Methodology and limitations

Four mutually exclusive content types, with a separate overlay for integrity concerns.

• NNewsNews — Reported news: event-driven, sourced, written in a reporting register.
• EEditorialEditorial — Editorial or opinion: argument, advocacy, commentary or analysis with a stated viewpoint.
• AAdvertisingAdvertising — Advertising, sponsored or promotional content, including affiliate and branded material.
• UTUntrustworthyUntrustworthy — Untrustworthy overlay. Human review only in Phase 1 — never assigned automatically.

Orientation bands describe where an item's framing sits relative to the political centre of the publisher's own jurisdiction — not a single global or US-centric baseline. An item scored conservative in one country may sit in a different band than a comparable item elsewhere, and that is the intended behaviour.

Orientation is never an input to reliability, source integrity, or manipulation risk, and orientation colours are deliberately muted so that no band reads as a warning.

Each classification stores a confidence value between 0 and 1. Items below the Phase 1 thresholds (0.75 for content type, 0.60 for orientation) are routed to a human review queue rather than presented as settled. Reviewer overrides are recorded in an audit log with the previous value, the new value, and the reviewer's stated reason.

The UT overlay is never assigned automatically in Phase 1. It can only be applied by a trained reviewer, and it will remain human-gated until the calibration gate for automated integrity signalling is met. This is a deliberate safeguard against automated defamation.

The platform stores headlines, publisher-provided summaries, canonical links, and the analysis it produces. It does not store or republish full article text. Every item links back to the original publisher.

A second pass extracts the checkable claims an item makes, records whether each is asserted by the outlet, attributed to a named source, predictive or a judgement, and attaches evidence. Evidence currently comes from cross-publisher coverage of the same event: independent outlets reporting the same claim is corroboration, not proof. Every evidence link is shown with the claim it supports, so you can follow the trail yourself.

Headline fidelity scores how faithfully a headline represents the text beneath it. Overstatement, dropped attribution, implied causation and unanswered question headlines lower the score. A low score is a statement about the headline, not about the outlet.

Source reliability is a running track record built from corroboration rate, sourcing quality and average headline fidelity, with confirmed untrustworthy items as a penalty. Political orientation is never an input to it. A conservative outlet and a liberal outlet with identical sourcing practice score identically.

The analysis pass can flag manipulation-risk cues, but a flag is a referral, never a rating. Nothing carries the untrustworthy designation unless a named reviewer confirms it with a written reason, which is recorded in the public audit trail. Framing is never a reason to flag an item.

Publishers can challenge any label, fidelity score or reliability figure through the appeals page. A reviewer answers every appeal in writing; where an appeal is upheld, the correction appears in the same audit log as any other override.

• English-language sources only; multilingual scoring arrives in a later phase.
• Classification reads headline and summary, not full article body, so nuance carried deep in a piece can be missed.
• Event clustering uses headline-token overlap, which under-groups events reported with very different wording.
• Evidence is limited to corroborating coverage from monitored outlets; primary documents and external record comparison arrive in a later phase.