How a rating is produced, end to end. Published so that any determination can be checked against the process that generated it.
An identified crawler retrieves the vendor's published legal and policy documents once per source per week. robots.txt is checked before every request and there is no override. A disallowed document is read manually by a person or not at all. Every retrieval stores the source URL, a timestamp, and a SHA-256 hash of the normalised text.
Only documents the vendor has itself published are assessed. Prior research, product changelogs, press coverage, and marketing material are out of scope.
Each criterion names the document types it requires. Criteria are grouped by
those requirements and each group is assessed against only the documents it
needs. Where a vendor publishes no document of a required type, the criterion is
recorded as NOT_PUBLISHED and scored 0 — a determination, not a
gap in the assessment.
Where a vendor maintains separate viewer-facing and organization-facing terms, only the organization agreement can satisfy an organization-facing criterion. A favourable term in individual-subscriber terms is disregarded.
Each criterion is assessed against the corpus and the published scoring conventions. Every determination records the score, a verbatim supporting quotation of twelve words or fewer, the reasoning, whether the finding was expressly stated or reached by inference, and a computed confidence value.
Confidence is computed by Truefield from seven observable conditions, not self-reported. Four are mandatory: the correct controlling document was identified, the relevant passage was retrieved, the evidence supports the finding, and no unresolved contradictory passage was detected. Failure of any one prevents an automatic determination.
Determinations are banded. A criterion validated for automation, with all mandatory conditions cleared, stands automatically and is sampled for audit. A determination resting on approved precedent stands automatically at an elevated audit rate. Novel interpretation, conflicting evidence, suspected manipulation, an appeal, or a conflict of interest routes to human exception review.
Automation eligibility is earned per criterion, per Standard version, by measured agreement between independent assessments — never assumed. A materially changed criterion returns to review until revalidated.
Each vendor receives its complete rating with the underlying evidence at no charge, not fewer than ten business days before first publication, with an invitation to identify factual errors and submit additional documents. One five-business-day extension is available for good cause. Publication follows whether or not the vendor responds, and whether or not it has paid Truefield anything.
Every published rating identifies the documents relied on, their sources, the retrieval date, and the Standard version applied. Absence of a document is reported as non-location as of a stated date, never as non-existence. Scope is disclosed on the face of the rating. No rating characterises a vendor's conduct, motives, or competence, and none states or implies that a vendor has violated, complied with, or failed to comply with any law.
Correction addresses objective factual error — a wrong URL, a superseded document, an extraction error — and is handled by expedited reassessment. Appeal addresses the application of the Standard and is decided by a reviewer other than the original decision-maker. Both are free. A corrected entry carries a dated public record of what changed.
Every snapshot assessed, with its hash and retrieval date, and every determination with its reasoning and confidence, is retained for seven years from the later of the determination, the expiry of any certification, or the final resolution of any related dispute. Any published rating is reproducible from primary evidence years afterward.