Release notes

Version 4.0.3

About Release

• Release date: April 3, 2026
• Compatible platform versions: Work.AI 10.3+

The Tara 4.0.3 release focuses on improving rule inputs, matching quality, data processing resilience, and model behavior. It also introduces updates to billing events, historical data upload, and adjudication logic.

• The rule object now includes all processed sanctioned entity names, including the watchlist entity (WLE) name, primary name, and alias, along with their corresponding name matcher scores, enabling more precise custom rule authoring to increase the automation rate.

• To prevent individual rule failures from interrupting Business Process execution, runtime exceptions during rule processing are now caught at the hit level, returning a NO_DECISION outcome with a failure narrative.

• A new Rare and Screened token overlap mode gives you stricter control over name matching and helps reduce false positives in high-risk workflows. The new Rare and Screened option is available in the Token overlap configuration field in the configuration form and is propagated through the Business Process to the Decision Model.

• Billing events now contain additional AI Agent metadata, including the AI Agent variation, name, and skill, which improves billing accuracy and helps ensure compliance with billing requirements.

• To make historical data ingestion more flexible and reduce failures caused by missing fields, continuous learning (CLDS) historical data upload provides improved support for Firco v6 and similar screening systems:

  • Added optional hit-level analyst decision fields (T_IS_RELEASED and T_DECISION_TYPE), so the upload process can preserve more granular review outcomes.

  • Added optional hit-level tag content (T_TAG_CONTENT), while still allowing Tara to derive tag content from the message when available.

  • Made the T_MESSAGE, T_NATURE, and T_CONTENT_FORMAT columns optional, so missing or blank values in these fields no longer cause a Business Process to fail.

• Payment information, such as payment amount, currency, and transfer direction, is now extracted from FUF and SWIFT content, enriching alert context and enabling more accurate rule-based adjudication.

• Data cleanup removes commonly used placeholder values from alert data during both CLDS historical data upload and real-time processing to prevent placeholder values from interfering with CLDS compatibility and match accuracy. By default, the cleanup logic deletes values such as Unknown, N/A, and *, replacing fully matched placeholder content with empty values before further processing.

• When the tag content is a substring of the screened address, Named Entity Recognition (NER) processing uses the full address as tag content to provide richer context and improve location entity prediction accuracy. The enhancement is designed for cases where the original tag content is embedded in a longer address and would otherwise not provide enough context for reliable location detection.

• To reduce incorrect location validation caused by partial or misleading address matches, geocoder lookup now contains two configurable location validation modes. The Strict option marks country and address as valid only when the full address is confirmed, whereas Approximate allows country validation for partial area-level matches.

• The rule object provides more complete and accurate names and name matcher scores, resulting in more reliable inputs for custom rule design and adjudication:

  • Added all available name matcher similarity scores for screened, lookup, and NER data compared against sanctioned entity names, primary names, and aliases.

  • Restored original token spacing when processed names are concatenated (for example, "lesterkub" → "Lester Kub"), which makes matched names easier to interpret in rules.

  • Changed score and name selection to follow input source preference (screened data first, then lookup data, then NER data) rather than the highest score, ensuring more consistent rule inputs.

• Synthetic data generation has been updated to v34, incorporating refactored generation methods and updated screening scenarios.

• Adjudication rules cover more client-specific scenarios to improve adjudication accuracy and increase the automation rate:

  • Used the NER source type when replacing the NAME type.

  • Added new textContainsAnyWleNames(), isNameInTheMiddle(), getMessageContentTagValue(), and isNamesEqual() rule functions to support more flexible rule logic.

  • Refined source handling and matched WLE logic to treat names as primary names when primary names are absent.

  • Added logic to update the sanctioned code map when a sanctioned name resembles an ID, improving downstream rule evaluation in edge cases.

  • Enhanced rule-processing stability by preventing null pointer exceptions on empty decision reasons when building a decision object.

• To prevent unbounded table growth in production environments, the pss_reapplication_hashcodes_v4 Data Store includes the decision_date column and supports configurable data purge to automatically remove records older than the defined retention period.

• Hit entity prediction delivers more accurate and explainable results, reducing incorrect hit text selection and unnecessary ambiguity in alert analysis:

  • Replaced the fuzzy matching score of hit text with a character matching ratio to improve the precision of direct text comparisons.

  • Introduced a preprocessing step that extends partially matched hit text to the nearest token in tag content to better identify direct matches.

  • Reduced unnecessary text splits by targeting specific hit text boundary handling and filtering ambiguous words.

  • Reduced the number of unknown hit entities, so this outcome occurs only when the compared names have no meaningful overlap.

• The Decision Model is upgraded to v6.0.9 with Name Matcher v1.2.3 and includes the following improvements:

  • Fixed an unhandled exception for the AN/ANT (Netherlands Antilles) and CS/SCG (Serbia and Montenegro) input country codes. Legacy country codes that were later split are now converted to the corresponding new country codes (for example, SCG → SRB and MNE) and matched against any matching parts in the results.

  • Fixed the issue where aggregated_features names were blank when screenedData.name was not provided. Processed WLE names can now be taken from nerData or securityData.

  • Added original and processed names from screenedData, securityData, and nerData to the output for use in rules.

  • Updated libpostal output to use the ISO country code and city extracted from the identified SWIFT address.

  • Enhanced organization matching for names with legal forms in different positions. For example, SUEKINO AG now matches INTERNATIONAL LIMITED LIABILITY COMPANY SUEKINO LTD.

  • Improved legal form handling by removing consecutive legal form abbreviations (for example, Company LLC Inc).

  • Improved handling of duplicate legal forms. Names such as Company LLC Holdings LLC trigger truncation at the first occurrence for cleaner matching.

  • Preserved legal form abbreviations followed by ambiguous company terms, such as AB Group or KS Holdings, to avoid overstripping meaningful name parts.

  • Improved handling of legal forms at the beginning of names to prevent incorrect text removal.

  • Enhanced organization matching for quoted company names within legal form wrappers. For example, International Joint Stock Company 'NAME Holdings Limited' and NAME HOLDINGS LIMITED are now treated as the same entity.

  • Improved handling of names with apostrophes and similar characters. Names such as O'Brien correctly match O Brien by normalizing quote-like characters to spaces.

  • Added support for alias indicators in names. Names containing phrases like t/as, trading as, aka, also known as, dba, doing business as, formerly, or f/k/a are automatically split, and both parts are matched separately. For example, ABC Corp t/as XYZ Services matches against both ABC Corp and XYZ Services.

  • Enhanced legal form comparison to properly handle organization names with multiple legal forms. Previously, only single legal forms were compared. All combinations are now checked for equivalence.

• The NER model is upgraded to v3.2.2 with the following improvements:

  • Enhanced securities, fund, and stock name recognition through targeted model fine-tuning to reduce false positives and incorrect NER predictions.

  • Improved organization detection by preventing ISO country codes, such as AE and AG, from being treated as business entity legal forms, reducing false ORGANIZATION classifications for LOCATION and PERSON entities.

  • Refined organization merger logic to skip merging adjacent entities when the leading legal form matches an ISO code, preserving correct entity types in mixed address and company text.

  • Improved entity classification accuracy for entities containing location markers (for example, street types, address abbreviations, known city, or region names) even when they are initially misclassified. This reduces false positives in text with many addresses.

  • Expanded detection of location-like patterns across both ORGANIZATION and PERSON entities for more reliable disambiguation between people, companies, and addresses.

  • Updated reference data to improve recognition accuracy.

Bug fixes

• Resolved the issue where the reason narrative for a single merged hit incorrectly included the SPLIT marker.

• Fixed the null pointer exception when required data was missing during time difference calculation in the PSS QC Report Business Process, which prevented the quality control report from being generated.

• Resolved the issue when a Business Process failure was routed to the Apply rules step while the rule object was missing, causing processing to get stuck.

• Fixed the null pointer exception in DroolsService that occurred when creating a container with null rules content.