Every document has two verification surfaces. The structural surface is what the document's shape reveals — clauses, fields, architecture. The biographical surface is the space between the document's claims and the real identities of the people inside it.
Prewash scrubs the first. Rinse scrubs the second.
Errors planted on the structural surface fall out in Prewash. Errors where a real person is made to say something their real self would never say only fall out in Rinse. Sentences can be individually true and still biographically impossible for the speaker attached to them. That gap is where fabricated quotes live — and it is invisible to any fact-check that grades propositions one at a time.
Rinse reframes verification from is this sentence true to is this sentence true for this speaker. Know your subjects before you quote them. The discipline is old. The protocol is new.
Any document where the meaning depends on who is speaking: transcripts, interviews, panel debates, press conferences, leaked audio, quoted social posts, podcast excerpts, on-screen chyrons, reported speech in news articles. If you could describe the document as "people talking," it is a Rinse job.
The user needs no prior knowledge of the document. The prompts are generic and work on any speaker-driven material. Run them in order. Do not skip cycles.
Before reading any claims in this document, list every named speaker, author, or subject. For each one, build a short factual profile from independent public sources — not from the document itself. Include:
- Place of birth / origin and current base
- Current role, title, and affiliation (with date of appointment)
- Previous roles relevant to the topic
- Signature positions, policies, or rhetorical phrases they are publicly associated with
- Biographical anchors: education, hometown, family, career path
- Any distinctive vocabulary or framing they are known for
Cite sources for every profile. Do not yet evaluate anything the document says. The goal of this pass is only to know who these people are, independent of the document.
Now read the document. For every factual claim each speaker makes — especially about geography, institutions, numbers, their own biography, their party or organisation's positions, or named third parties — check whether it is consistent with the reference profile you built for that speaker.
Rules:
- Flag any claim that does not fit the speaker's known self, even if the claim would be independently true in isolation. A true fact in the wrong mouth is still a finding.
- If a single sentence contains multiple independent claims that each fail the profile check, list each claim as a separate finding.
- For each speaker, list every geographic, institutional, and biographical anchor they invoke across the document, and check whether those anchors form a coherent set for this speaker. An incoherent set is itself a finding, even when each anchor could be individually defended.
- Treat inconsistency with the speaker's profile as a red flag equal in weight to a factual error.
Finally, check the document for internal contradictions across three layers:
- Speaker-vs-profile: summary of Pass 2.
- Speaker-vs-self: does each speaker's argument hold together as an argument, not just as a sequence of individually defensible sentences? Do the anchors inside one speaker's turn align with one another?
- Document-vs-self: does the document contradict itself? Look specifically at labels, titles, captions, chyrons, dates, and any metadata that appears more than once. A speaker introduced under two different affiliations, a title that does not match a role, a date floor that conflicts with a stated fact — all count.
Report all three layers separately. Rank findings by severity at the end, and distinguish hard errors (factually wrong) from coherence failures (not wrong in isolation but wrong for this speaker or this document). Do not soften findings. Do not collapse the passes. If a claim passes all three checks, say so.