Adverse media programmes break in three ways that compound until analysts stop trusting the queue entirely.
A subject monitored as "Tariq Al-Rashid" is mentioned in an Arabic-language article as "طارق الراشد" and in a Russian wire as "Тарик аль-Рашид". String-match tools return zero hits on both. The adverse mention goes undetected, and the programme owner finds out from the regulator.
A monitoring tool fires a score. What produced it? Without the source passage, mention offset, and confidence breakdown in the alert payload, an analyst cannot assess the hit, a reviewer cannot override it, and no audit trail exists for the decision that follows.
Monitoring a common name across a broad corpus without coherence filtering produces thousands of alerts per day, most of them for different people. Analysts learn to ignore the queue. The one genuine adverse mention arrives and goes unactioned for the same reason.
Media mentions are resolved to monitored entities across Arabic, Cyrillic, Hebrew, CJK, and Latin — without translating names at index time. Transliteration variants, partial names, and initialised forms are handled at retrieval time, preserving cross-script recall without introducing ambiguity.
Register a standing query at onboarding for each monitored subject. As the corpus ingests new media — news wires, regulatory filings, court records — the query evaluates automatically and fires an alert only when the resolved mention matches the operator-defined query predicate and clears the coherence threshold.
Each alert ships with a pinned graph_version. Opening a session against it returns the source document, publication date, passage excerpt, passage_offset, and the per-component scores that resolved the entity. An analyst sees exactly what was said, where, and how confident the resolution was — before deciding whether to act.
Coherence filters out mentions of different people who share a common name before an alert is raised. Only mentions free of blocking evidence — wrong geography, conflicting role, mismatched dates of birth — pass the threshold and enter the analyst queue.
A standing query registered once fires automatically whenever a resolved, coherence-cleared adverse mention arrives for the monitored subject. No batch job, no polling, no re-running the workflow.
Each alert fires via webhook with its graph_version, merge_confidence, and resolving_attributes. Opening a pinned session against that graph_version returns the source passage. The analyst sees the exact text that triggered the alert, the passage_offset in the document, and the per-component scores that resolved the entity — before making any decision.
"Per-mention provenance so a triage analyst sees the source passage — not just a score."
Entity "Tariq al-Rashid" resolved to cluster #C-8841. Standing query #AMQ-2291 registered against the adverse-media corpus — financial crime, regulatory enforcement, court proceedings.
Arabic-language article ingested. Entity mention "طارق الراشد" extracted and resolved to cluster #C-8841 with merge_confidence 0.91 — via cross-script retrieval, no index-time transliteration. Operator query predicate matches "خضع للتحقيق" (subject of investigation).
Query #AMQ-2291 re-evaluated against the new mention. Alert #ALT-7714 delivered via webhook: { alert_id, subscription_id, mission_id, fired_at, query_summary, graph_version: "v_20260512_0842", merge_confidence: 0.91, merge_coherence: "likely", blocking_evidence: [], resolving_attributes }.
session_open called against the alert's graph_version: v_20260512_0842. Evidence call returns source document, publication date, passage excerpt, passage_offset { start: 412, end: 501 }, and source_summary. Source passage now in hand.
Analyst reviews the source passage, confirms the mention refers to the monitored subject, and records the adjudication. Decision: ESCALATE TO COMPLIANCE. The source passage and resolution evidence are attached to the decision record — permanently.
Cross-script entity matching at query time. Names are stored in the script they arrived in; transliteration variants resolve at retrieval.
standalone librarySource passage, passage offset, and per-mention evidence returned by the evidence call. Responses are stable within a pinned session.
Continuous reshaping as media evidence arrives. Cluster boundaries update as new mentions resolve — without destructive merges.
Open an investigation with session_open against the alert's graph_version. Cluster boundaries update without destructive merges; adjudication decisions remain reproducible within retention.
A demo runs against a sample of your monitored subjects and a representative media corpus, and returns resolved adverse mentions with the source passages your analysts would see in production.