How it works
Every dembli answer follows the same path: the official publication is stored before it is read, the data is put in a comparable form, the comparison is explained signal by signal and everything is dated. This page shows what each stage guarantees and how your team checks it.
01 · Acquisition and proof
The official publication, stored before it is read
Every source is read from its publisher. The file received is identified by its SHA-256 and stored before any parsing, and it is never altered afterwards. Every fact served points to the record, the publication and the file it came from.
- Every screening leaves a receipt: who queried, when, the versions used and the parties reported.
- The certificate is the receipt signed with Ed25519, bound to the publications consulted and their files' SHA-256.
- The period's dossier for the regulator is a signed file whose last line carries the SHA-256 of everything before it.
How to check
Check a certificate outside the platform, with dembli's public keys and the publisher's own files.
Illustrative example · invented data
Screening receipt
- Query
SCR-DEMO-001- Made at
2026-11-03 09:05- Profile
example-profile v3- Publication
sha256 9f2c…e41a- Signature
Ed25519 · JWS
Verifiable outside the platform with the public key
02 · Normalization
Validated documents and names comparable across scripts
Before comparing, every item is put in the form in which it can be compared, without losing what the source published.
- Document numbers from more than 15 national schemes are validated by the issuer's check digit. A number that fails stays unverified, never negative.
- Structured identifiers, such as an LEI, an IMO number or a Brazilian CNPJ root across head office and branches, are normalized before lookup.
- Names in other scripts, such as Cyrillic, Arabic and Hangul, are compared in their Latin form. A transliterated name never counts as an exact match, and the explanation says it was transliterated.
- The alternative names and spellings the publisher itself gives are part of the comparison.
How to check
The explanation of every result shows which document was validated and whether a name was transliterated.
Before comparing
LEI 5493001KJTIIGC8Y1R12valid ISO 17442Иван ПетровIvan PetrovDocument 4821-7unverified
03 · Matching
A deterministic comparison, explained signal by signal
In screening, the same input under the same versions always gives the same answer. The search starts from the document; the name comes after.
- Every signal comes back with its name, value, weight and effect, beside the contradictions and what was left out, each explained in a sentence.
- Five classes, from NO_MATCH to EXACT_MATCH, say how likely it is that the subject is the listed candidate. None of them measures risk.
- A missing attribute does not count as a contradiction, and a comparison that could not be made does not count as zero.
- Profiles belong to your organization: lists, the lowest class reported and the PEP window, each profile versioned and recorded in the receipt.
How to check
Send a batch with the outcome you expect on each line: each comes back found, withheld by the profile, missed, false alert or cleared.
Screening result
- Name
- compatible
- Date of birth
- not stated by the source
- Country
- compatible
- Document
- not given
POSSIBLE_MATCH Possible match · review
Match confidence, not risk. The decision is your team's.
04 · Time
What was known on any date
A screening or a party's record can be read as it stood at a past instant, with the publications served at that moment.
- A past-date screening returns what a screening on that day returned, and the receipt names the instant and the publications served.
- Every publication is dated from the moment the platform begins reading it, and earlier periods follow the dates the official register itself states.
- Receipts are kept for six or ten years, as your organization's retention duty requires.
- Monitoring tells what each publication changed for every customer followed.
How to check
Rerun today the screening of a past date and compare it with the receipt kept from that day.
Screening on a past date
- Query as of
- 2026-10-15
- Publication served
- Example list · 2026-10-12
- Result that day
NO_MATCH- Result today
POSSIBLE_MATCH
The list changed on 2026-11-01; each day's receipt names the instant and the publications.
05 · Measured quality
What the sources leave out is stated too
The platform measures its own work and reports each source's completeness, including when it is low.
- Every list reports how many of its parties carry a date of birth and a document, so your team knows how much each signal can help.
- A new matcher configuration ships only with a reference baseline measured under it on the real corpus.
- Your team measures its own profile with test batches, as in the matching stage.
How to check
Ask for any list's completeness: how many parties carry a date of birth and how many carry a document.
What holds at every stage
Unknown is not negative
When the source does not state a value, the answer says it is unknown. Absence is never presented as proof.
News is not an official fact
A press mention is attributed to its outlet and kept apart from official records.
The decision is your team's
dembli delivers facts, explained matches and proof. Accepting or refusing follows your institution's rules.
This page describes what each stage guarantees and how to check it. The matcher's code is not published; every answer shows the signals and weights it applied, which is what your audit needs.