WritfindFindings, not flags.

Our approach

How a report is built

  1. 01

    Screen public sources

    We search for adverse information within the investigation scope, screen the results and retrieve relevant pages. Any use of a search snippet is marked.

  2. 02

    Investigate and write by matter

    Related events are written together in one chapter. Where further research could change the report, we look for evidence about unresolved outcomes or conflicting accounts.

  3. 03

    Link findings to source passages

    Before a report is assembled, quoted passages are checked against the saved source text. A text match does not by itself prove that a claim is supported.

  4. 04

    Make the chronology visible

    Key dates from the report form a timeline. Each selected date leads back to its sentence and the matter it belongs to.

Reports distinguish events from the dates they were reported or recorded. Information is limited to the source-reading date shown in the report.

Reading the report’s status labels

Category results use Flagged, Clear and Undecided. Unavailable content and unread sources are marked separately.

  • FlaggedAdverse matters were found.
  • ClearThe category search was completed and no adverse matters were found.
  • UndecidedThe category has not been fully assessed.
  • Not surveyedThe relevant report content or source text is unavailable.

The nine categories

Reports show the search result or assessment status for each category.

  1. Fraud
  2. Securities violations
  3. Money laundering
  4. Sanctions
  5. Bribery and corruption
  6. Organised crime
  7. Tax evasion
  8. Terrorism financing
  9. Other adverse matters

How the benchmark scores a report

Reports on three companies were compared using the same investigation scope. A model assessed them without tool names; contested calls were settled by hand. The published comparisons include three graded Writfind report versions. Each Deep Research report was generated once.

Recall
Matters found divided by the reference set. The reference set is every adverse matter any of the three tools reported, pooled, deduplicated and admitted only if a cited page supports it.
Precision
Matters accepted divided by matters accepted plus matters rejected. A reported matter is rejected when it is not adverse, is about a different entity, or falls outside the brief.
F1
The harmonic mean of precision and recall, combining coverage of the reference set with the proportion of reported matters accepted.
Evidence support
Evaluated passages whose central claim is supported by a cited source, divided by all assessable passages. The central claim covers the event, parties and outcome; this does not mean every detail is supported. Passages with no readable cited page, and passages describing only the investigation, are excluded.
Unsupported passages
An evaluated passage whose central claim is not supported by any readable cited source.
Passages without citations
An evaluated passage about the subject that gives no source citation. Passages describing only the investigation are excluded.
Unreachable source URLs
A cited URL that returned no page when the grader tried to read it.
Sources marked unread
A cited source marked as unread in the sample report, with only its search snippet available. This count was not recorded for the Deep Research reports.