Methodology
How the assessment works.
How AI Trust Scan turns public evidence into a TRUSTYCYBER Score, and where human-led independent assurance goes further. Every published result records the methodology version that produced it.
The pipeline
From public evidence to a defensible result.
The scan is automated, and every step records what it did, so a reader can check the result against its sources.
Resolve the entity and locate its public sources: AI disclosures, trust material, policies, and security and privacy documentation. The confidence of the entity match is recorded.
Read every source and extract claims. A claim without a source reference and an exact excerpt is rejected before it can influence anything.
Every claim carries an evidence class, from vendor-published through externally corroborated to registry-verified.
Application code, never the AI, computes the weighted dimension scores, the TRUSTYCYBER Score, the grade and the procurement recommendation, deterministically.
Findings, conditions and buyer questions, each linked to its evidence, with evidence coverage and confidence stated alongside the score.
The score
Eight weighted dimensions produce one score.
Every scan grades the same dimensions, with fixed weights per scope profile (vendor, product or organisation). A dimension that genuinely does not apply is excluded from the calculation rather than scored as a failure, and the score is always published with its grade, evidence coverage, confidence and assessment date. The score and the procurement recommendation are related but different: a well-scored organisation can still carry conditions that must be resolved in writing. Agentic AI systems are scored on a ninth dimension, Agent, with its own fixed weight, so an AI agent is assessed on agent-specific evidence rather than borrowed credit.
See a live report →Dimensions and weights (vendor profile)
- Organisation & AI governance: 15%
- AI system: 15%
- Customer data: 15%
- Security foundation: 15%
- Model & provider transparency: 10%
- AI supply chain: 10%
- Independent assurance evidence: 10%
- Legal & contractual: 10%
Evidence classes
Every claim records where it came from.
Each claim is labelled with how strong its source is.
Registry verified
Confirmed against an official third-party register, such as an accredited certification body’s public registry.
Externally corroborated
Stated by a source independent of the vendor, not merely repeated across the vendor's own pages.
Vendor published
Stated by the vendor on a vendor-controlled domain, with no independent corroboration found.
Third-party reported
Reported by an analyst, journalist or reviewer without independent verification of the underlying fact.
Reasoned inference
Not directly stated, but reasonably inferred from stated facts. Always flagged as inference, never presented as a claim.
Apparent contradiction
Two or more sources make claims that cannot both be straightforwardly true. Both are shown, with the conflict noted.
Not publicly evidenced
The topic was actively searched for and no relevant statement was found. Never used for topics that simply were not examined.
Stale or superseded
True as stated, but the source shows signs of being outdated relative to other sources or the retrieval date.
Technically observed
Seen directly in DNS records, certificate history or security headers rather than stated in a document. Collected by code, never assigned by the model, and kept separate from what the vendor says. A provider observed but named nowhere becomes a disclosure-gap finding.
Certifications
How certifications are verified.
A certification claim is walked up a verification ladder: claimed by the vendor, evidenced in a document, corroborated on the issuing registry, scope verified, and current. Each rung is checked and shown separately. A certificate that cannot be found on a registry is reported exactly that way.
The verification ladder
- Vendor claimed
- Evidence cited
- Registry corroborated
- Scope verified
- Currency confirmed
Limits
Every scan result carries this text, verbatim:
This automated result is based on publicly available information at the time of scanning. It is not an audit, certification, legal opinion or assurance engagement. Publicly unavailable evidence may materially change the conclusion.
The scan assesses publicly evidenced assurance posture. It does not test whether controls operate effectively, and it does not establish that a vendor is safe, compliant or suitable for a particular use. It is a single point-in-time scan of English-language public sources. It is not continuous monitoring, it is not a penetration test, and it never accesses authenticated content.
Where independent assurance goes further
When a decision needs more than public evidence, a human-led engagement examines evidence the public scan cannot see, tests it against defined criteria, and produces findings a qualified auditor determines and signs. The scan's open questions typically become the assurance scope. See assurance services or check your fit.
Versioning and independence
Scoring rules, weights, evidence classes and this page's disclaimer are versioned. Every published result records the methodology version that produced it, so an older report is read against the rules that made it. Scores cannot be purchased, and scan results are published on the same basis for every organisation assessed. Where objectivity cannot be maintained, work is not represented as independent.
See the methodology in a live report.
Run a free AI Trust Scan on a vendor, product or organisation. Every figure in the result links back to its evidence.