Confidence & methodology
How Ahel scores confidence
Every answer Ahel returns is built from sourced findings, and every source is tagged with a confidence tier and a percentage. This page explains what those tiers mean, how the number behind them is derived, and how to cite a finding so your reader — or your auditor — can check it. Confidence is a signal, not a guarantee; the point is that you can always see how strong the ground under a claim is.
The four tiers
These are the exact tiers and colours you see on the source strip under a chat answer. Each source is sorted into one of them.
An independent source backs the claim, or a verification pass returned a supporting verdict. The strongest tier Ahel assigns.
How to read it: Safe to treat as established, with the source named alongside. Still cite the source and date — Verified means well-evidenced, not infallible.
A credible match that falls short of independent corroboration — one good source, or a strong-but-imperfect identity match.
How to read it: A working assumption, not a fact. Worth a second source before you publish or make a decision on it.
Surfaced but not scored: a weak or fuzzy match, an inconclusive verdict, or a source that returned no confidence value.
How to read it: A lead to chase, never a conclusion. Treat it as a pointer for further checking.
A verification pass found a source that contradicts the claim. This is an active negative signal, not just an absence of evidence.
How to read it: Evidence points the other way. If a claim is Refuted, lead with that — it is often the most newsworthy result.
How a score is derived
A tier is never a single magic number. It comes from three signals that Ahel records on every finding, so the reasoning stays auditable end to end.
1 · Source confidence × identity confidence
Ahel keeps two confidences apart, and that distinction is the heart of the method. Source confidence is how sure a given source is of the datum itself. Identity confidence is something different and just as important: how sure we are that the record is about the same subject you asked about, rather than a namesake. A record can be perfectly accurate and still be the wrong person — so a strong claim from a weakly-matched identity is held back, not promoted. The tier reflects both.
2 · Match quality
Each finding also carries how it was matched — exact (a precise, unambiguous identifier match), fuzzy (a close but inexact match), or inferred (derived rather than directly stated). An exact match on a unique anchor — a registration number, a verified email, a wallet address — carries far more weight than a fuzzy match on a common name, and the tiering treats it that way.
3 · Corroboration across independent sources
One source is a data point; several independent sources agreeing is evidence. When a claim is checked, Ahel records how many sources were consulted and how strongly they agree — an agreement measure across the receipts. Independent corroboration is what lifts a claim toward Verified; a lone source, however confident, lands lower.
4 · Verdict stance
When a finding is run through verification, each source returns a stance on the claim: it supports, refutes, or is inconclusive. Stance outranks the raw percentage. A supporting verdict lifts a source to Verified; a refuting verdict marks it Refuted outright, regardless of how confident the underlying source was — because a contradiction is the result you most need to see.
Sources & provenance
Every source on the strip is named so you can follow it. Named public authorities appear under their own name; everything else is grouped into neutral record categories.
Named public authorities
Where a result rests on an official, citable register, Ahel names it directly — for example the Estonian Business Register (Äriregister), the GLEIF legal-entity index, the OpenSanctions consolidated lists, Nasdaq, and national company and court registries. These names are the citation: you can point a reader or an auditor straight at the authority of record.
Record categories
Not every finding maps to a single named register. Where it doesn’t, the source is grouped into a clear category that describes what kind of record it is — Primary records (Ahel’s own ground-truth layer), Identity records, Contact records, Business registry, and Social profilesfor the subject’s own public accounts. The category tells you the nature and weight of the evidence without overstating its origin.
What we publish, and what stays proprietary
We publish the source name or category, the confidence tier, the percentage, and — where the source exposes one — a link to the underlying record. The exact methods behind Ahel’s own proprietary ground-truth layer are not disclosed. That boundary is deliberate: it protects the integrity of the data without weakening your ability to cite the finding, because the tier, the score, and the named authority are all in front of you.
Using this in your work
For journalists, investigators, and compliance analysts, a finding is only as good as the way you record it. Cite three things together: the source (the named authority or record category), the tier and percentage, and the date you ran the lookup — sources change, and a confidence score is a snapshot of what was known when you asked.
Citing a finding
“Registered in Estonia under reg. code 16498765 — Verified, 96%, Äriregister, checked 22 Jun 2026.”
“Listed on a sanctions programme — Refuted, 0%after checking OpenSanctions, no current match, checked 22 Jun 2026.”
- Lead with the tier, not the percentage alone — “Verified” and “Probable” carry meaning your reader can weigh; a bare number does not.
- For decision-critical claims, don’t publish on a single Probable or Unconfirmed source. Use it to direct further reporting and confirm independently.
- A Refuted result is a finding in its own right. Record it and the source that contradicted the claim — it is often the strongest line in the story.
- Re-run before you publish if time has passed. The tier is current as of the lookup date, and registries update.
More on the underlying terms in our terms and privacy policy. Questions about a specific result? [email protected].