Every claim, traced to its evidence. Here is exactly how.

1. 01UploadThe PDF goes from the browser straight to storage. Nothing is emailed or copied around.
2. 02ParseLlamaParse reads each page in agentic mode: text, headings, tables, figure regions and a page screenshot. Layout boxes are kept so a passage can be pointed at later.
3. 03Cross-checkThe PDF’s own embedded text layer is extracted and diffed against the parse. Numbers the parser dropped, numbers it produced that are not in the text layer, suspect tables and low-confidence pages are flagged with a severity score.
4. 04Vision checkFor flagged pages, CRx reads the page image next to both texts and returns a verdict per page (parser correct, parser wrong, unclear) with per-token findings. Nothing is applied automatically.
5. 05Human decisionA reviewer accepts the parse, applies a specific correction, or swaps in the PDF text. Reviewed pages stay reviewed.
6. 06ChunkPages become section-aware passages of roughly 300–500 tokens with character offsets. Each table is one passage; each figure is cropped and described by CRx so it can be searched.
7. 07IndexPassages and page images are embedded (Voyage) and stored in a vector index (Pinecone) under the brand’s own namespace. Nothing crosses brands.

Every passage keeps its document, page, section path and character range. That is what makes “page 989, paragraph 3, sentence 1” a fact rather than a guess.

CRx extracts verbatim, contiguous spans: efficacy, safety, mechanism, epidemiology, dosing, comparisons, statistics. Headings, calls to action, slogans and approved content blocks are excluded. Each span is then located in the actual document by text match; anything the model paraphrased instead of quoting is dropped rather than guessed.

Claims are keyed by their normalised text, so re-checking after edits keeps the references a reviewer already selected. A writer can also mark any sentence as a claim by hand, or remove one with “Not a claim”.

1. 01ExactThe normalised text matches an approved claim, or character-bigram similarity exceeds 0.95. The approved references are reused. Shown in green.
2. 02NearEmbedding similarity of at least 0.92, and CRx confirms the same product, endpoint, magnitude and direction, population and comparator. The approved references are reused with a “verify wording” note. Amber.
3. 03RewriteSimilarity between 0.75 and 0.92, or CRx finds a material difference. CRx proposes the approved wording adapted to the sentence, with a one-line explanation. Accepting it replaces the copy and attaches the approved references. Blue.
4. 04NewNo approved claim is close. The claim goes to the library for evidence.

Only claims with status Approved take part. Candidate and proposed claims never influence a check.

1. 01RetrieveThe claim is embedded and matched against passages (top 50) and page images (top 10). Bibliographies and bare headings are excluded.
2. 02RerankCohere Rerank narrows the pool to the ten most relevant passages. Without a Cohere key, the dense order stands.
3. 03GradeCRx reads all ten and grades each against the claim exactly as worded, returning a verbatim supporting excerpt and one or two sentences of reasoning.
4. 04LocateParagraph and sentence numbers are computed from the stored offsets of the passage, not asked of the model.
5. 05Keep fiveThe five best candidates are stored. Scores of zero are dropped.

View source opens the page with the passage boxed on the page image, the excerpt, the full reasoning and the full passage, and prints as an evidence exhibit: source, page, paragraph, claim, excerpt, reasoning, date.

Approved claims come in three ways: upload an MLR-approved piece and CRx extracts each claim with the citations it carries; paste claims one per line with their citations; or ask What can the brand say about ___ and CRx drafts candidate claims from the library, each pinned to one passage with a verbatim excerpt and graded against it (anything under 70 is discarded).

Statuses: Candidate (drafted, unreviewed) → Proposed (curated, ready for MLR) → Approved → Superseded or Withdrawn. Approval is a human action by a brand admin. Only Approved claims are indexed for matching, so nothing drafted by a model can ever be treated as approved by accident.

Writers draft in a plain rich-text editor. Approved content blocks (boxed warnings, safety information) drop in with @ and carry their own references; when the library publishes a new version, projects are told an update is available and can take it, unless the project is already approved or exported, in which case its blocks are frozen. Flow State hides everything but the paper.

Reviewers comment on ranges, mention colleagues, and resolve threads. Every action lands in the project’s activity feed and the brand’s audit log with who and when.

Export produces a Word document: the copy with citations numbered by first use (AMA superscript or Vancouver bracketed), a reference list built from each source’s citation line, and an annotated appendix listing each claim with its selected references, the supporting excerpt and its location. Exporting freezes the content blocks at the versions that went out.

Tenancy. Every row carries the brand id; every query is scoped to it; files and vectors are keyed by it. There is no cross-brand search, and no brand is visible to anyone who has not been invited to it.

Access. Invitation-only. Brand admins invite by email and assign a role (admin, writer, reviewer, viewer); a company email domain can be pre-approved. Sign-in with Google, Microsoft, passkeys or an emailed link; no passwords. Enterprise single sign-on can be set up per brand on request.

Deletion. Deleting a document removes the file, the parsed pages, the passages and their vectors. Deleting a workspace removes everything under its brand id. Details in the privacy policy.

CRx does not approve claims and does not decide what is compliant. It finds, grades, and shows its work so that the people who are responsible can decide faster and with the evidence in front of them. A score is the model’s assessment of one passage against one sentence; the reasoning next to it is there to be checked, and the boxed page is there to be read.

Models can be wrong. The design assumes it: excerpts are verbatim, locations are computed, approvals are human, and every action is logged.