Skip to content
AlphaCitation
Request a briefingSign in

Label Ledger, for regulated markets

What AI claims about your medicine. Checked against the regulator’s label.

ChatGPT, Claude, Gemini and Perplexity answer questions about your medicines every day, outside your medical, legal and regulatory review. Label Ledger puts a fixed set of questions about your medicine to those engines through their APIs, checks what the answers claim about approved uses and populations against the label that governs the market where one does, and keeps the answers, and every case it opens, on a hashed, independently timestamped record.

Request a briefingRead the standard behind the rule

Which label decides, market by market

For claims about approved uses and populations, as the engine runs today.

For each market, which regulator decides a claim and what is shown beside it
MarketDecides the claimShown beside it
United StatesThe US FDA, from the prescribing informationThe EU product information, as the EMA publishes it
European Union, Iceland, Liechtenstein and NorwayThe European Commission’s authorisation, read from the EU product information the EMA publishes for centrally authorised medicinesThe FDA’s prescribing information
Every other market, including the UK and the GulfNo regulator’s label is read as deciding thereBoth, each labelled as not governing that market

A foreign regulator’s label never becomes a finding about another market. Where no regulator’s label governs, you see what the answers say beside both benchmarks, clearly labelled, rather than a verdict the record cannot support.

Why this exists

Every communication your company issues passes medical, legal and regulatory review. AI answers don’t.

When someone asks an assistant whether your medicine is right for a pregnant patient, how it compares with a competitor, or what it is approved to treat, the answer is generated on the spot, with no reviewer of yours, no record you hold and no obligation to match the label. It reaches the audiences your communications reach, without any of the controls that govern them.

What it detects

Five ways an answer can depart from the record.

  • Claims the label doesn’t support

    AI credits your medicine with a benefit the label it is checked against doesn’t carry, for example one that is real for the drug class or a sibling product, wrongly attached to yours.

  • Off-label population exposureHighest severity

    AI presents your medicine as appropriate for children or for use in pregnancy, the two populations the check reads, where the label excludes them. This routes to patient safety.

  • Your fact, credited to a competitor

    A benefit your label carries that AI credits to a competitor instead.

  • Real benefits AI under-tells

    On-label benefits that AI omits or buries, so the picture the answers give is narrower than the approved one.

  • Your own prohibited claims, repeated

    Claims you have designated as never to be made, checked against a list you provide and kept separate from the regulatory checks above.

How it works

From a captured answer to a closed case, every step on the record.

  1. Capture

    Each answer a scan collects is stored verbatim with the exact question, the model that produced it and the time it was returned.

  2. Check against the governing label

    Claims about approved uses and populations are checked against both the US and the EU labels, each read and dated on its own, and decided only by the label that governs the market the question was asked in, where one does. Where the two disagree, the disagreement is reported as one, never resolved into a single answer.

  3. Open a case

    Each detected discrepancy, except an under-told benefit, opens a case at once: routed to its risk domain (patient safety, promotional and regulatory, comparative, market access or narrative drift), given a severity, and held open until someone on your team owns and resolves it.

  4. Verify before it counts

    A case stays an unverified candidate until AlphaCitation verifies its finding and a named verifier approves publication, and a finding checked against a label publishes only when that label’s regulator governs the market. Until then it is never counted, ranked or headlined, it is labelled a candidate wherever it appears, and an open candidate is never sent to your ticketing system.

  5. Close only on proof

    A case is marked fixed only when a fresh scan shows the engines have stopped saying it. If a later scan finds it again, the case reopens itself.

  6. Timestamp the record

    Captured answers and case events are linked into hash chains. Once a day, if a chain has changed since its last timestamp, its head is timestamped by an independent RFC 3161 authority, so a later change to anything before it would no longer match.

Built for the teams who own this risk

One record, read by each function that answers for the medicine.

Pharmacovigilance and Patient Safety
Excluded-population exposure, surfaced when AI starts implying a use the label rules out.
Regulatory Affairs
An uncontrolled external surface making product claims, monitored against the label that governs the market where one does, with each answer recorded.
Medical Affairs
Where AI credits a benefit the label does not carry, or under-tells one it does, with the answer’s sentence and the label reference side by side.
Legal
A hashed, independently timestamped record of exactly what was said and when, and of everything done about it.
Corporate Affairs and Market Access
Competitive misattribution and narrative drift: benefits AI credits to a competitor, and registered claims AI contradicts or misconceptions it repeats.

The evidence chain

Independently timestamped, hashed and linked.

A record of what AI said should rest on more than our word.

Every captured answer is stored with a hash of its content and linked to the record before it, and case resolutions are recorded the same way, beside the observations that opened the case.

Once a day, if the record has changed since its last timestamp, its latest state is timestamped by an independent authority using an open, published standard (RFC 3161), the standard also used to timestamp signed documents and software. Each timestamp is evidence, from a party with no stake in the outcome, that the state it covers existed at that time.

Illustrative: the shape of a sealed case record

CASE_EVENT · resolvedSealed

sha256 7f2a8c91e3d4f206b5a17c8902ef4d61b3c8a90f…

chain_hash 3b91fa20c76e8419d0527bc1f93a6de40128c7e…

anchored chain head, daily RFC 3161 timestamp

Each finding’s evidence packet, exportable as a PDF, quotes the sentence it rests on verbatim, with the engine, model and capture time, the record it was checked against, and the capture’s content and chain hashes.

What that means in practice: if an engine’s answer to a scan question said your medicine was safe in pregnancy on a given date, and a later scan’s answer no longer said so, both answers are kept verbatim with the time each was returned, and the case records what your team did in between.

What the findings rest on

The rule behind every label finding is published, and open to test.

Open standard

Which Evidence Settles Which Claim

For each kind of claim an AI answer makes about a medicine, the standard sets out which evidence can settle it, for which market and as of which date. Label Ledger’s publication gate applies its rule that a foreign regulator’s record never settles a claim about another market. Version 3.0 is published under CC BY 4.0 with machine-readable rules and conformance test vectors. A pre-registered study measured how closely six AI models applying its classification annex agree with reference labels, and its benchmark is open.

Read the standarddoi:10.5281/zenodo.23594909 · benchmark doi:10.5281/zenodo.23602110

How we keep it honest

Not regulatory advice

Label Ledger is a monitoring and evidence system. It surfaces and proves what the engines say and where it departs from the governing label. The regulatory judgement on what to do about it stays with your team, and so does your own review process.

The regulator’s record and the claims you register yourself are kept structurally separate: anything resting on your own claims is labelled as yours, never presented as independent verification. Where no source can settle a claim, we raise no finding rather than guess.

Label Ledger is not a pharmacovigilance system and does not replace adverse-event reporting.

Request a confidential briefing

See it run on one of your own molecules.

Tell us one product and the markets that matter. We come back with a confidential briefing: what the engines say about that product now, checked against the label that governs each market where one does, with the captured answers and the evidence chain in front of you.

  1. We review your request.
  2. We discuss your brands, markets and priorities.
  3. You receive a written scope covering deliverables, cadence and price.

Prefer email? Write to analyst@alphacitation.com.

Configured to your products, competitors and markets. Covered by NDA on request.