Release notes

Version 4.7

About Release

• Release date: September 11, 2026
• Compatible platform version: Work.AI 10.3.0+

The Evan v4.7 release introduces more fine-grained Manual Task access control, a new Google (Site Search) input data provider, a decision reapplication step in the Core Business Process (BP), and support for file ingestion across providers. The release also replaces the Data Purge BP with an SQL-based purge mechanism, upgrades article parsing with Defuddle and Trafilatura, and adds configurable review requirements and autogenerated investigation comments. Other notable improvements include ML model enhancements, LLM fallback processing, support for large-batch investigations, refined crime category detection, and clearer visibility into AI-assisted dispositioning.

Features

• Evan now lets you configure whether specific user groups should be allowed or restricted from viewing and handling Manual Tasks within AMM BPs. The new feature eliminates the need to create new schemas, use hardcoded values, or carry out additional development for each new BP.

  Permission assignment is done using the Show Manual Review task only for user groups setting in the configuration interface or the _sys_user_groups input data field, the latter taking priority when both methods are used simultaneously.

• The list of data providers for sourcing input data has been extended with the Google (Site Search) provider, giving you a way to search only within a defined list of websites. The provider inherits all settings from the existing Google API provider, except screening keywords, partial name search, and location search. In addition, it features the new Search web-sites field where you can specify or upload exact websites or patterns for search.

• The Core AMM BP includes the new Decision Reapplication step that executes before investigations are split into articles. Where a prior decision reapplication is available for an entity, AMM reuses that decision and skips the call to the ML model. The enhancement reduces redundant processing and ensures that previously applied decisions are reused consistently in relevant investigations.

• Evan now supports the execution of the File Ingestion Business Processes with providers beyond External Source, enabling more flexible integration with third-party data sources.

  The CSV File Monitoring settings section is now always visible in the configuration interface, regardless of the selected provider. The change makes it easier to configure and review file monitoring settings across providers.

• The data cleanup approach has been changed from a mechanism based on the dedicated Data Purge AMM Business Process to the Work.AI-native SQL-based purge. The update provides a more scalable and reliable way to remove obsolete data while simplifying purge operations.

• The Poslight parser, which is no longer supported, has been replaced with Defuddle, while Trafilatura is now used as the primary parser. The enhancement reduces parsing failures and supports more accurate downstream screening and analysis.

• Evan now allows displaying additional information in manual review tasks in Workspace as a collapsible section below the article content, helping experts to review article details more efficiently and with greater convenience. You can switch the section display with the Show/Hide additional information toggle and navigate it using independent scrolling.

• The new Do you want Evan to automatically generate comment on close investigation popup? option has been added to the Human in the Loop step in the configuration interface. When the option is selected, Evan automatically generates a comment when you close an investigation, reducing manual effort and assignment review time.

• You can now set Evan to require Manual Task reviewers to specify review reasons and decision explanation comments at the article level or when closing investigations. The dedicated settings are available on the Human-in-the-loop step in the AMM configuration interface:

  • If the Require Review Reason when closing an investigation? and Require Decision Explanation when closing an investigation? options are set to Yes, reviewers cannot save or close an investigation until the review comment and decision explanation are provided.

  • If the Require Review Reason for article decisions? and Require Decision Explanation for article decisions? options are set to Yes, reviewers cannot save or close an investigation until review comments and decision explanations are provided at the article level.

Improvements

• To improve screening accuracy and content analysis, enhancements were introduced to the ML model. The updated model uses a Cross Encoder for improved focality detection, provides more reliable extraction of Moody's article data, and enhances name extraction and matching.

• The AMM Batch Processing BP now supports large batches of investigations more reliably. When merged investigation results exceed RabbitMQ message size limits, AMM can use a configurable reference output mode that stores the merged results in an object storage, such as S3, and returns a URL or a file pointer instead of the full payload.

• Articles dispositioned by a Large Language Model (LLM) now display the Auto (AI Assisted) label as the reviewer, while those processed by the original ML model continue to display Auto. The update provides better visibility into how disposition decisions are generated across Manual Tasks and reports.

• To improve processing reliability and reduce disruptions caused by external AI service issues, a fallback mechanism was introduced for article disposition and age detection. When an LLM API call fails, processing automatically continues using the original ML model. The failure is also recorded in the event log for monitoring and troubleshooting purposes.

• The crime category matching logic has been updated to make crime type detection more consistent:

  • The Other crimes category has been removed from the configuration interface, and the Escalate article if no crimes were detected checkbox is now enabled by default. The resulting logic is as follows:

    • If an article is ML-adverse and a configured crime category matches, the article follows the standard AMM processing flow.

    • If the article is ML-adverse but no configured crime category matches, escalate_article_with_no_crimes = true results in the Needs Investigation status.

    • If the article is ML-adverse but no configured crime category matches, escalate_article_with_no_crimes = false results in the False Positive decision.

    • If crime detection is not applicable or no adverse signal is present, the remaining standard disposition logic applies.

  • The AMM model input contract has been updated by removing the detect_crime_types Boolean field.

  • Spanish crime type categories are now supported in model configuration, crime classification, and configuration interface settings.

  • Crime types are now also recognized by the Evan LLM model.

• For AMM Manual Tasks opened outside Workspace, the Save and Continue and Close Investigation actions are now disabled. The improvement prevents unintended changes while keeping investigation data accessible for review in the read-only mode.

• The Redirect articles filtering policy has been renamed to Articles with redirect failed to make its purpose clearer, helping you identify and review articles more efficiently.

Bug fixes

• Resolved the issue when adversity scores were not consistently aligned with the focal party in processed articles. Now, party names are passed as explicit scoring context with name focality weighting and suppression of articles where the entity is not a primary subject.

• Fixed the issue when the Quality Control (QC) Business Process created empty Manual Tasks for investigations where all articles were excluded. Now, Manual Tasks are created only when reviewable articles are available.

• Resolved the issues causing the QC BP to fail with the following exception: IllegalStateException: Variable 'task_result' has wrong type STRING after Manual Task submission.

• Fixed OpenAPI schema validation failures caused by duplicate path shapes for the start-record-raw and get-record-result calls and by missing item definitions in ArticleResponseDto.matchedEntities.

• Resolved the issue when minimized LLM-generated summaries displayed additional content after the ellipsis. Now, truncated summaries are shown correctly.

• Eliminated the issue when Evan failed to identify articles containing Terms & Conditions or privacy disclaimers as False Positives. Now, these articles are classified correctly.

• Resolved the issue when LLM summary generation failed due to a JSON parsing error when the gpt-5-nano model returned empty content. Now, summary generation is handled correctly in these scenarios.

• Eliminated the problem when the decision reapplication mechanism treated null and empty field values as different, incorrectly detecting changes in investigation data.

• Resolved the issue where reopened AMM investigations lost their original additional_information content. Now, the reopening message is prepended to the existing information instead of replacing it.

• Eliminated intermittent split per article failures on large batches (500) by adding guard rails and improved error logging.

• Fixed the issue when the Force False Positive Disposition restriction was bypassed, causing investigations to be closed without the disposition enforced or stored.

• Eliminated the LLM article disposition failures when the gpt-5-nano model was used.

• Fixed the issue when the Articles that have fewer than 50 words filtering policy was displayed in the configuration interface, even though the Automatically escalate articles that have fewer than 50 words? is set to No.

• Fixed the display and selection issues with the WC1 API key secret component that occurred when Google was selected as the search provider.

• Resolved the issue when the File Ingestion Business Process rejected files containing only the content_link field. Now, validation accepts files that include either article_url or content_link.

• Eliminated the Google translation 400 Bad Request error.

• Resolved the issue when the manualTaskUrl field in webhook notifications failed to open the configured custom Workspace queue. Now, the URL correctly directs you to the intended filtered queue.

• Resolved the deserialization failure caused by the matchedEntities field generated as List<String> in the evan-amm-api client instead of List<MatchedEntity>.

• Resolved the issue when the Collect articles step in the Core AMM BP grouped duplicate articles only from the first similarity group returned by the NLP model. Now, duplicate pairs from all similarity groups are processed.