phronesis · laserbrain
laserbrain AI
A model handed a document and a question returns fluent prose either way — whether every sentence came from the document, or whether one came from nowhere. Nothing errors. The answer reads the same. This finds the sentence that did not come from the source, and shows you the words it used to decide.
Everything runs in this page. No model, no key, no request — which matters here more than elsewhere, because the input is by definition someone’s document.
the source it was given
the answer it returned
quotations that are not in the source
“this transforms our position in the market”
A quotation says these exact words are in the document. Curly quotes, apostrophes, dashes, spacing and case are normalised — those are typography, not misquotation — and an ellipsis is split so a properly elided quote is not flagged. It asks whether the words are in the document, not whether the person named said them.
names that are not in the source
Vantiv Holdings
Names survive paraphrase. A sentence can be rewritten forty ways and “Vantiv Holdings” is in every one of them or in none — so a name from nowhere is evidence in a way the score below cannot be. It finds names from nowhere, not names in the wrong place: the same names in the opposite claim pass clean.
figures that are not in the source
91
A figure cannot be paraphrased. Either the source contains it or the answer invented it — which is why this is reported on its own rather than folded into the score below.
Not a probability and not a truth score — the fraction of content words in the answer that appear in the source. Click a sentence for every term.
sentence by sentence
the frozen reference answer
Freeze an answer and every later one is compared against it — claim by claim, so said differently and says something different stay separate. It is never revised on its own: an instrument that re-derives its reference from the latest answer follows the thing it is measuring and reports that nothing moved. Stored in this browser only.
what it is not — read this before trusting a number
Faithful paraphrase reads as unanchored. An answer that restates the source correctly in its own words scores badly here. That is a false positive, it is common, and it is exactly where good writing lives.
A fabrication built from the source’s own words reads as anchored. Recombining the document’s vocabulary into a claim it never made scores perfectly. That is a false negative and it is the more dangerous of the two. No lexical method can see it.
So this is a screen, not a verdict: it finds the sentence worth a person’s attention and says why. Calling it a hallucination detector would be the same overclaim as calling a keyword tally “coherence” — which this line made once already and had to walk back. Both limits above are pinned as passing tests; if either is ever quietly fixed, the build fails until this page is rewritten.
A sentence is flagged below 60% anchored terms, and counted as clean at or above 85%. A quoted span is read as a quotation at 3 words or more — below that it is usually emphasis, and checking it would produce noise instead of findings. All three are choices, printed here rather than buried, because a threshold nobody can see is a threshold nobody can argue with.
both checks, over http
The page is the demo. The check belongs where the model call happens — a pipeline, a CI step, a gate before deploy — so both run as endpoints. No key, no state, nothing stored.
POST https://docs-drift.degibug.workers.dev/v1/anchor
{"source": "...", "answer": "..."}
POST https://docs-drift.degibug.workers.dev/v1/answer-drift
{"reference": "...", "current": "..."}