Releases

This is a major release spanning administration, dispersion screening, and mapping. The biggest changes give administrators direct control over platform configuration and model choice, add a watchable multi-agent conversation mode, and bring animated 3D dispersion plumes to the map.

• Developer documentation. Several months of stranded developer documentation have been published and given an accuracy pass, along with an automated check to catch it going stale again.
• Admin settings. Administrators now have a dedicated settings page covering 151 runtime configuration options, including a collapsible advanced section, with every setting documented in place.
• Model selection and compliance. Tenants can now choose their own models and supply their own provider credentials, with residency and compliance rules enforced as hard filters per tenant. An automatic fallback chain redirects a request to another model rather than failing it if a preferred model is unavailable.
• Convened conversations. A conversation can now hand off work to a watchable sub-agent running in its own conversation thread, bounded by administrator-configurable spend and concurrency ceilings (up to 10 concurrent) instead of an open-ended run.
• Dispersion screening and 3D plumes. Dispersion results can now render as animated 3D volume plumes directly on the map, driven by vertical wind profiles, with clearer refusals when a scenario falls outside the valid range of the screening method. Exposure threshold tables are now generated directly from EPA reference data instead of being entered by hand.
• Spatial and mapping tools. Distance and area queries now push down to the spatial database for large datasets, with geographic transforms handled automatically. The map gained a right-click menu to ask about a point or copy its coordinates, a redesigned legend and feature popover, and basemaps that no longer require a separate API key.
• Verifiable answers. Answer checks now require a table cell to carry real data before it counts as a claim, catch status-only responses that never actually answer the question, and keep citation markers out of the visible text. Every export and chart now carries a lineage trail back to its data sources, and workbook exports include a sheet documenting how the file was made.
• Data sources. Eunice can now keep vintage-retaining local mirrors of external data sources, so a snapshot from a point in time stays available even after the source changes upstream. New sources include municipal geography (neighborhoods and council districts) and federal spending data, and outdated catalog entries are marked superseded and demoted below their replacement in search results.
• Charts and document exports. Chat and generated documents now render math notation correctly, and regenerating a document creates a new version rather than overwriting the original, preserving its history.
• Reliability and security. Session handling is now hardened against cancellation errors mid-turn, background jobs log their output instead of discarding it, and terminal sessions enforce stricter host and credential checks, including SSH host-key pinning. Email integration is now disabled by default, and sandboxed code execution blocks outbound network access unless explicitly allowed. Security hardening also extended across authentication and network egress.
• Compute and execution. Data analysis and spreadsheet tools now run in isolated worker processes with bounded reuse, reducing memory pressure, and large data transfers move through a lazy, byte-budgeted delivery path instead of loading everything upfront.
• Search and retrieval. Document search now combines multiple ranking signals, raising answer pass rates roughly sixfold in evaluation testing, and preserves section headings and document identity more reliably during indexing.
• Tool discovery and memory management. The platform now surfaces relevant tools based on live usage signals rather than static lists, and idle conversation memory is reclaimed automatically instead of held indefinitely, based on what is already durably stored.
• Interface and skills. Skills can now be imported directly from a zip file or a GitHub URL, and the chat interface received visual polish, including cleaner tables, a frosted active tab, and improved composer sizing.

This release focuses on environmental screening accuracy and speed, new data quality and entity tracking tools, and deeper integration with connected systems, alongside security and reliability improvements throughout the platform.

• Wildfire smoke and dispersion screening gets more accurate and far faster. Smoke plumes now follow the real wind field and flag when smoke will arrive at a location of interest, forecast raster processing dropped from 85.6 seconds to 7.6 seconds, and new NOAA smoke and fire detection feeds plus the national PurpleAir monitor network are now available on the same maps.
• Chemical release screening now accounts for changing wind conditions. Dispersion screening tracks wind that shifts over the course of a forecast instead of assuming one fixed direction, continuous releases stay anchored to their real source, and the actual transport speed is disclosed rather than estimated.
• Maps are more accurate and easier to read. Three dimensional terrain display, corrected street level basemaps, satellite imagery fixes, and clearer color legends make maps easier to interpret, and a failed map tile no longer renders a blank result silently.
• Data quality tools flag join risks before they cause bad results. Datasets now surface which columns are ready to join with each other, spatial joins can match on shared attribute values, and users can look up which table and column holds a specific value.
• A shared entity and fact system tracks records across data sources. Facts recorded from different sources are reconciled with conflicts surfaced rather than silently overwritten, and a bulk data ingestion path lets larger datasets be brought in and indexed automatically.
• Connected systems support two way work tracking. Users can create, comment on, and transition tickets in a connected work tracking system directly from a conversation, with per user credentials and identity resolution so “my tickets” reliably resolves to the signed in user.
• Sensitive actions now require a second approver. High impact operations route through a dual control approval flow, and a new safeguard prevents outside data pulled into a conversation from forcing or fabricating a write on a user’s behalf.
• Tenant and user data stays correctly scoped. Several fixes closed gaps where one tenant’s data, model context, or datasource resolution could leak into another conversation or user’s session.
• The assistant now discloses failures instead of hiding them. Truncated results, capped answers, and schema mismatches are surfaced to the user instead of returned as if complete, and error messages point to the actual problem, such as a missing column, instead of a generic failure.
• Reusable skills can be authored and managed from the interface. Users can create, edit, and run custom skills from a dedicated page, with import and export support and a “Run in chat” shortcut.
• Workspace and chat interface improvements. Multiple conversations can now run in separate tabs at once, the collapsed panel gained a reopen control, and general layout and button placement were cleaned up throughout.

This release focused on making Eunice’s answers verifiable: every answer now shows what was checked and how it was produced, and a new fact checking pass catches unverified data before it reaches the user. It also improved reliability in data handling and added a direct path from data grids to charts.

• Verifiable, inspectable answers. Every answer now carries a trust summary showing which claims were checked, what data sources were used, and what steps Eunice took to produce it, with a detailed view of the exact tool calls, code, and results behind each step.
• Automatic fact checking against real data. Before an answer is shown, Eunice checks that every data value in it (numbers, identifiers, table cells) actually appears in the retrieved evidence; if not, she retries the answer once and otherwise shows a visible warning rather than presenting unverified data silently.
• Data tables render from real data, not memory. When Eunice presents a table of retrieved data, it is now generated directly from the underlying dataset rather than typed out from memory, closing off a class of fabricated table values.
• Optional show your work and review modes. Users can turn on a setting to see Eunice narrate her own steps as she works, and administrators can enable a review mode where a second pass checks whether the original request was actually satisfied before a final answer is returned.
• More reliable data handling. Fixed an issue where binned or grouped data columns could fail to save or crash when viewed, and corrected a false positive in the fact checking system that flagged ordinary descriptive phrases like “a 10 row preview” as unverified.
• Chart access from data grids. Users can now open a data grid directly into the chart workspace, including editable spreadsheets and specific selected cell ranges, with the chart option reachable from full screen grid view as well.

This release rebuilt spreadsheet editing on a live formula engine, so uploaded workbooks keep their formulas and formatting when edited in Eunice. It also fixed several import issues that could corrupt or misdisplay spreadsheet data.

• Full formula fidelity spreadsheet editing. Uploaded Excel workbooks now open on a live formula engine: any cell is editable by the user or by Eunice, formulas recalculate automatically, and edits can be reverted.
• Cell and range formatting preserved and editable. Fills, bold, italic, underline, font color and size, alignment, borders, column widths, and row heights now display and can be edited directly in the grid, matching the original workbook’s appearance.
• Select a range in chat and edit it directly. Users can select a block of cells in a spreadsheet and ask Eunice to read or edit exactly that selection, values and formatting alike, without describing the range in words.
• More reliable spreadsheet imports. Fixed issues causing blank rows to display as zeros, valid dates to display as errors, and complex multi-table sheets to lose data outside the first detected block; every detected block on a sheet is now preserved.
• Faster edits and reopenable exports. Cell edits now complete in a single round trip so dependent formulas update immediately, and any exported Excel file can be reopened directly in Eunice’s grid editor instead of only downloaded.

This is a major feature release. It adds a full spreadsheet editing workspace and a living document workspace with built in provenance, alongside a set of data lineage, security, and reliability fixes across the compliance data catalog.

• Spreadsheets now recalculate and import faithfully. Uploaded Excel workbooks are automatically recalculated on upload, so formulas show real computed values instead of stale error placeholders. Every imported formula is now verified against the source data before it is accepted, cells that could not be read faithfully from certain file formats are clearly disclosed rather than silently dropped, and exporting a workbook re-emits live formulas correctly.

• Spreadsheets are now a full editing workspace. Users can open a dataset as a live grid, transform and filter columns, and sort, with accurate row counts in place of placeholder rows. A history panel tracks every change with the ability to revert the most recent edit, and workbooks are now fully visible and searchable in the data catalog alongside every other file type.

• Documents can be created, edited, and exported with full data provenance. A new document workspace lets users build living reports with embedded charts, maps, and tables that stay synced to their source data, track edit history, and import existing Word documents. Generated reports can include an automatic appendix documenting the exact data sources, methods, and models behind them, so a report stands on its own under outside review.

• More reliable source selection and data lineage. When more than one copy of a dataset exists, such as a local table and an external source, the platform now prefers the complete, most current copy instead of treating them as unrelated results, and discloses when a returned result is a partial sample rather than the full dataset. A catalog defect that could let a stale, orphaned copy of a dataset silently serve outdated rows has been fixed, and datasets now carry a genuine as of date wherever one can be determined.

• Expanded ENERGY STAR compliance data and automatic checks. The platform now ingests the full library of current and historical ENERGY STAR specification documents, extracts numeric eligibility thresholds from them, and links each spec version to the specific models and partners it affects. A new validation pass automatically flags implausible results, such as impossible coordinates or an unexpectedly empty join, before they reach the user.

• Administrative and review functions are now access controlled. Several admin and peer review functions, including the ones that trigger new tool creation, were previously reachable by any signed in user. They now require the appropriate role.

• Reliability fixes across search, documents, and maps. Document and knowledge search no longer truncates results before they reach the model, legacy Excel files that previously came back empty now read correctly, and updating a map layer now refreshes its data instead of leaving stale content on screen. Long conversations are now summarized without losing tool detail, and chart rendering, sorting, and legends are more stable while editing.

• Personalized responses and more capable tool loading. The assistant now remembers corrections and preferences a user has given her and applies them in later conversations. Related conversations are automatically grouped into projects with a one page summary of where things stand, and the assistant can load additional specialized tools mid conversation as needed instead of being limited to a fixed set.

This release expanded Eunice’s data tools across spreadsheets, environmental datasets, and mapping. Users gained more ways to edit, filter, and visualize data, alongside a broader library of ENERGY STAR and water data.

• Environmental data catalog expansion. Eunice now indexes all 52 ENERGY STAR certified product categories (up from 12) and adds a nationwide utility data service availability layer showing where building owners can access electric, gas, steam, and water usage data for benchmarking, plus a full ENERGY STAR Portfolio Manager benchmark dataset covering more than 1.8 million properties.
• Editable data grids with familiar spreadsheet tools. Users can filter and resize columns, promote a dataset or filtered subset to an editable working copy, and build charts directly from grid data, with charts and grids cross filtering each other.
• Water data queries by area. Streamflow and gage height queries now support searching by bounding box, county, or watershed, not just by state, making it easier to pull every monitoring station in a defined area such as a storm’s footprint.
• Saved maps with time based layers. Users can save a map from the workspace, reopen it later with a thumbnail preview, and layer multiple time based data sources, such as weather forecasts and dispersion plumes, on independent timelines instead of one shared control.
• Chart reliability fixes. Fixed charts hanging when switching chart types, defaulting to unusable identifier columns, and failing to load for certain uploaded or managed datasets.
• Interface polish. Date values now display correctly on chart axes instead of raw numbers, repeated chart updates no longer create duplicate cards, and small map previews show only view controls instead of the full editing toolset.