Every claim, anchored to its evidence.
Evidence Librarian watches the literature for you, recommends what belongs in your library — and explains why — and tells you what every new paper means for the positions you already hold — with each verdict quoted verbatim from the source.
Mean opioid consumption through 72 hours was 38% lower with liposomal infiltration (p < 0.001). A multicenter randomized trial across 14 sites (N = 418). Time to discharge did not differ between groups (p = 0.41).
Your evidence process is running on heroics.
Small medical information and communications teams carry hundreds of publications, thousands of claims, and a review calendar that never pauses. The tools underneath are a shared drive, a reference manager, and a color-coded spreadsheet somebody maintains by hand.
The alert email lands.
Two or three new publications a week, every week. Someone reads each one, summarizes it against a template, and the team meets to decide what gets in — a process that lives in inboxes and memory.
The statement gets re-researched. Again.
Nobody fully trusts the last version of the scientific platform, so every new deck starts from a fresh literature search — the same evidence, re-reviewed from scratch, by whoever drew the assignment.
The rounded number slips through.
The source says 0.4 days; the slide says “about half a day.” The study said inpatients; the claim didn’t. These are the misses that surface in review — or worse, after it.
Nobody remembers which SRDs are stale.
A new publication changed the picture three weeks ago. Which response documents cite the old one? The honest answer, in most teams: whoever updated it last might know.
One loop, from journal to approved response.
Watch
Saved search profiles — your products, terms, and study types — run against the literature on a rolling window you control. New publications arrive as candidates, not chores. Got a paper a colleague emailed? Paste the link, the PMID, or the DOI and it joins the same queue.
Judge
Each candidate comes with a recommendation — add or skip — the reasoning, and the passages it rests on, quoted verbatim from the abstract and each annotated with what it proves. You approve or decline. The rationale is kept either way.
Curate
Approved papers arrive carrying the abstract verbatim, their study design read off PubMed’s own metadata, and every summary field openly marked pending — the model drafts one on request, and anything the abstract doesn’t cover stays pending rather than invented. They roll up into a library of statements with plain-language wording, an evidence-strength score weighted by study type, and a currency date.
Enforce
New evidence flags the statements it affects. Superseded references flag everything citing them. A statement cannot hold approved status without an active supporting reference — and loses it the moment its last one goes.
Learn
When your librarians overrule a recommendation and say why, the reason becomes precedent. Accumulated precedents can be distilled — with your lead’s sign-off — into standing rules the triage follows from then on.
Judgment you can inspect, provenance you can click.
Recommendations that show their work.
The inbox doesn’t just fetch papers — it reads them against what your library already holds. A letter to the editor with no new data gets a “skip” and the reason. A national database study that extends your single-center cohort gets an “add” and the gap it fills.
You decide the net: a rolling surveillance window per search profile — the last N days, recalculated on every fetch — or explicit dates for a fixed sweep. Letters, editorials and commentary are filtered out unless you ask for them; retraction notices and errata are excluded from every profile, always, because a retraction belongs in reference health, not in the queue as evidence.
Your cap is a cap on new papers, not on results looked at: ask for twenty and the fetch keeps searching until twenty genuinely new candidates exist or PubMed runs out, then tells you all three numbers — new, duplicates skipped, results found. Everything already reviewed, approved or declined, is remembered, so nothing comes back twice.
Papers that arrive some other way go in by hand. Paste PMIDs, PubMed links and DOIs in any mix — pasted citation text, bullets and spreadsheet commas are ignored — or attach a text or CSV file. The import reports what it recognized, what it deduped, which DOIs PubMed couldn’t resolve and which lines held no identifier at all. Imported papers are ordinary candidates from there on, distinguished only by a provenance chip.
Triage runs live. Start the queue and it works in small concurrent batches behind a running count — each card filling in with its recommendation, its rationale and its quoted passages the moment its own verdict lands, so you read the early ones while the rest are still going. A candidate the model fails on stays untriaged, shows the error on its own card, and is skipped by the remainder of the run; nothing pauses, nothing loops, and the closing line counts the failures alongside the adds and skips.
Every recommendation is a starting point for your judgment, never a substitute for it.
It learns the way your team decides — on the record.
The AI’s judgment isn’t a black box you adapt to. It’s four layers of plain text you can read on the inbox itself: the base judgment everyone can see and your lead can amend, team rules your admin sets, personal rules each librarian tunes for their own products, and precedents — past decisions where a person overruled the AI and wrote down why.
Only explained overrides teach the system. And when the same lesson keeps repeating, one click drafts it into a standing rule — citing the exact decisions it came from — for your lead to edit and ratify. Nothing changes how the AI judges without a person choosing it.
Not just the paper. The passage.
Every statement in the library is linked to the exact sentences that substantiate it — quoted verbatim from the source, and machine-verified so a paraphrase can never masquerade as a quote. One click, and the reviewer is reading the highlighted passage itself. One more, and they’re on the PubMed record it came from.
Anchors come from two places: a reviewer quoting the passage themselves, and an impact reading accepted during inbox triage, which lands its quote on the statement when the paper is approved into the library. Either way the quote is checked character-for-character against the stored source. If it isn’t really there, it doesn’t save.
The source text isn’t freely editable either. Correcting a transcription error goes through a governed route: a written reason is required, the before and after are both recorded, and every anchor and intake passage on that publication is re-validated against the corrected text. Anything that no longer resolves is flagged for a person — never quietly dropped.
In 1,842 propensity-matched pairs, mean length of stay was 1.3 days with liposomal infiltration versus 1.7 days with standard infiltration (p < 0.001). There was no difference in 90-day readmission (2.8% vs 3.1%, p = 0.58).
Abstract — Results, ¶1 · verified verbatimWhen the evidence moves, everything citing it knows.
Mark a publication superseded or retracted, and the effect cascades: every statement that leaned on it flips to needs review, every response document fed by it is flagged, and the reason is written down — automatically, immediately, with nothing left to memory.
Approval has teeth, too: a statement cannot reach approved status without at least one supporting reference that is both included in the library and still active, and it loses that status the moment its last one goes. Newly included publications that match a statement’s topic surface on its card as suggestions — link them or dismiss them, but they don’t sit there unnoticed.
Response-document health is never a stored field somebody has to remember to update. It is recomputed every time the registry is read, from the publications actually feeding each document — so a candidate you approve in the inbox flags the affected document in the same second, with the offending publication named in the reason.
Where your evidence is thin, on one screen.
The dashboard reads the library back to you: how the outcome directions divide — favorable, neutral, mixed, unfavorable, each with its share — a breakdown by product, and a heat map of outcome direction against your category dimension, where the cells darken with the number of publications behind them.
The pale column is the point. A near-empty row for a procedure area is next quarter’s surveillance profile, written for you. Filter first, or restrict the view to evidence actually cleared for use, and the picture is scoped to what you could genuinely cite tomorrow.
The colors mean the same thing here as everywhere else in the product — the outcome palette is fixed, so a chart read on the dashboard reads identically on a publication row.
A trail your quality team will actually like.
Statements move through a real review flow — authored, submitted, assigned, approved or sent back with comments. Every transition is recorded server-side: who, when, what changed, and whether it was a person or the system acting.
Overruling the AI is part of the record too: your lead decides whether a written reason is optional or required when a librarian goes against a recommendation — and either way, the decision, the decider, and the reasoning are kept.
Self-approval isn’t forbidden — small teams need the escape hatch — but it’s marked, prominently and permanently. Transparency over ceremony.
The conversation is on the record too. Each publication carries an append-only discussion thread — author and timestamp, never edited, never deleted — so the field question that prompted a second look is still there next year, next to the decision it produced.
- Aug 28 · 09:12reference linked — national cohort (anchor: Results ¶1)
- Aug 28 · 09:14submitted for review → assigned: M. Alvarez
- Aug 28 · 11:02status changed: Submitted → Approved
- Aug 29 · 07:00flagged by reference health — supporting reference superseded
Guardrails first. Then speed.
Human in the loop, always
The system recommends, drafts, and flags. People approve. Every consequential action waits for a decision — and keeps the reasoning attached to it.
Nothing claimed without a source
Approval gates make unreferenced statements structurally impossible, not just discouraged. Assembly itself never invents prose: every assembled sentence is a verbatim field of a publication, the text of an approved statement, or your own boilerplate. Where a section has to be written by hand — the prescribing information, say — the document says so rather than pretending a machine sourced it.
Memory you can read
The AI’s memory is the explained portion of your audit trail — rules your team wrote and overrides your team justified. Nothing more, and every word of it is on screen.
Transparent search
Saved search profiles make your surveillance strategy explicit and repeatable — the same terms, the same scope, the same filters, every week, on the record.
Verbatim or it doesn’t save
Anchors and evidence passages are checked character-for-character against the stored source before they attach. The AI can quote your library; it cannot paraphrase into it.
Corrections are governed
Source text can be fixed, but only with a reason on the record — and every quote resting on it is re-checked, with the failures raised rather than removed.
No silent deletions
A declined paper stays visible with the reason it was declined, and can be restored to the queue whenever the picture changes. Six months on, you can still see what you passed on.
The manual is in the product
An eight-step walkthrough ships inside the app: what each stage is, why it matters, exactly how it behaves, and where the AI does and doesn’t decide. Onboarding isn’t a PDF.
Your instance. Your vocabulary.
- A dedicated, isolated environment for your organization — your data shares infrastructure with no one.
- Your words, not ours. The category dimension is renamed inside the product — “surgery type” becomes “indication”, or “tumor type” — and the new label follows through every filter, table, form and chart. The data underneath doesn’t move.
- Access by role. Each area — inbox, library, statements, response documents, dashboard — opens only for the roles you grant it, and a person’s navigation shows only what they can actually open.
- Shaped to your workflow — your document names, your classification criteria, your review path, your triage rules. Tailoring is measured in conversations, not change orders.
- Delivered and supported by Acceleration Point, with the medical affairs judgment to configure it well — sixteen years inside these teams, not adjacent to them.
Bring one week of your real inbox.
The best evaluation is your own workflow. In a working session, we set up your search profiles, run your actual literature through the loop, and build a response document from the evidence it surfaces — so you see your Tuesday, not our demo.
Request a working session