Skip to content

Release Notes

Version 3.0.0

DSAIL Platform 3.0.0 introduces DSAIL Language 1.3 with multi-instance rules, AMBIGUOUS verdicts for self-contradicting documents, universal rulesets built on a shared corpus of claims, and an Admin hub for monitoring and diagnostics. It also brings bulk delete across the workspace, unaffiliated datasets, a unified Activity feed, project version rollback, and a rebuilt unit system with counts, ratios and calendar durations.

New Features

  • DSAIL Language 1.3: DSAIL 1.3 is the language version of this release, and its headline addition is multi-instance rules. A rule can now be about the several things a document describes, such as each loan in a credit memo or each supplier in a vendor review, rather than one answer per document. Declare an entity, bind claims to it with of <Entity>, and quantify over its instances with ForAll, Exists, AtLeast, AtMost, ExactlyOne and CountWhere. The full language is available inside an entity scope (arithmetic on entity claims, claim-to-claim comparisons, counting, and mixing document-level claims with entity quantifiers), verified by the same Z3 engine as every other rule. Extraction asks its questions once per instance and preserves contradictions between instances instead of reconciling them away, and the ruleset wizard authors multi-instance rules across the basic, RAG and structured extraction methods. 1.3 also consolidates the changes made since 1.1: a single numeric type replaces integer and real; enum, declared ordered or unordered, replaces symbol; numeric values carry units; boolean subtypes govern how streamed measurements combine; and the language gains IF/CASE conditional expressions, counting quantifiers and set operations. Declare version 1.3; (the compiler also accepts version 1.2;); rulesets that still declare 1.0 or 1.1 no longer compile and must be updated. See the Rules About Many Things tutorial and the full DSAIL Language reference.
  • AMBIGUOUS verdicts: When a document both supports and refutes the same claim about the same entity, the assertion now reports AMBIGUOUS (with an amber badge on the Runs and Ruleset Studio surfaces) instead of collapsing to UNKNOWN. A rule's verdict is YES only when its resolved assertions all agree on TRUE, NO only when they all agree on FALSE, UNKNOWN when none resolved, and AMBIGUOUS otherwise. When an entity scope has no instances, an instance has no value, or a unit cannot be converted, the UNKNOWN carries an explicit reason.
  • Universal rulesets and the corpus: A governed vocabulary of shared claims, plus the axioms that connect them, lets the platform reason about rules written by different people from different sources together. A rule that names a registry claim (@namespace.name) shares one answer with every other rule that names it, so the question is asked once per run. Rulesets declare the corpus rulesets that govern them in the wizard's Governed by list, and portion marking against classification guides is supported as a first-class use case. Three new screens sit on the corpus: Corpus Health and the Vocabulary Workbench in the workspace sidebar, and Analysis inside a project. See The Corpus and the SCG Portion Marking tutorial.
  • Admin hub: Monitoring and Diagnostics now live together under an Admin entry in the workspace sidebar, shown only while the active group is support_admin. Old /monitoring/* and /diagnostics/* links redirect to their /admin/ equivalents.
    • Monitoring is streamlined to its API and Utilization sections alongside JVU, LLM and project usage, with tooltips on every chart; the System Health and "coming soon" sections are gone.
    • Diagnostics gains three tabs. Health shows per-component status cards with a link into that component's logs and a banner when telemetry is degraded. Logs filters by source, severity, component and free text, with a Live mode that pauses while the tab is hidden. Audit records who ran which health check or log query, and when; that trail is also available on the Activity page. Filters survive switching between tabs.

Improvements

  • Bulk delete for projects, rulesets and datasets: Every list page has per-row checkboxes, a select-all control and a Delete selected action. A shared confirmation modal requires typing delete, lists every item that will go and summarizes the cascade. Deletion is all-or-nothing: a pre-flight check refuses the whole batch and names each item blocked by an active or pending run or an in-progress dataset generation. Single-item delete uses the same modal and the same guards. Selection counts and deletes only the rows visible under the active search and sort, and the controls are hidden for read-only groups.
  • Unaffiliated datasets and the workspace Datasets page: Datasets join documents and rulesets in being able to exist outside a project and be added to one later. A new Datasets entry in the workspace sidebar lists every dataset with its project chip and count badges. A dataset already used in a run stays with its project; duplicate it instead.
  • Unified Activity feed: The workspace Activity page is a single newest-first timeline interleaving run summaries with changes to documents, rulesets and datasets, replacing the Feed and Logs subtabs. Filters for user, project, entity type, activity type, time window and free text narrow each other from left to right, and every card opens the page it describes. The project-level Activity page is removed in favor of the project filter.
  • Project version rollback: A project can be rolled back to a previous version, restoring its database state, with an export produced as part of the rollback.
  • Units: Numeric values carry a unit and conversion happens inside the rule at comparison time. The built-in standards are now independent libraries (Counts & Ratios, SI & metric, Calendar durations, Currencies, US customary & imperial, Digital information), each toggled per project from the Unit Library page. Counts (items, each, dozen, thousand and so on) and ratios (%, bp, ppm, per mille) are separate dimensions, so a bare number is never read as a percentage. The new Calendar durations standard treats a month as 28 to 31 days and a year as 365 to 366 days. A converter factor may be a range; a comparison across a range is decided only when the whole range agrees and otherwise resolves Unknown with a reason. A project converter overrides a standard's converter for the same unit pair.
  • DSAIL editor: Ctrl+Space offers your declared names first, then keywords and built-ins with call-shaped snippets; hovering a declared name shows how it was declared; generic word-based suggestions are off. Colours are regrouped by meaning and meet WCAG AA contrast, and the editor gains bracket-pair colouring, indentation guides, bracket matching and code folding.
  • Claim readings: A boolean claim can declare whether it is about the text (silence means False) or about the world (silence means Unknown), removing run-to-run variance for questions that read both ways. Undeclared claims behave as before.
  • Documents and project transfer: Uploads complete immediately, because OCR and parsing now run when a document is first used rather than at upload. Project import runs asynchronously with progress reporting, decompresses the archive once and batches its inserts, and the UI tracks several in-flight imports at once; export runs off the request path. Very large documents no longer fail on an internal message-size limit.
  • Interface: All fonts are bundled with the UI, with Inter as the primary typeface, so text renders correctly in air-gapped deployments. Delete operations report their outcome with toasts.

Bug Fixes

  • Deletion guards: Deleting a single project, ruleset or dataset referenced by an active or pending run is rejected instead of cascading the run away. A dataset still generating records cannot be deleted, with a staleness timeout so a crashed generation cannot lock it forever. Lists refresh consistently after create, generate and delete, and the per-rule trash icon in the Studio is always visible and keyboard accessible.
  • Multi-instance rules: Rules using arithmetic, claim-to-claim comparisons or CountWhere inside an entity scope can be saved and run. An entity that no assertion quantifies over is rejected at validation with an explanation instead of resolving UNKNOWN mid-run. Wizard output is compiled before it is accepted, a single-claim entity is rewritten into a quantified claim, questions are generated per instance rather than for the entity itself, and the Run Details "DSAIL Code" column is populated for quantified assertions.
  • Units: A bare 0.8 compared with an 80 "%" literal now resolves Unknown with a missing-converter warning instead of being read as 0.8%, previously wrong by a factor of 100.
  • DSAIL editor highlighting: local and the symbolic operators are coloured, scientific and leading-dot reals highlight as one number, Bool. subtypes and posture annotations are coloured only in their proper positions, unit literals are tokenized with their number, and a tokenizer failure that could strip highlighting from a whole rule is fixed.
  • Monitoring metrics: Expected 409 responses are excluded from API and LLM error counts, average JVU per project divides only by active projects, and the Resources section no longer shows a previously selected range.
  • Activity and logs: Quick time presets use local time, and ruleset run logs are attributed to their project.
  • Large projects and documents: Imports with thousands of documents no longer fail with 502, large exports no longer stall the API, OCR failures are surfaced and retried instead of being reported as success, replacing a document version's extracted text clears its derived state, and structure extraction no longer discards LLM results wrapped in markdown fences.
  • LLM providers: Rule extraction and scoring no longer force temperature 0 on Gemini-3 models, which degraded their output quality; other models are unchanged.

API Changes

  • Bulk delete: Bulk-delete requests accept up to 100 ids and return 409 with per-item blocking reasons when any item has an active or pending run or an in-progress dataset generation.
  • Admin endpoints: All /diagnostics/* and /monitoring/* endpoints require support_admin as the active group; the previous auditor-or-support-admin gate on monitoring is removed. GET /logs/system validates source and names the valid values in its 400, and the diagnostics log endpoint returns 422 for an unknown severity or component.
  • Unit library: GET /unit-library/standards lists counts_ratios, si, calendar, currency, us_customary and digital as separately activatable built-ins. GET /unit-library/si-standard returns SI units only; percent and basis points moved to counts_ratios. A Converter.factor may be a two-element [low, high] range, which may not carry an offset. A converter for an existing unit pair overrides it instead of being rejected. Saved unit libraries and compiled artifacts that reference the legacy merged standard key are expanded to its components on every read and stored normalized on save; no existing conversion changes its factor or offset, and point-valued converters behave exactly as before.
  • Datasets: project_id is optional on DatasetCreateRequest and nullable in dataset responses, GET /datasets includes unaffiliated datasets, POST /datasets/{dataset_id}/project/{project_id} and its disassociation path manage affiliation (rejecting double association, name clashes, and a dataset with runs), and WorkspaceStatsResponse gains total_datasets.
  • Project import and documents: POST /projects:import returns 202 Accepted with a job payload pollable at GET /projects:import/{job_id}, and POST /projects:import-conflicts returns an import_token so the archive is uploaded once. GET /documents/{id}/structure returns 404 until the document has been processed.
  • DSAIL grammar: Tightened so that every valid DSAIL program compiles; constructs that could not be compiled were removed.

Documentation

  • New pages: A top-level Security page covering the threat model, architecture, identity and access control, data protection and audit posture; concept pages for The Corpus, Corpus Health, the Vocabulary Workbench and Analysis; and two tutorials, Rules About Many Things and SCG Portion Marking.
  • Updated pages: The DSAIL Language page covers entity scopes, the unit standards (calendar durations, range factors, counts and ratios, and the spellings deliberately left out) and claim readings. The Runs page explains AMBIGUOUS and how a rule's verdict is derived. Getting Started covers the workspace Datasets page and the Activity feed. The Authentication reference is retitled and points to the Security page, System Requirements notes the customer's responsibility for the self-hosted environment, and the unit-library API descriptions describe the standards catalog.

Version 2.7.0

Version 2.7.0 introduces a major workspace reorganisation centred on unaffiliated documents and rulesets, a new support_admin role with diagnostics access, bulk-delete APIs for projects, datasets, and rulesets, and a redesigned Activity page alongside several navigation and usability improvements.

New Features

  • Unaffiliated Documents and Rulesets: Documents and rulesets can now exist outside any project and belong to at most one project at a time. New workspace-level Documents and Rulesets pages surface all assets across the workspace — including unaffiliated ones — each with a count badge in the sidebar.
  • Add to Project / Remove from Project: Every document and ruleset row includes an Add to project control to associate an unaffiliated asset with a project. Documents can be removed from a project via Remove from project in the project Documents page ⋮ menu (silver documents and documents backing a ruleset or dataset are permanent members and cannot be removed); rulesets can be disassociated from the workspace Rulesets page via the project chip's × control.
  • Duplicate Document / Ruleset: A new Duplicate action in the ⋮ row menu creates a copy of a document or ruleset (defaulting to [name]-copy) that can target any project or none, enabling easy reuse across projects.
  • New support_admin Role: Added a support_admin group that grants access to the Diagnostics module and (alongside the auditor role) the Monitoring dashboard, without requiring membership in a customer data-owning group.
  • Diagnostics API Module: Added a new /api/v1 diagnostics router exposing platform health and log endpoints, gated on the support_admin active group.

Improvements

  • Workspace Navigation Redesign: The left sidebar now includes dedicated Documents and Rulesets top-level entries (replacing the former Create/Import flyout and separate Insights entry), and the former Home page is renamed to Dashboard.
  • Activity Page Subtabs: The workspace-level Activity page is reorganised into two subtabs — Feed (run summaries, formerly the Insights page) and Logs (system activity log) — with a shared project filter. The same two subtabs appear inside a project's Activity view.
  • Read-only Enforcement for Affiliated Assets: A document or ruleset that belongs to a project is read-only when opened from the workspace-level page, with a link to open it in its owning project for editing.
  • Rebranded AI Suggestion Prompt in Ruleset Wizard: The prompt "Let us suggest title and context" has been updated to "Let Jaxon suggest title and context" to reflect the assistant's name.

Bug Fixes

  • Aligned Data Icons on Project Cards: Fixed icon and count alignment in project card chips so icons no longer shift and counts display with consistent width.

API Changes

  • Bulk Delete Endpoints: Added DELETE /projects/bulk, DELETE /datasets/bulk, and DELETE /rulesets/bulk endpoints, each accepting up to 100 IDs and returning 200 (all deleted), 207 (partial success), or 409 (blocked due to active runs).
  • Monitoring Endpoints Widened to support_admin: The /monitoring REST API endpoints previously restricted to the auditor group now also accept users acting as support_admin.
  • New Project Association and Duplicate DTOs: Added DocumentProjectAssociationResponse, DocumentDuplicateRequest, DocumentDuplicateResponse, RulesetProjectAssociationResponse, RulesetDuplicateRequest, and RulesetDuplicateResponse to support the new association and duplication workflows.
  • New ProjectRef DTO: Added a ProjectRef object (containing project_id and an optional name) returned on document and ruleset responses to indicate which project an asset belongs to.
  • API Version Bumped to 2.7.0: The REST API version string and OpenAPI spec description have been updated from 2.6.1 to 2.7.0.

Documentation

  • Getting Started Guide Updated: Revised to reflect the renamed Dashboard, new workspace Documents and Rulesets pages, unaffiliated asset concepts, the merged Activity/Feed layout, and removal of the separate Insights and Create/Import menu items.
  • Documents Concept Page Updated: Added sections covering project affiliation, the Add to Project / Remove from Project controls, silver-document restrictions, and the distinction between workspace-level and project-level Documents views.
  • Rulesets Concept Page Updated: Added sections covering project affiliation, the Add to Project control, workspace vs. project Rulesets views, read-only behaviour for affiliated rulesets, and the Duplicate workflow.
  • Runs Concept Page Updated: Clarified that runs against a dataset require a project-affiliated ruleset, and that unaffiliated rulesets can still be tested interactively in the Ruleset Studio; updated Activity page references to reflect the new Feed subtab naming.

Version 2.6.1

This release delivers targeted bug fixes that improve error clarity across dataset and attachment endpoints, along with a new API method for browsing dataset version history.

Bug Fixes

  • Fixed HTTP error masking on dataset endpoints: Validation errors (e.g., unsupported file type, encoding errors, or dataset not found) on the CSV import and dataset synthesis endpoints now return their proper HTTP status codes and messages instead of a generic 500 error.
  • Fixed attachment ownership lookup for documents: Document attachment verification now correctly resolves the document identifier to its internal primary key before comparing, preventing false "not found" errors when retrieving attachments by document ID.

API Changes

  • New endpoint – List dataset versions: A new list_dataset_versions API method returns all versions of a dataset in descending order (newest first), including version metadata but excluding full document payloads.

Version 2.6.0

Version 2.6.0 brings a new Admin Telemetry Dashboard, streamlined Ruleset Studio workflows, improved large-batch run stability, and a collection of usability fixes across datasets, projects, and document processing.

New Features

  • Admin Telemetry Dashboard: Added a new Monitoring dashboard for platform administrators with tabbed panels for JVU Usage, API Usage, LLM Usage, and Resource Usage — each featuring date-range filtering, project scoping, KPI summaries, time-bucketed charts, and top-N breakdowns.
  • Generate All / Regenerate All in Ruleset Studio: Added one-click "Generate All" (when no DSAIL exists) and "Regenerate All" (when DSAIL already exists) buttons to the rule editor, automatically chaining DSAIL generation and claim (question) generation in a single action.

Improvements

  • Streamlined Ruleset Wizard flow: The Ruleset Wizard now automatically opens the newly created ruleset in the studio as soon as rule and DSAIL generation completes, removing the previously required manual confirmation step. The wizard step formerly labelled "Configure Extraction" has been renamed to "Create Rules & DSAIL" to better reflect what occurs during that stage.
  • Simplified slot removal in datasets and rulesets: Slots can now be removed at any time without first manually re-binding or clearing dependent content. In datasets, documents bound to a removed slot are automatically dropped while other documents are preserved; in Ruleset Studio, claims bound to a removed slot are automatically reassigned to the first (default) slot. Removal is blocked only when it would eliminate the last remaining slot.
  • Large batch run stability: Batch runs processing 20 or more documents now store per-document results in object storage rather than inlining them in workflow messages, preventing failures caused by the 4 MB gRPC payload limit.
  • PDF conversion reliability: Document processing now enforces a 30-second timeout on the primary markdown converter and automatically falls back to pypdf (with OCR) when a timeout or empty output occurs, preventing hung or silently failed ingestion jobs.

Bug Fixes

  • Ruleset Studio claims loading fix: Fixed an issue where the claims panel was left empty when opening Ruleset Studio because claims were not loaded for the initially auto-selected rule until the user manually clicked another rule.
  • Fixed premature claim generation in Ruleset Studio: Resolved a bug where "Regenerate All" would skip DSAIL generation and jump straight to claim generation when a DSAIL already existed, due to a stale "completed" workflow status being picked up instead of the newly started run.
  • New Run modal scrolling fix: The New Run modal now correctly scrolls its form content on small screens, with the header and action buttons (Cancel / Start Run) remaining fixed in place.
  • Removed duplicate letter tile from Project list rows: Fixed a bug where a letter avatar tile was incorrectly displayed at the start of each row in the Projects list.

API Changes

  • New monitoring endpoints: Added a suite of endpoints under /api/v1/monitoring — including /monitoring/jvu-usage, /monitoring/api-usage, and related LLM and resource usage routes — returning structured, time-bucketed telemetry data with KPIs, breakdowns, and filters for date range, granularity, and project scope. The /monitoring/jvu-usage endpoint is auditor-only.
  • New internal endpoint – Ruleset config: Added GET /api/v1/internal/rulesets/{ruleset_id}/config to serve compiled ruleset configuration to Temporal workers, avoiding the need to inline large configs in workflow inputs.
  • New internal endpoint – Store document result: Added POST /api/v1/internal/runs/{run_id}/document-result to offload per-document batch run results to object storage, keeping Temporal workflow history payloads small for large batch runs.

Documentation

  • Updated ruleset slot documentation: The Rulesets concept page now documents that removing a slot auto-reassigns any bound claims to the default slot, and that the last remaining slot cannot be removed.

Version 2.5.0

Version 2.5.0 introduces a redesigned workspace navigation experience, a new Home dashboard and Insights feed, and streaming attribution that shows exactly which document chunks drove each verification result.

New Features

  • Two-level left navigation: A persistent workspace sidebar (Home, Projects, Create, Import, Insights, Activity, Documentation, API Reference) now swaps to a project-scoped menu (Documents, Rulesets, Datasets, Runs, Activity) when you enter a project, keeping context-relevant actions always at hand.
  • Home dashboard: A new workspace Home page surfaces headline stats (total projects, rulesets, and runs this week) alongside your favorite and recently updated projects.
  • Project favorites: Star any project to pin it to the Home dashboard; favorites are saved per user and can be managed via the new POST /projects/{project_id}/favorite and DELETE /projects/{project_id}/favorite endpoints.
  • Insights page: A workspace-level Insights feed aggregates run summaries across all projects, with pass/fail filter chips and direct links to individual run detail views.
  • Streaming attribution: A new streaming attribution pipeline processes documents in cumulative chunks and tracks which source text segments drove each extracted variable's value, enabling fine-grained provenance reporting surfaced in a new Attribution Viewer on the Runs page.

Improvements

  • Renamed extraction strategy — "Section Typed": The "Section Classified" strategy in the Ruleset Wizard is now labeled "Section Typed," with updated description text clarifying that it identifies the purpose of each section before extracting rules using purpose-specific prompts.
  • Ruleset Wizard — combined "Create Rules & DSAIL" step: Steps 6 (Extract Rules) and 7 (Generate DSAIL) have been merged into a single step; DSAIL generation now starts automatically as soon as rule extraction completes.
  • Project navigation: Entering a project swaps the sidebar menu, displays the project name at the top, and provides an ← All Projects link and breadcrumb for quick switching between projects.
  • Cross-project Activity page: The workspace-level Activity page now shows activity across all projects with a project filter; the in-project Activity view shows the same log pre-scoped to the current project.
  • Cumulative chunk processing: Document chunking has changed from a sliding-window approach to cumulative (additive) processing, improving attribution accuracy by evaluating progressively larger portions of the document.

Bug Fixes

  • Ruleset Wizard — DSAIL policy defaults to neutral: The default DSAIL policy in the Ruleset Wizard is now neutral, and a prompt bias that previously skewed results toward pessimistic outcomes has been corrected.

API Changes

  • Workspace statistics endpoint: New GET /projects/stats endpoint returns aggregate workspace statistics including total projects, rulesets, runs this week, and average pass rate over the last 30 days.
  • Enriched run summary fields: RunSummaryResponse now includes project_id, project_name, ruleset_name, dataset_name, started_at, completed_at, and duration_ms.
  • Provenance data on Runs API: The /runs endpoint now exposes attribution and provenance graph data alongside standard run results, including flip events and rule resolution provenance.

Documentation

  • Getting Started guide updated: Documents the new workspace menu, Home dashboard, project-scoped menu, and Insights page, and updates the Projects and Activity sections to reflect the new navigation.
  • Runs concept page updated: Notes that run results surface on the workspace-level Insights page with links back to the run detail view.

Version 2.4.0

DSAIL Platform 2.4.0 introduces a comprehensive document processing overhaul with intelligent structure detection, section classification, and coverage analysis, alongside a redesigned Projects landing page that serves as the new entry point with enhanced navigation and project management capabilities.

New Features

  • Document Coverage Analysis: Added coverage reporting that shows which sections of policy documents have been processed by rule extraction, with coverage status indicators (covered/thin/skipped).
  • Enhanced Document Structure Parsing: Added comprehensive document structure detection for PDF, DOCX, and plain-text documents with deterministic extractors for PDF bookmarks and DOCX heading styles, plus OCR support for scanned PDFs.
  • Section Type Classification: Added automatic classification of document sections into types (REQUIREMENT, EXCEPTION, DEFINITIONAL, PROCEDURAL) using LLM analysis.
  • Extraction Logging: Added detailed logging capabilities for rule extraction processes to track progress and diagnose issues.
  • Flat Document Chunking: Added intelligent document chunking that preserves section boundaries and maintains context for better rule extraction.
  • Projects Landing Page: Added a dedicated Projects page as the new entry point, showing project cards in a grid layout with search and sort capabilities for returning users, and a welcome screen for first-time users.
  • Reusable Toast Notifications: Added a new toast notification system for better user feedback across the platform.

Improvements

  • Enhanced LLM Integration: Improved LLM utility functions and model management for more reliable AI-powered document processing.
  • Structured Pipeline Processing: Enhanced the structured extraction pipeline with better signal handling and processing flow.
  • Project Cards Enhancement: Project cards now display entity counts (documents, rulesets, datasets, runs) and last updated timestamps to help users quickly assess project activity.
  • Navigation Improvements: The navbar logo now links back to the Projects page, and the app automatically redirects users to select a project when none is active.
  • Enhanced entity browsing: Added search, sort, and card/list view toggle functionality to Documents, Rulesets, and Datasets pages for improved navigation and organization.
  • Project search: Added search functionality to the project dropdown to filter and find projects by name.
  • Rule name display in test panel: Added a read-only field showing the selected rule name in the Ruleset Studio test panel for better context while testing rules.
  • Ruleset Studio default posture: Changed the default DSAIL code posture from pessimistic to neutral in the Ruleset Studio page.

API Changes

  • Extended REST API: Added new endpoints and data transfer objects for documents, rules, and extraction runs to support the enhanced functionality.
  • Projects API Enhancement: Added updated_at, doc_count, ruleset_count, dataset_count, and run_count fields to the list projects endpoint.

Documentation

  • Getting Started Guide: Updated project workflow documentation to reflect the new Projects page and improved onboarding experience.

Version 2.3.0

This release introduces powerful new document analysis capabilities including structure parsing, rule provenance tracking, and coverage reports, along with enhanced evidence support in verification results.

New Features

  • Document Structure Analysis: Added comprehensive document structure parsing with support for PDF bookmarks, DOCX headings, and OCR-based structure detection for scanned documents.
  • Rule Coverage Reports: Added coverage analysis that maps extracted rules back to document sections, showing which sections are well-covered, lightly covered, or skipped during rule extraction.
  • Rule Extraction Provenance: Added detailed provenance tracking that records which document sections each extracted rule originated from, including extraction logs and metadata.
  • Section Classification: Added automatic classification of document sections into types (REQUIREMENT, EXCEPTION, DEFINITIONAL, PROCEDURAL) using LLM analysis.
  • Evidence in Run Results: Added evidence field to run result claims that provides verbatim excerpts from source documents supporting each claim's evaluation.

Improvements

  • Ruleset Studio Page: Enhanced the ruleset management interface with new document structure visualization and provenance display capabilities.
  • Large File Upload Support: Added support for uploading larger documents with extended timeout handling and automatic structure parsing during upload for improved performance.
  • LLM Error Handling: Improved error handling for LLM connections to gracefully handle network failures and provide better user feedback when language models are unavailable.
  • Multi-slot Dataset Document Selection: Documents are now sorted alphabetically and already-assigned documents are filtered out from the selection dropdown, with a clear message when all documents are assigned.
  • Extraction Logging: Enhanced extraction process logging with detailed tracking of rule extraction steps and metadata.
  • Intelligent Text Chunking: Added intelligent text chunking capabilities for processing large documents in manageable segments.

Bug Fixes

  • LLM Model Discovery: Fixed issue where LLM proxy connection failures would crash the application instead of gracefully falling back when model lists are unavailable.

API Changes

  • Documents API: Added new endpoints and data structures for accessing document structure, rule provenance, and coverage reports.
  • Rules API: Extended rule objects to include provenance metadata and extraction context information.
  • API Field Rename: Renamed evidence field to llm_explanation in run results API responses for better clarity.

Documentation

  • SEAD4 Tutorial Update: Updated the Models step to reorder and rename fields - 'Configuration Model' is now listed second and 'Run Model' is renamed to 'Inference Model' and listed first.
  • Tutorial Button Labels: Updated button labels in getting started guide and tutorials to match current UI ("Start Rule Extraction" → "Extract Rules", "Start DSAIL Generation" → "Generate DSAIL").

Version 2.2.0

This release introduces a comprehensive claim quality scoring system that evaluates rule reliability and provides automated remediation suggestions, along with enhanced user interfaces and expanded API capabilities.

New Features

  • Claim Quality Scoring System: Added comprehensive claim quality assessment with variance, ablation, absence alignment, and counterfactual testing to evaluate rule reliability and suggest improvements.
  • Automated Remediation Suggestions: Added intelligent diagnosis and remediation recommendations for problematic claims, displayed in read-only DSAILEditor format.

Improvements

  • Claim Scoring Evidence Visualization: Added UI display of claim scoring evidence and test results throughout the platform interface.
  • Enhanced Ruleset Studio: Improved functionality and user interface in the Ruleset Studio page.
  • Updated Runs Page: Enhanced features and better user experience on the Runs page.

API Changes

  • Claim Scoring API Endpoints: Added new REST API endpoints for claim scoring functionality across LLMs, rulesets, and runs services.
  • LLM Management APIs: Added new REST API endpoints for managing and interacting with LLM configurations.

Version 2.1.0

This release focuses on improved user experience across the platform with new search, filtering, and information features, along with enhanced multi-document support for rulesets.

New Features

  • Document search: Added a search function to the Documents page that filters documents by name in real-time, with an expandable search input and clear visual feedback when no results match.
  • Document view modes: Added card/list view toggle controls to the Documents page, allowing users to switch between compact list view and detailed card view layouts.
  • Interactive concept information: Replaced static documentation links with modal dialogs on the Datasets, Documents, Rulesets, and Runs pages that explain concepts inline without leaving the platform.
  • Multi-document ruleset support: Rulesets now support multiple named input slots for claims that reference different documents (e.g., comparing a contract against an amendment), with slot management in the ruleset editor and slot-specific claim binding.
  • Enhanced dataset structure: Datasets now organize documents into logical records with sequence numbers and input slot assignments, supporting multi-slot rulesets with proper document-to-slot mapping.

Improvements

  • Modal keyboard navigation: Added Escape key handling to close modals throughout the platform, including the ruleset wizard, dataset modals, document editor, and run detail views.
  • Rule organization: Rules in the Ruleset Studio are now sorted alphabetically by name for consistent ordering and easier navigation.
  • Document name display: Improved document name presentation in variance test results and evaluation tabs by removing file extensions for cleaner display.
  • Navigation guard enhancements: Converted browser confirmation dialogs to styled modal dialogs for unsaved changes warnings, providing better visual integration with the platform design.
  • API documentation improvements: Updated OpenAPI specification version to match platform release and enhanced API reference page with multiple viewing options.

Bug Fixes

  • Timezone handling: Fixed UTC timestamp parsing to properly respect browser timezone settings when displaying dates and times throughout the platform.
  • Document placeholder text: Updated the create document modal placeholder to remove file extension guidance, reflecting that extensions are stripped from display names.
  • DSAIL validation improvements: Enhanced error handling for LLM rate limiting with specific error types and user-friendly messaging when providers return persistent 429 responses.
  • Search functionality: Implemented proper search input focus management and keyboard shortcuts for the new document search feature.

API Changes

  • Dataset structure updates: Added input_slots, target_ruleset_id, sequence, and input_slot fields to dataset DTOs to support multi-document workflows and improved dataset organization.
  • Claim binding: Added input_slot field to claim creation and update endpoints, allowing claims to be bound to specific input slots within multi-slot rulesets.
  • Run request format: Enhanced run endpoints to accept named document slots in the input_documents object, supporting both single-slot ({"input": {...}}) and multi-slot ({"contract": {...}, "amendment": {...}}) request formats.

DSAIL Language v1.1

The DSAIL grammar was narrowed to constructs that compile without probabilistic evaluation. New operators include Xor, Implies, IsMember, IsSubset, length, and the % (modulo) arithmetic operator. Distribution-based constructs have been removed in favor of deterministic logic.

Version 2.0.2

This release introduces comprehensive audit logging, a redesigned ruleset creation wizard, document editing capabilities, and enhanced data management features along with numerous bug fixes and UX improvements.

New Features

  • Activity logging system: Added a unified activity tracking system with audit logs for all platform operations (document uploads, ruleset changes, run execution) and system logs for technical diagnostics. The new Activity page provides filterable, searchable logs with project-based access control and real-time updates.
  • Document markdown editing: Documents can now be edited directly in the platform using a full-screen markdown editor with search functionality. The editor supports both read-only viewing and editing modes, with version control automatically creating new document versions when changes are saved.
  • Document import enhancements: The document import interface now supports CSV files (each row becomes a document), has a unified import modal for all file types, and includes better validation with support for files up to 100MB and PDFs up to 5 pages with automatic OCR.
  • Redesigned ruleset creation wizard: The ruleset creation process now uses a unified 8-step wizard supporting both policy-driven extraction and manual creation workflows. The wizard includes theme suggestion, background processing for long-running tasks, and improved progress tracking.
  • Enhanced dataset management: Added CSV import for bulk document creation, automatic generation progress tracking, and improved organization of gold vs silver documents with better sorting and filtering.

Improvements

  • Better navigation and project management: Projects are now sorted alphabetically, duplicate names are prevented, navigation guards warn before discarding unsaved work, and switching projects properly redirects from context-specific pages.
  • Improved ruleset studio experience: Added resizable panels, DSAIL posture controls (optimistic/neutral/pessimistic), enhanced rule creation workflow, better error handling for empty rulesets, and improved validation feedback.
  • Enhanced run management: Added click-to-copy for run IDs, better error messaging for timeouts and failures, improved result organization, and run type-aware tab switching after creation.
  • UI consistency improvements: Updated all instances of "DSL" to "DSAIL" throughout the platform, standardized button terminology ("Create" instead of "New"), improved hover states and tooltips, and enhanced visual feedback with darker text colors for better readability.
  • Performance and reliability: Implemented batch workflow throttling to prevent interactive runs from being blocked, improved certificate renewal robustness, better handling of concurrent operations, and enhanced error recovery mechanisms.

Bug Fixes

  • Authentication and access control: Fixed OAuth callback handling that was causing state parameter issues, resolved login page avatar display problems, corrected group membership handling for new users, and improved session management.
  • Document and dataset operations: Fixed file upload error handling for rejected files, corrected document name display by removing file extensions in most contexts, resolved issues with large document processing, and improved CSV import validation.
  • Ruleset and run stability: Fixed rule creation workflow to immediately show new rules without requiring page refresh, resolved DSAIL validation for empty content, corrected variance calculation for multi-mode results, and improved workflow status tracking.
  • Platform reliability: Enhanced mTLS certificate renewal to handle authentication errors gracefully, fixed database connection issues in logs when services are unavailable, resolved temporal workflow cleanup, and improved error propagation from background tasks.

API Changes

  • Enhanced audit logging: Added project-based filtering to audit log endpoints and improved log entry metadata including project associations and operation context.
  • Document management: Added new endpoints for markdown content retrieval and editing (GET/PUT /{id}/versions/{ver}/markdown) and document renaming capabilities.
  • Run filtering: Added project-based filtering to runs API to ensure users only see runs for their selected project context.
  • Authentication improvements: API now accepts Authentik API tokens directly as Bearer tokens, bypassing the complex OAuth2 client credentials flow and simplifying programmatic access.

Documentation

  • Updated API documentation: Embedded interactive Swagger UI and ReDoc directly in platform documentation with version-aligned OpenAPI specifications and improved authentication examples.
  • Enhanced concept guides: Updated all documentation to reflect the new wizard workflow, corrected terminology throughout (guardrails → rulesets, DSL → DSAIL), and added comprehensive tutorials including a new SEAD-4 security clearance tutorial.
  • Improved getting started: Revised the getting started guide with current screenshots, updated workflow descriptions, and clearer step-by-step instructions for all platform features.

Version 2.0.1

Version 2.0.1 is a patch release focused on deployment flexibility, authorization hardening, UI reliability improvements, and new documentation.

Deployment & Infrastructure

  • AWS Marketplace AMI: Production-ready AMI for one-click deployment via the AWS Marketplace.
  • Automated Docs Publishing: Platform documentation at docs.jaxon.ai is now published automatically via pipeline or locally with deployment/docs/publish_docs.sh.

Authorization & Security

  • System Admin Group Protection: The authentik Admins group is now read-only in the platform — members can view but not create or modify assets, preventing accidental asset ownership by infrastructure administrators.
  • Dynamic Group Hierarchy: Replaced hardcoded group allowlist with a parent-group hierarchy for application access, allowing flexible team structures without code changes.
  • Authentik Blueprint Fix: Added automatic Authentik worker restart after deployment health check to resolve user-group binding 403 errors on fresh deploys.
  • Enhanced Network Security: Implemented mTLS (mutual TLS) authentication across all internal service communications for improved security isolation.
  • Session Management: Improved logout handling and added idle timeout warnings with clean session termination.

Ruleset Improvements

  • Publish Validation: Publishing a ruleset now validates that it contains at least one rule and all rules have DSL code, preventing incomplete rulesets from being frozen as published versions.
  • Wizard Polling Reliability: Rewrote the rule extraction and DSL generation polling in the Ruleset Wizard to use a centralized usePoller hook, eliminating race conditions and stale-state issues during long-running extraction workflows.
  • Activity Monitoring: Added real-time notifications and progress indicators for rule extraction and DSL generation workflows.
  • Interactive Runs: Added support for interactive runs with parallelized processing and downloadable result files for batch operations.
  • Variance Testing: Implemented variance test capabilities for comprehensive ruleset validation and analysis.
  • Claims Management: Enhanced claims system with author tracking and duplicate prevention for audit consistency.

User Interface

  • System Logs: New unified logging page with filterable audit trail showing platform activity, guardrail runs, and system events with expandable details and auto-refresh polling.
  • Collapsible Sidebar: Made the navigation sidebar collapsible to maximize workspace area.
  • Document Management: Added ability to upload and manage documents directly through the UI with improved text extraction for PDF and DOCX files.
  • Datasets UI: New interface for managing datasets with full CRUD operations.
  • Project Import/Export: Added functionality to import and export projects with proper handling of naming conflicts.

API Enhancements

  • REST API Examples: Added code examples for creating runs via curl, Python, and JavaScript with proper authentication handling.
  • DSAIL Validation: Added endpoint for validating DSAIL code syntax and structure.
  • LLM Configuration: Expanded default model support including Google Gemini and additional Claude models.
  • Text Extraction: New endpoint for extracting text from PDF and DOCX files.

Bug Fixes

  • Project-Scoped Runs: Runs are now filtered by the active project, fixing an issue where the Runs page displayed data from other projects.
  • Documents Page Deduplication: Resolved a duplication issue on the Documents page caused by a merge conflict.
  • UI Build Dependencies: Fixed build failures by regenerating package-lock.json to include missing html-to-image dependency.
  • Authentication Flow: Fixed routing issues when users are logged out and improved handling of authentication state transitions.
  • Rule Edit Modal: Resolved issues preventing changes from being persisted in the edit guardrail window.

Documentation

  • SEAD-4 Tutorial: New end-to-end tutorial walking through SEAD-4 compliance verification — from document upload through ruleset creation, testing, and variance analysis.
  • Updated Screenshots: Refreshed all platform screenshots to reflect the current UI.
  • Authentication Reference: Expanded documentation on the OAuth2/PKCE authentication flow and Authentik integration.
  • API Reference: Updated API documentation with current endpoints and authentication examples.
  • Best Practices: Updated guidance on variance testing and removed second-person phrasing for better clarity.

Version 2.0

Version 2.0 is a major platform update that simplifies the ruleset model, introduces AI-powered rule creation, and redesigns the user interface around a focused document verification workflow.

Simplified Ruleset Model

The platform now focuses exclusively on rulesets powered by DSAIL formal logic. The previous guardrail types (Consensus, Consistency Checking, Critique & Revise, Entailment, LLM-as-a-Judge, and Human Review as standalone) have been replaced by a unified Ruleset/Rule model. Every ruleset contains one or more rules, each encoding a single policy requirement as DSAIL code evaluated by an SMT solver for mathematically provable results.

Ruleset Creation Wizard

A new wizard extracts rules from policy documents automatically. Paste a compliance document or policy summary, and the wizard analyzes the text, identifies individual requirements, and generates both natural language rule descriptions and DSAIL formal logic code. This replaces the manual rule-by-rule configuration workflow.

Ruleset Studio

The integrated Ruleset Studio provides a unified environment for viewing and editing rules, DSAIL code, and extraction questions. Each rule's components are visible in context — natural language text, formal logic, and the questions the platform uses to extract relevant information from documents.

REST API

A new REST API provides full programmatic access to all platform operations — managing projects, documents, rulesets, datasets, and runs. Interactive API documentation is available at /api/docs (Swagger UI) and /api/redoc (ReDoc), generated automatically from the OpenAPI specification.

Redesigned Navigation

The interface is now sidebar-driven with four primary sections:

  • Documents — First-class document management with versioning and text content support
  • Rulesets — Create, view, and manage rulesets with the creation wizard and Ruleset Studio
  • Datasets — Organize test data and link documents for accuracy measurement
  • Runs — Unified testing interface with accuracy and variance testing capabilities

Documents

Documents are now a first-class entity with full versioning support. Create documents by pasting text directly or uploading files. Every edit creates a new immutable version, providing a complete audit trail.

Runs

The Runs interface replaces the previous Playground as the unified testing environment. Run rulesets against individual documents or entire datasets. Test types include:

  • Single document tests — Quick compliance verification against one document
  • Accuracy tests — Measure ruleset performance against gold-standard datasets with F1, Precision, and Recall metrics
  • Variance tests — Statistical analysis across multiple iterations to assess ruleset consistency

Previous Releases

Version 1.3.0

Dataset Management & Testing

  • Dataset Creation & Management: Comprehensive dataset system for organizing test data. Create datasets from scratch, import from CSV, or export to CSV for external analysis. Manual record editing with full CRUD operations.

  • Synthetic Data Generation: Parallel bulk generation of synthetic test records (1/10/100 at a time) using configurable LLMs. Advanced candidate scoring with D-score diversity metrics ensures high-quality, diverse training data. Background job processing with real-time progress tracking.

  • Test Execution Framework: Background test job system for running comprehensive accuracy and variance tests against datasets. Real-time progress monitoring with cancellation support. Jobs persist across navigation for long-running test suites.

Quality Assurance & Validation

  • Accuracy Testing: Evaluate ruleset performance against gold-standard datasets with detailed per-rule metrics including F1 score, Precision, and Recall. Overall ruleset scoring with confusion matrix analysis to identify systematic errors.

  • Variance Testing: Statistical variance analysis across multiple test iterations to measure ruleset consistency. Identify non-deterministic behavior patterns and evaluate reliability across different inputs and conditions.

  • Test Results Visualization: Interactive test result displays with sortable tables, per-rule performance breakdowns, and detailed accuracy metrics. Export test results for reporting and quality assurance workflows.

Enhanced User Experience

  • Record Management UI: Intuitive interface for viewing, adding, editing, and deleting dataset records with appropriate field validation.

  • Background Job Monitoring: Real-time status cards for generation and test jobs with progress bars, completion metrics, and cancel/done actions. Jobs continue running during navigation with persistent state.

  • Data Quality Tracking: Automatic labeling of records as Gold (uploaded/manual) or Silver (synthetic) quality for downstream filtering and weighted evaluation strategies.

Version 1.2.0

Authentication & Authorization

  • Authentik Integration: Enterprise-grade authentication system with SSO support and comprehensive user management.
  • Group-Based Access Control: Security groups function as teams that own assets at a project level, with flexible assignment based on organizational structure.

Configuration Management

  • Version Control for Rulesets: Track changes to ruleset configurations with complete version history and ability to revert to previous versions.

Policy Rules Enhancements

  • Verification Rulesets for Questions: Add secondary verification layer to Policy Rules questions, enabling configurable rulesets to validate LLM-generated answers for improved accuracy and reliability.

Version 1.1.2

Enhanced User Experience

  • Sorting for Rulesets lists: Quickly organize items in the rulesets view.

Platform & Infrastructure

  • Kafka Topic Creation Flow: Adjusted creation process to avoid topic creation bootstrapping issues.
  • Kafka Image Update: Switched to the Apache-maintained image due to Bitnami deprecation.

Version 1.1.1

Enhanced User Experience

  • Speech-to-Text Inputs for Textareas: Microphone-powered STT for textareas across the UI for faster, hands-free input.

AI-Powered Automation

  • RAG-Based Pipeline for Policy to Rule Extraction: Retrieval-augmented option for converting policy documents into structured rules.

Expanded LLM Support

  • Support for GPT-5: Configure and run rulesets with GPT-5.

Reliability & Quality

  • Bug Fixes & UI/Usability Improvements: General stability enhancements and minor UI polish.

Version 1.1.0

Core Features

  • Batch Processing: Enhanced Python client with support for batch request processing, enabling efficient bulk verification operations.

Client Library Improvements

  • Pre-Flight Validation: Jaxon client validates target rulesets before execution, preventing wasted resources on invalid configurations.
  • Enhanced Batch Processing: Improved batch operations that continue processing valid items while clearly identifying and logging validation failures.

Platform & Infrastructure

  • Kubernetes Support: Full Kubernetes deployment support with Helm charts for production-scale orchestration and container management.
  • Enhanced Python Client: Improved client library with batch processing capabilities.

Version 1.0.0

Core Features

  • Rulesets: Verify LLM outputs with Policy Rules (Neurosymbolic) rulesets.
  • Integrated Logging & Dashboard: Real-time dashboards for monitoring inference.
  • LLM Generalization: LLM proxy for decoupled usage of major LLM providers.

APIs and Integrations

  • Python Client: Installable Python client for direct integration into customer applications.