — not a guess dressed up as knowledge.

How is the quote actually checked?

Every candidate fact is tried before it is admitted, in three steps ordered from cheapest to most expensive:

1. The exact-substring check. Is the quote a character-for-character substring of the source document? This is deterministic string comparison — it costs nothing, and it cannot be sweet-talked. A model that invents or "cleans up" a quote fails here, every time.
2. Schema validation. Does the fact parse into its typed shape — the right fields, the right kinds of values? Malformed facts are rejected before any further spending.
3. The entailment check. A separate, cheaper model reads only the quoted spans — never the whole document — and judges whether they actually state the claim. Because the verifier can't see anything else, it can't "helpfully" find alternative support the extractor never cited.

Only facts that survive all three steps become part of what the radar knows about your document. Everything else is discarded or kept solely as a flagged, unverified candidate for a human to inspect.

Why not just use confidence scores?

Because a confidence score is a self-assessment, and you can't audit a self-assessment. Models are routinely confident and wrong; a number between 0 and 1 tells you how the model feels, not whether the fact is true.

A quote inverts that relationship. You — or a reviewer, or a skeptical juror — can open the document, find the sentence, and check it against the claim yourself. The evidence is the interface. And when our public evaluation page goes live, the grounding rate — the share of admitted facts that pass the exact-substring check — is measured mechanically on every run. Until those measurements exist, the only claim we make is about the mechanism itself; we won't quote numbers we haven't measured.

What this buys the radar

Everything downstream inherits the guarantee. When ResearchRadar alerts you that a new paper matches your work, the match traces back through facts that each carry their quotes — so "why am I seeing this?" always has an answer you can read, not a similarity score you have to trust. That's the difference between a radar and a firehose: not more signal, but signal you can audit.

We're building this in the open, and early access opens in small batches — the waitlist below is the way in.