Skip to main content
Methodology

How we check your AI’s work.

Every ASURIQ verdict traces back to this page. Three scored categories, a published standard, and a hard line about what we do and don’t claim. This is the citable, versioned account of the method.

v1.1Effective 11 August 2026
AI answers
You ask ChatGPT, Claude, or Gemini
ASURIQ checks
269 databases queried automatically
Three scores
Correct? Sourced? Current?
Badge appears
Right next to the answer
1. Standard lineage

Not invented for a product.

Source evaluation has a standard: the Admiralty Code, formalised as NATO AJP-2.1. It rates evidence on two independent axes — how reliable the source has been historically, and how well the specific information is backed up by other sources. Originally built for military and intelligence assessment, it’s still the working standard in open-source intelligence and structured fact-checking today.

ASURIQ’s three scoring categories extend that lineage rather than replace it. When a party disputes a verdict, the answer isn’t “trust us” — it’s this page: here is the standard, here are the categories, here is the record behind this specific score.

Three-category scoringEvidence axesIndependenceWhat we measureFact-check integrationDatabase coverageYour own sourcesLegal standingVersioning
2. Three-category scoring

What the badge actually measures.

Every check answers three questions. Each question gets its own score. Only one of them drives the badge color.

82of 100
Is this correct?
Drives the badge color
61of 100
Is this well-supported?
Adds a note if low
90of 100
Is this current?
Risk amplifier
Is this correct?
Factual accuracy, internal coherence, and balance of perspective. Combined as a weighted geometric mean — if factual accuracy is zero, the whole score collapses regardless of how coherent or balanced the answer is. This is the only category that determines the badge color.
Is this well-supported?
Whether the answer cites its sources and covers the topic completely. A correct answer with no citations is still correct — low evidence quality adds a note to the badge but never turns it amber or red. This answers a different question than truthfulness: not "is this right?" but "did it show its work?"
Is this current?
How recent the information is and whether the domain requires extra care. Context acts as a risk amplifier: the same truthfulness score gets tighter requirements in medical or legal domains than in general trivia. Context does not generate its own badge state — it adjusts the bar for truthfulness.
Five badge states
Verified
Checks out against the evidence
Noted
Correct, but sources are thin
?
Contested
Evidence disagrees with the answer
Flagged
Contradicts the record
Unknown
Not enough data to judge

Green is the most common state, because most correct answers should be green. When they aren’t, the thresholds are wrong, and we fix them — not the other way around. The badge rewards accuracy. It does not manufacture doubt.

3. The evidence axes

What feeds the three categories.

Three categories answer your questions. Six evidence axes are how we get the answers. Each axis measures one specific property of the evidence behind a claim.

Feeds: Truthfulness
W
How much evidence exists and where it comes from
Corroboration density
Not a raw count of sources. Measures how evidence is distributed across independent origins, so five outlets repeating one wire report count as one voice, not five.
C
Did sources find this independently?
Cumulative coincidence
Whether independent sources landing on the same conclusion exceeds what chance alone would produce. The more unlikely the agreement, the stronger the signal.
S
How much this matters
Stakes
Higher-stakes claims warrant deeper evidence before a verdict. A trivia fact and a medical dosage claim receive different scrutiny — this axis encodes that difference.
Feeds: Evidence quality
B
Who benefits if this is believed?
Beneficiary concentration
Measured structurally — ownership records, disclosed funding, financial interest. Never inferred motive. The question is who gains, not why they might.
E
How concentrated is the evidence supply chain?
Epistemological capture
If most evidence traces back to one funder, one publisher, or one supply chain, the apparent breadth of corroboration is misleading. This axis measures the structure of the evidence ecosystem, never the truth of any claim inside it.
Feeds: Context
T
How recent is the evidence?
Temporal proximity
Whether evidence is close enough to the event in time, and whether enough time has passed for independent sources to verify rather than merely repeat the first report.
4. Independence scoring

Five sources, one confirmation.

Agreement only counts when sources arrived at a claim separately. Five outlets citing the same press release are one confirmation repeated five times — not five confirmations.

Who owns the sources?
Ownership clustering
Sources are grouped by common owner or funder before anything is counted. Shared ownership shows up as one voice, not several.
How far from the original report?
Origin-chain depth
How many hops of re-reporting separate a source from the original account. A claim traced to a single first report reads differently from one confirmed by parties who investigated independently.
Who gains if this is believed?
Beneficiary concentration
Whether the sources backing a claim are structurally positioned to benefit from it being believed. Measured by disclosed interest, not inferred motive.
5. What we measure, and what we refuse to

Structure, always. Opinion, never.

We measure
How many independent sources agree
Who owns or funds those sources
How far each source is from the original report
How the evidence is distributed over time
We refuse to measure
×Whether a source is biased
×Whether coverage is being suppressed
×Motive, intent, or character
×Anything not traceable to disclosed, structural facts

“This claim’s four corroborating sources sit under two ownership groups” is a fact nobody can contest. “This source seems biased” is a judgment that can be argued with. ASURIQ reports the first kind and never the second — the vocabulary is the boundary, not a courtesy.

6. Fact-check integration

Third-party signals, kept separate.

When established fact-checkers like Snopes or PolitiFact have already reviewed a claim, ASURIQ surfaces their finding alongside the badge — but never bakes it into the score.

SnopesIndependent fact-check
False
“The Great Wall of China is not visible from space with the naked eye”
WikipediaMyth detection
Common misconception
Wikipedia lists this as a widely held misconception

Fact-check results appear as a separate callout card in the detail view, visually distinct from ASURIQ’s own scoring. They are independent data, queried from the Google Fact Check API and Wikipedia in real time. A 60% word-overlap relevance gate ensures only genuinely related fact-checks are surfaced. ASURIQ never outsources its verdict — third-party findings are context, not authority.

7. Coverage

What we reach. What we don’t.

Every free check queries 269 databases automatically. Detail cards add 3 commercial sources for a fuller ownership and sanctions picture.

Science and health
PubMed
ClinicalTrials.gov
FDA
WHO
Europe PMC
arXiv
Financial and regulatory
SEC EDGAR
FRED
BLS
World Bank
Government and legislative
Congress.gov
DOI Foundation
Reference and general
Wikipedia
Wikidata
CrossRef
OpenAlex
GDELT
Google Fact Check
Commercial (detail card only)
OpenCorporates
OpenSanctions
Nominatim/OSM

Where no adequate source exists for a domain, ASURIQ says so rather than guessing. “We cannot verify this from public sources” is not the same as “this is unverifiable,” and neither is the same as “this is false.” Absence of evidence is reported as a coverage gap, never treated as evidence of absence.

8. Our own sources

Who verifies the verifier.

The same independence question we ask of every claim, asked of ourselves. Most of the databases above are not owned by any single commercial interest: SEC EDGAR, FDA, WHO, BLS, and Congress.gov are government bodies. PubMed, Wikipedia, Wikidata, CrossRef, arXiv, Europe PMC, and OpenAlex are run by public agencies, nonprofits, or academic institutions. That structure is published here, not asserted — the same ownership-transparency standard applied to ourselves first.

10. Versioning and citation

This page is a version, not a live document.

Methodology changes are versioned, not silently edited. This page states which version was in effect for any certificate that cites it.

ASURIQ Methodology v1.1. Silent Ampersand LLC. Effective 11 August 2026.
https://asuriq.dev/methodology
Your data stays yours
Prompts stay with your provider. We see analysis metadata only.
Keys never stored
One-way hash for authentication. Your credentials pass through. Never persist.
~2 second responses
Single-model cognitive tools return in about 2 seconds.
37 of 100 flagged
We ran 100 ChatGPT answers through verification. See the study →

See it applied to a real record.

Every verdict links back to this page