Start with the source.
Reporting cites the exact source revision it used. Original language, translations, retrieval times, and timestamped transcripts remain connected. When source material changes, earlier reporting keeps its original evidence trail.
More than one model. More than one view.
Screening and scoring consider multiple configured models. We retain individual outputs, the actual provider and model, the rubric, and the source revisions. Numeric scores use the median of comparable dimensions; spread and categorical disagreement stay visible. Failed or incomplete analysis never counts as approval.
Context before labels.
Bias, tone, and framing are assessments of specific material within a stated context. Aggregates need sample counts and time ranges. Country-specific political taxonomies are versioned and are never treated as interchangeable global scales.
Editorial responsibility stays human.
Staff approve public assessments and AI-assisted reporting. Trusted contributors can publish passing original submissions to the community feed. Featuring remains a separate editorial decision with a recorded reason.
A clear record of change.
Published revisions, corrections, source changes, model runs, ranking decisions, and moderation actions carry history. Private drafts and internal moderation data remain permission controlled.
Prominence has an explanation.
The initial quality policy weighs sourcing at 35%, originality at 25%, clarity at 20%, and freshness at 20%. Freshness halves every 48 hours. Incomplete scoring uses chronological community placement.
Current published methodology
{
"bias": "Country-specific political taxonomies and framing, tone, sourcing and stance are separate dimensions. Outlet and writer aggregates show sample counts and model disagreement; scales across countries are not equivalent.",
"scoring": {
"aggregation": "median of comparable dimensions; disagreements retained",
"clarity": 0.2,
"freshness": 0.2,
"freshness_half_life_hours": 48,
"originality": 0.25,
"sourcing": 0.35
},
"screening": "Two distinct model identities must pass the exact revision. Disagreement or unavailable inference requires human review. AI drafts always require an editor.",
"source_policy": "Original evidence is retained by revision. Public citations point to exact observed versions. Translations and machine analysis identify their actual provider, model and input revision.",
"version": "1"
}