Release notes

Version 4.1.0

About Release

• Release date: June 10, 2026
• Compatible platform versions: Work.AI 10.3+

The release adds support for alert- and hit-level rules, enhances decision reapplication (CLDS), introduces email notifications for quality check (QC) and CLDS reports, and improves message content parsing, reporting, and billing.

• You can now manually disable billing by creating a dedicated pss_disable_billing secret entry in Secrets Vault. This also prevents billing-related error messages from cluttering the logs.

• A new IND (Individual) and ORG (Organization) ownership rule escalates cases where a sanctioned individual's name is fully contained in a screened organization name, helping identify potential ownership or control relationships between individuals and organizations.

• Time-based data collection in QC and CLDS reports is improved:

  • For the Past N days option, current-day records are excluded.

  • For the Custom date range option, a specific time range is now applied.

  • Time zone support is added. By default, the Control Tower UTC time zone is used. All database queries, QC reports, and CLDS reports use and display timestamps in the configured time zone.

• To avoid redundancy and the risk of inconsistencies, the derived object is removed from the rule model. Its data is redistributed to dedicated match objects (idMatch, addressMatch, countryMatch, and dobMatch), which serve as the authoritative sources during rule evaluation. As a result, Tara v4.1.0 is not backward compatible with previous rule files.

  • DOB matching now calculates and sets the match result based on the configured threshold, rather than relying on rule logic.

  • ID matching correctly sets the source (screened vs NER) and ID subtype directly on idMatch, with consistent use of pipe-separated values.

  • Address matching evaluates all available address sources (screened, NER, and geocoder-enriched data) instead of stopping at the first match and sets an address confirmation flag.

  • Country validation invalidates only specific countries that fail geocoder lookup, rather than clearing all country data when a single value is invalid.

  • Split hit entity resolution in the decision engine now uses entity data from match objects instead of the removed derived objects.

  • Rules referencing $hit.derived.*, for example, $hit.derived.subType or $hit.derived.countryConfirmed, are no longer supported and must be updated to use the appropriate match object.

• QC and CLDS report generation capabilities are extended to send report links to specific email addresses. Email notification settings are available on the Output step of the configuration form.

• NER document creation and result processing for multi-tag content are enhanced to prevent false escalations caused by incorrectly merged entities:

  • Tag content is deduplicated during document creation.

  • Entities of the same type are no longer merged across multipart tags during NER result processing.

• The rules framework offers clearer structure, flexibility, and greater control over decision outcomes:

  • Hit-level and alert-level rules are now separated into the hit_rules.drl and alert_rules.drl files, each with its own agenda group ("hit" and "alert"), ensuring better organization, predictable execution, and easier troubleshooting.

  • Alert-level decision comments are defined in alert_rules.drl, with the getStatisticalAlertLevelComment() method used to generate them. Alert rules populate the MessageDecision comment using the DecisionStatistic object.

  • In addition to standard decision statuses like RESOLVE and NO_DECISION, custom statuses, for example, PASS, HALF_PASS, L1, or L2, are supported for both HitDecision and MessageDecision.

  • You can now define custom hit and alert rule files in the Model step of the configuration form. When provided, these custom files replace the default rules, so you have full control over rule logic and decision outcomes.

• For easier understanding and maintenance, the Enrichment object is split into LocationEnrichment (external address and geolocation data) and EntityEnrichment (security and entity-related data). Backward compatibility is preserved as existing data is automatically converted with no impact on results.

• To reduce the risk of unintentionally releasing relevant hits, the decision reapplication provides enhanced blacklisted country detection. When enabled, hits containing a blacklisted country are no longer automatically released, and newly added sanctioned countries are consistently enforced across both new and previously processed data.

• CLDS now adapts based on recent analyst decisions, so outcomes are more accurate and consistent. When Update CLDS status based on most recent analyst decisions is enabled, Tara evaluates recent analyst actions to dynamically adjust how CLDS handles hits. Repeated Release and Resolve decisions can allow CLDS to release a hit even if earlier decisions prevent it, while repeated Escalate decisions can block automatic release. The Last N analyst Release/Resolve and Last N analyst Escalate configuration fields control this behavior.

• The Swift ADK parser is updated to handle message content where the same tag appears multiple times, which results in more accurate data extraction and fewer gaps in downstream processing.

• The FUF message content parser now supports partial tag matching. It identifies and extracts tag values even when tag names differ only by a numeric suffix, for example, ABC001 and ABC002. If no exact match is found, the parser attempts a partial match and, for multiple matches, combines values into a single result. The enhanced logic ensures more reliable tag extraction for both single and multiple tag occurrences and supports downstream processing such as NER.

• The message content parser is enhanced to identify tag content based on hit text when tag names are missing or do not match, ensuring more reliable data retrieval. This fallback mechanism applies when a single matching tag is found, enabling reliable parsing even with incomplete or inconsistent message formats.

  The parser also improves detection of sanctioned countries and cities in name fields by normalizing and cleaning name values and inferring missing location data where possible.

• The message type in the input contract is now optional and uses a String type instead of an Enum, reducing dependency on predefined values and simplifying integration with client data. Supported values include SWF and ISO_20022; others are deprecated. For backward compatibility, a default value of PAYMENT_TRANSFER can be used where needed.

• Billing Service integration ensures more consistent and accurate data handling. The service call uses the Java Native Worker BillingInfo object instead of BillingServiceInfo, aligning with the standardized data model. The Business Process variation title is now retrieved via controlTowerBpService.

• The AI Agent configuration is upgraded with the Decision model v6.1.0 and Name Matcher v2.0.1 and includes the following improvements:

  • Refined matching scores. 1.0 is returned only for exact matches after preprocessing, for example, "John Patrick Smith" vs "John Patrick Smith".

  • Normalized "Last, First [Middle]" name formats. Names containing a comma are converted into the standard order before comparison, for example, "Jackson, Corey A." → "Corey A. Jackson".

  • Add non-equivalent name pairs. Pairs such as "amin" and "amir" are included to prevent false positives.

  • Extend the organization stop-word list. Terms such as "groups" are added to improve organization name matching.

  • Expanded candidates for the MD prefix. Additional variants are generated by replacing "MD" with "Mohammed" and both options are evaluated.

  • Applied token coverage penalties. Scores are reduced when a watchlist name is embedded within a much longer screened name.

  • Adjusted scoring for name-in-company patterns. A mild penalty is applied when a watchlist name matches only part of a longer organization name.

  • Improved stop-word handling in preprocessing. Word boundaries are enforced to avoid matching stop words inside larger terms.

  • Restricted country code removal. Removal is applied only when the code appears at the end of the name.

  • Refined number-word removal. Removal is skipped when it would leave only a very short token (≤ 2 characters).

  • Corrected handling of legal-form abbreviations. Trailing name parts are preserved when company identifiers like "Corp" or "Ltd" are present.

  • Updated Python package import paths by moving modules from wf_name_matcher.matching.* to wf_name_matcher.domain.matching.*.

  • Added suffix matching. Matches are recognized where one token is a suffix of another, for example, "Gazprom" vs "Belgazprom".

  • Reduced acronym-based false positives. Acronym letters must match the order of the corresponding words.

  • Fixed scoring for near-identical organization tokens. Similarity rules are now used instead of treating such tokens as abbreviations.

  • Corrected single-word matching behavior. Over-penalizing matches are avoided against longer names.

  • Improved stop-word-only matching logic by reducing false positives when only common terms match with high confidence.

  • Resolved edge cases where no high-confidence token pairs are found during stop-word-only matching.

• The NER model is upgraded to v3.2.6 with the following changes:

  • Improved person name recognition. Titles and honorifics, for example, Dr, Prof, or MD, are now included as part of the name. For example, "Dr. Jane Smith" is extracted as a single entity.

  • Ensured accurate confidence scores for organizations. Organization entities now reflect the model's actual confidence, providing consistent scoring across entity types.

  • Updated regex code for ID_LEI to ensure both alphabetic and numeric characters are present.

  • Fixed the preprocessing regex issue that caused incorrect splits inside UUIDs.

  • Introduced Java 8 compatibility. The model is now available as a dedicated Java 8 bundle alongside Java 21.

  • Improved entity recognition in specific scenarios:

    • German e-commerce payment references to avoid duplicate organization detection

    • SWIFT indicator tag content to ensure classification as a named entity

    • Mixed entity blocks in SWIFT fields to separate person, organization, and address correctly

  • Corrected confidence scores for indicator-tagged entities. Actual model probabilities are now used instead of inflated values.

  • Improved vessel name detection. Structural prefixes, for example, MT or IMO, are no longer included as part of the vessel name.

  • Improved recognition of financial organizations in securities payment messages. Entities like BLACKROCK and GOLDMAN SACHS are now correctly extracted as ORGANIZATION when surrounded by financial identifiers.

  • Improved handling of "trading as" patterns. For example, "DANIEL SMITH TRADING AS CONSULTING MARKETING" correctly returns both PERSON and ORGANIZATION.

  • Improved recognition of romanized Chinese personal names and addresses.

  • Fixed the entity offset issue by correcting begin positions for salutation-prefixed names, for example, "MRS JANE SMITH".

  • Removed invalid entities. Punctuation-only outputs, such as a lone comma, are suppressed.