Methodology

Principles

• The Trust Health Score is a deterministic function of classified public evidence and a published formula version. The same inputs and the same version always produce the same number.
• An AI model classifies and summarises. It never produces, adjusts, or overrides the score.
• Coverage Confidence is computed separately and reported separately. It is never folded into the score.
• Where coverage is Insufficient, the numeric score is suppressed. It is not estimated, and it is not shown with a caveat.
• Every input to a score is retrievable as evidence, with its source and collection date.

Inputs

Public items about the configured clinic and its locations, collected from review platforms, public social posts, public search results, and public answer-engine responses. Each item carries its source, its publication date, its collection date, and its language.

Classification

Each item is assigned to one or more themes from a clinic taxonomy reviewed in English, Turkish, German, and Arabic, together with a sentiment value and any applicable risk signals. Classification is applied per item and is retained with the item, so a change to the taxonomy is visible as a re-classification rather than a silent shift in the number.

Scoring

The score aggregates classified items over a defined window, weighting by recency, source reliability, and theme severity. The exact weights belong to the formula version, and every score displayed in the product is labelled with the version that produced it. A change of version is a documented event; historical scores are never silently recomputed under a new version.

Status: the version 1 weights and thresholds are being finalised with the product team and will be published in full here before the score is shown to any customer. Until they are, this page describes the framework rather than the constants.

Coverage Confidence

Determined by explicit thresholds over the volume of evidence, the number of distinct sources, the recency distribution, and the language coverage relative to the configured patient markets. It reports one of a small set of states, the lowest of which suppresses the score.

Risk states

A clinic is in exactly one of Calm, Watch, Alert, or Crisis. Transitions are driven by rules over the deterministic measurements, never by a model, and every transition records the signal that caused it.

Comparison

Competitor comparisons apply the same formula version, the same window, and the same coverage rules to every clinic in the set. Where a competitor’s coverage is insufficient, the comparison is reported as uncomparable rather than treated as a zero.

Known limits

• It measures what the public record says, which is not the same as clinical quality. It must never be presented as a quality rating.
• Public platforms are not evenly distributed across patient markets, so coverage will differ by market — which is why coverage is reported.
• Verinten does not verify whether a public statement is true.
• Classification is accurate, not infallible. Every theme opens onto its evidence so a misclassification can be seen and reported.

Reporting a problem with the methodology

Send a description and, where possible, the evidence identifiers concerned through the contact form. Methodology corrections are versioned and recorded in the change history below.