Methodology
Two Levels of Confidence
Every event on Ukraine Truth is assessed at two levels. These serve different purposes and should not be confused:
Event Corroboration (primary)
How many independent, reliable sources report the same event? This is the prominent indicator shown on every event page. It is computed algorithmically from the source registry and cannot be influenced by any single source's framing.
Source Tone (secondary)
How does each individual source frame its reporting? An AI model reads the source article and classifies the language as certain, likely, uncertain, or speculation. This reflects the outlet's own editorial confidence — not whether the claim has been independently verified. A source can report something with "certain" tone and still be wrong.
Event Corroboration
Each event's confidence tier is determined by a weighted reliability algorithm. Every source in our registry has a reliability rating (high, medium, variable, or low). When multiple independent sources report the same event, their reliability weights are summed to produce a corroboration score.
Verified
Weighted score ≥ 3.0. Multiple high-reliability independent sources confirm the core facts. For example, if Kyiv Independent, Ukrinform, and Ukrainska Pravda all report the same event, it receives Verified status.
Likely
Weighted score ≥ 1.5. At least two independent sources agree on the core facts. High confidence but additional sources would further strengthen the assessment.
Contested
Sources conflict on key details. We present the competing accounts and let readers evaluate.
Uncorroborated
Single-source reporting with no independent corroboration. Most events (~92%) begin at this level and may be upgraded as additional sources are matched through our pipeline. We include these when the source is credible, but clearly label the uncertainty.
Debunked
Multiple trusted sources have contradicted the claim. We keep debunked events visible so readers can see what was disproven and why.
Reliability Weights
High = 1.0, Medium = 0.5, Variable = 0.3, Low = 0.0. State-controlled media (e.g. TASS, RT) always contribute zero weight regardless of reliability rating.
Source Tone
Each claim on an event page includes a "source tone" label. This is generated by an AI model reading the source article and classifying the language the outlet used. It is not an independent verification — it reflects how the journalist or outlet itself framed the information.
A "certain" source tone does not mean the event is verified. An event with a single "certain" claim from one outlet is still "uncorroborated" until independent sources confirm it.
Structural Source Badges
We classify sources by structure, not ideology: Independent Newsroom, State Agency, Think Tank, Independent Newsletter, Aggregator, or Commentary Platform. The badge explains what kind of institution is publishing the claim.
We avoid ideological labels in source scoring. Reliability outcomes are based on measurable performance in our claim database.
Performance Metrics & Formula
Source performance is computed from claim-level data: corroboration rate, contradiction rate, time to confirmation, and upgrade/downgrade frequency.
Public Formula
Performance score = 0.50 × corroboration rate + 0.35 × (100 - contradiction rate) + 0.15 × confirmation-speed score, normalized to a 30-point scale.
Corrections, Reviews, & Appeals
Ratings are reviewed quarterly, after major correction events, and upon verified source appeal. All appeals must include documented evidence.
When new information contradicts a previous assessment, we update the event confidence tier and preserve the correction trail. We do not silently alter records.
What We Don't Do
- We don't editorialize. We present sources and let readers decide.
- We don't publish operational military details that could endanger lives.
- We don't claim certainty where none exists. "We don't know" is a valid answer.
- We don't treat all sources as equal. A wire service with correspondents on the ground is not equivalent to an outlet with low transparency and weak corroboration performance.