Sign inSee APLOMB verify
The protocols

Fourteen protocols, fourteen precise triggers.

APLOMB does not « verify everything ». Each protocol fires only if a deterministic pattern is present in the question or in the model's answer. Here are the fourteen, what fires them, what they prove, and above all what they do not. Thirteen query a public source live; only one, the statutory interest rate, relies on an internal table, and we say so on its line.

The trigger is deterministic, not guessed

No source is called « just in case ». A company number must be present and valid, a medicine must be named, a DOI must be cited. That is what makes verification reproducible: same question, same sources queried. It is also an owned limit, and a simple one: if nothing is identifiable, nothing is cross-checked, and APLOMB says so instead of pretending.

Law

Légifrance

PISTE API (DILA) · codes and statutes
Trigger
Every reference (article + code or statute) cited in the answer is cross-checked live, one by one.
What it proves
That the cited text exists, under that number, with that exact wording. Signed receipt with official extract and link.
What it does not prove
That the article applies to your situation. A real text can be invoked wrongly.

Companies

SIRENE

recherche-entreprises.api.gouv.fr
Trigger
A French company number (9 or 14 digits) present in the question and validated by the Luhn checksum. Two at most.
What it proves
Existence, administrative status, directors, published figures.
What it does not prove
The company's actual health, nor the truth of what is said about it.

BODACC

opendatasoft (DILA) · commercial notices
Trigger
Same trigger: an identified company number.
What it proves
The existence, or absence, of a published insolvency proceeding.
What it does not prove
A difficulty not yet published in BODACC.

Finance

Statutory interest rate

Internal table, rate published in the Journal officiel
Trigger
The words « taux » and « intérêt légal » in the question.
What it proves
The rate applicable for the current half-year, for individuals and professionals.
What it does not prove
No stable API exposes this rate: it is an internal, versioned table, to be revised twice a year. We would rather say so.

Health

BDPM

French public medicines database
Trigger
A medicine named in the question. In-memory index of the official file, refreshed daily: zero network call per request.
What it proves
That the medicine exists and holds a marketing authorisation.
What it does not prove
That it suits you. APLOMB makes no therapeutic recommendation.

ANSM shortages

French medicines safety agency
Trigger
Same trigger: a named medicine.
What it proves
A declared supply tension or shortage.
What it does not prove
Actual availability at your pharmacy.

FINESS

data.gouv.fr · health facilities
Trigger
A FINESS number explicitly cited.
What it proves
The declared identity and address of the facility.
What it does not prove
Its quality, nor its actual activity.

RPPS

French health directory (ANS) · FHIR
Trigger
An RPPS number explicitly cited (11 digits).
What it proves
That the professional is registered, and under which profession.
What it does not prove
Their competence, nor their availability.

ICD-11

World Health Organization
Trigger
An ICD-11 code cited in official format.
What it proves
The official label of the code.
What it does not prove
A diagnosis. A code is not medical advice.

Odissé

Santé publique France
Trigger
The terms « couverture vaccinale » or « taux de vaccination ».
What it proves
The official published figure.
What it does not prove
A causal reading of that figure.

GHO

Global Health Observatory (WHO)
Trigger
The term « espérance de vie » (life expectancy).
What it proves
The official indicator published by the WHO.
What it does not prove
An individual projection.

OpenFDA

FAERS · US pharmacovigilance
Trigger
« adverse effects » or « pharmacovigilance », AND an identifiable medicine. The French-name to international-name mapping uses a short, safe table, never guessed.
What it proves
The volume of spontaneous reports recorded in the United States.
What it does not prove
Causality. A report is not proof of an effect, and the data is American.

Scientific literature

Crossref, Europe PMC, ClinicalTrials

public registries
Trigger
A DOI, a PMID or an NCT identifier cited in the question OR in the model's answer.
What it proves
That the publication really exists under that identifier. A precise identifier not found in the registry is flagged: that is the marker of a fabricated reference.
What it does not prove
The quality of the study, nor that its conclusions support the claim.

Rumours

ClaimReview

Google Fact Check Tools · AFP Factuel, Le Monde, Reuters
Trigger
The verifiable claims extracted from the model's answer.
What it proves
That a published, dated fact-check signed by a newsroom contradicts, corroborates or qualifies the claim. A contradiction produces the Contested verdict.
What it does not prove
The absence of a fact-check proves nothing. Many true or false claims have never been fact-checked.

Every confirmed source leaves a receipt

A successful call produces a receipt: the source, the exact query, the timestamp, the HTTP code, the SHA-256 hash of the response and the public link. These receipts are sealed in the certificate. A third party can replay the call and compare.