Methodology
The Wire's corroboration score is the central editorial instrument of the platform. Everything we publish carries a score; everything we decline to publish failed to reach one.
Every tip is parsed into a structured record consisting of asserted entities (organizations, individuals, monetary amounts, dates, locations), supporting evidence (documents, recordings, photographs), and a free-text narrative. New tips are matched against existing tip clusters using fuzzy entity resolution. A cluster is a set of two or more tips that assert sufficiently overlapping facts about the same underlying event.
The Six Components
The number of distinct submitters in the cluster. Submitters are de-duplicated using submission tokens, behavioral fingerprints, and content-similarity heuristics. A single submitter who tips twice does not count as two sources.
The structural variety of sources. A cluster containing one ministry insider, one external contractor, and one civil-society observer scores higher than three insiders from the same office. Diversity is computed across role, institutional position, and stated relationship to the facts.
The time between the earliest and latest corroborating tip. Spread that is too short suggests coordinated flooding; reasonable spread suggests independent witnesses surfacing over time. The function is non-monotonic — both very short and very long spreads are penalized.
The number of distinct submission origin regions inferred from non-identifying signals. Geographic spread is a corroboration signal because most coordinated influence operations originate from a narrow geographic footprint.
The presence, type, and verifiability of attached documents. Documents are scored on internal consistency, provenance signals (metadata, watermarks), and external corroboration. Photographs and audio are scored separately for tampering signals.
The confidence with which the cluster's asserted entities match across submissions. Two tips that both name the same ministry but disagree on the official, amount, and year will score lower than two tips that align tightly on all three.
Adversarial Controls
Burst detection. A sudden surge of tips alleging similar facts within a narrow time window is penalized rather than rewarded. Real-world events that legitimately produce burst tips are reviewed manually before scoring.
Near-duplicate detection. Submissions whose narrative text exceeds a similarity threshold to existing submissions are flagged. Identical or near-identical wording across "independent" sources is the most common signature of coordination.
Behavioral fingerprinting. Without logging IP addresses or persistent identifiers, we capture aggregate behavioral signals that do not identify individuals but can detect a single submitter posing as multiple sources.
Document-tamper detection. Images and PDFs are scanned for compression artifacts, metadata inconsistencies, and known forgery signatures. Documents that fail these checks reduce the documentary-evidence component to zero.
State-actor screening. We monitor for fact patterns consistent with known intelligence-service messaging templates. Submissions that match these templates are held for extended editorial review regardless of surface plausibility.
What the Score Does Not Measure
The score is not a probability of truth. It is a measurement of corroborating evidence under our weighting scheme. A finding scored 74% means "this asserts a single coherent fact pattern, supported by multiple independent sources with documentary evidence, and is unlikely to be a coordinated fabrication." It does not mean "there is a 74% chance this is true."
The score also does not measure legal admissibility, the editorial gravity of a finding, or its public interest. Those judgments are made by editors after the score is computed.