New capabilities and improvements to Elnora, published weekly.
Between September 30 and October 7 we gave the agent working access to file version history, made file saves and uploads appear live for everyone in your organisation, saved image regions beside Knowledge Base images, and had the agent build a workbook when you ask for a spreadsheet.
You can ask the agent to list the saved versions of a workspace file, open an earlier one and compare it with the current text, then restore it. Versions saved in a different text encoding compare cleanly line by line.
When a teammate saves, restores or uploads a file, the file tree and open viewer update in your workspace straight away. An unsaved plate map draft stays as you left it, and your own saves show their confirmation without reloading the editor.
Regions the agent draws on a Knowledge Base image, together with its scale calibration, save to a regions file in the image's own folder. Setting the scale and drawing regions in one request writes a single file, which the agent reopens next time you work on that image.
When you ask for a spreadsheet, the agent converts your table into a workbook under the file name you gave. Several tabs, formulas, charts or edits to an uploaded workbook go through the workbook editor, and a CSV is the fallback.
Between September 22 and September 30 we taught Elnora to find and number the layers in a cross-section, added particle size distributions in real units and per-channel image intensity, made agent edits merge with concurrent changes, and made file previews read instrument files in their original encoding.
Elnora finds the layers in a cross-section image by tracing boundaries that run across its full width, then draws each candidate layer as a numbered region over your image. When it is unclear which layer you mean, the agent asks before reporting thickness or porosity.
Particle and region measurements return a size distribution and a table with one row per object, in micrometres once the image has a scale. You can filter objects by area in square micrometres, and the result states which objects were left out of the outlines and why.
Image summaries give intensity statistics for each colour channel, and for each channel of a multichannel microscopy stack, with transparency kept out of the numbers. A greyscale image saved in colour is recognised as a single channel, and the agent measures it directly.
When a teammate or another conversation saves a file while the agent is editing it, the agent reads the newer version, checks what its own change would do to it, and merges both before saving. This covers protocols, plate maps, sequence annotations and other workspace files.
The file viewer detects the encoding of each text file, so instrument CSV exports show micro, degree and plus-minus signs as written. Sequence, alignment, structure and notebook previews read files the same way the agent does during analysis.
Between September 16 and September 22 we improved image analysis in Elnora. Area fraction and porosity measurements now state the method and threshold behind each value, scale calibrations are saved with the image along with their source, and segmentation separates touching objects into individual regions.
Area fraction and porosity measurements run through the built-in region tools, over the whole image, a selected region or a single layer. Each reported value names its method and threshold, and repeated measurements of the same image are reconciled into one consistent answer.
When the agent sets an image's scale, it tells you where the value came from: the file's metadata, a scale bar label, or a length you provided. The calibration is saved beside the image, and the image viewer and every later length measurement use it.
Between September 9 and September 16, organization deletion moved to two deliberate steps, with an approval page that states plainly when it cannot confirm a link's status. File version restore now preserves binary attachments exactly, and the agent tells you directly when it cannot read something you attached.
Terminating an organization workspace and permanently deleting it are now two separate actions, each with its own confirmation step. Removal always passes through the standard waiting period first, rather than skipping straight to permanent deletion in one action.
Restoring a previous file version or committing a working copy now copies the underlying bytes exactly, rather than passing them through a text decoder. Images, PDFs and other binary attachments keep their original content and exact size through a restore, so an uploaded scan or document reopens correctly afterward.
When a document attached to a conversation cannot be extracted, the agent names which attachment is unreadable and continues answering the rest of your question. You can then re-upload the file or try a different format, rather than a task ending with no answer at all.
Between September 2 and September 9, image analysis works across every supported imaging format, and file previews parse semicolon-delimited spreadsheet exports. Single-cell quality-control workflows now adapt to low-count samples, and organization deletion requests process more reliably, with confirmation pages that state exactly what was removed.
The image toolkit reads and analyzes the full range of formats in the imaging catalog, including whole-slide, volumetric and specialized microscopy files. Region selection and measurement in the viewer work the same way regardless of file format.
File previews and the statistics tool recognize semicolon-delimited files with comma decimals, a common format for spreadsheet exports from European locales, and split every value into its own column. Delimiter detection scores consistency across rows, and a spreadsheet's sep directive line is recognized and dropped automatically.
Doublet detection in single-cell RNA sequencing quality control scales its internal component count to the number of genes that survive filtering, completing on shallow or low-complexity samples. When an automatic threshold cannot be found, the workflow falls back to an expected rate and records which method set it.
Between August 19 and September 2, agent tasks retry automatically when a model call returns no output, and skills can open the reference documents they cite while working. Uploads now decode across more text encodings, and a file saved with no folder name lands in your personal space automatically.
When a model call returns no text and no tool calls, the fallback chain moves to the next model rather than ending the task. Work already gathered earlier in the turn is kept, so the conversation carries on rather than restarting from scratch.
Elnora's built-in skills can now open the reference documents they cite while a task is running, rather than relying on memory of their contents. This covers supporting material bundled with dozens of skills across lab automation, structural biology and materials science.
Extraction now detects a file's text encoding automatically and decodes older or less common formats alongside standard text, rather than treating them as unreadable. A file that is genuinely binary is reported as such rather than turned into garbled text.
Search across the knowledge base takes a full question and returns the passages that answer it, each tied to the file it came from. You can name a figure when you ask for it, and each panel of a multi-condition analysis is written under its own name.
Search across the knowledge base now takes a full question and returns the passages that answer it, each tied to the file it came from. Opening that file in full is one step further, so a method detail is confirmed against the original wording.
You can now name a figure when you ask for it, and an analysis run across several conditions writes each panel to your workspace under its own name. A figure needed for a report or a slide is then found by what it shows.
Plate readings go straight into a dose-response fit, with the curve, the potency value and the points behind it returned together. Elnora is connected to DepMap for gene dependency and to TCIA for public imaging collections, searchable by modality.
Plate readings now go straight into a dose-response fit, with the curve, the potency value and the points behind it returned together. The number that decides whether a condition is worth repeating comes back in the same conversation as the raw plate.
Elnora is now connected to DepMap for gene dependency, with symbols resolved through HGNC, so a lab nickname or an older alias reaches the same record as the approved symbol. Results narrow to a single cell line or tissue type.
Imaging collections in TCIA are now searchable by modality and by the metadata recorded alongside each study, and Elnora reads what it finds. Whether a public dataset exists for a planned comparison is settled in the conversation, before anyone starts a download.
Plate maps record what each well holds, and the instrument's readout file binds to that map so analysis runs on one table keyed by condition. Materials lists carry catalog numbers that link to live product pages, and a draft deck or workbook arrives as an editable file.
Plate maps now record what each well holds, and the instrument's readout file binds to that map. Analysis runs on a single table keyed by condition, which means a number in a result can be traced back to the well and the sample behind it.
Each condition is now compared against the control on the same plate, and a dose-response curve is fitted from the same table. Wells excluded from a fit are listed with the reason, which means a curve can be reviewed by a second scientist.
A materials list carries catalog numbers that link to the manufacturer's live product page, with distributor entries resolving to the product they resell. An order can be placed from the protocol itself, which keeps the list and what sits on the shelf in step.
A conversation that ends in a draft deck or workbook produces an editable file, built sheet by sheet with formulas that carry across sheets and charts drawn to match the data. The draft goes straight into a group meeting and is edited there.
Files written by the agent land in the shared knowledge base under the folder they belong to. Elnora writes the run file each connected liquid handler, plate reader, imager and cytometer accepts, then reads the output back, with every run rehearsed before it starts.
Elnora writes the run file each connected liquid handler, plate reader, imager and cytometer accepts, then reads the output back. Every run is rehearsed in simulation and approved by a scientist first, which means a plate is set up the way the protocol describes it.
Files written by the agent land in the shared Knowledge Base under the folder they belong to, which puts a result in front of the rest of the lab the day it is produced. A daily sweep archives stale documents and folds duplicates together when auto-tidy is on.
Elnora carries your saved details, your own skills and your named protocol templates into each new conversation. Three specialist agents share the work, cell line shortlists are assembled from measured evidence, and a saved style profile shapes the reports, decks and figures it produces.
Elnora now carries your saved details, your own skills and your named protocol templates into each new conversation. Name a template and the next experiment is drafted to that structure, which means a group's house format holds from one protocol to the next.
Cell line shortlists are now assembled from measured evidence, with each candidate carrying the record it was drawn from and its gene entry in Ensembl. The reasoning behind a choice is visible before a plate is seeded and reads well in a methods section.
A saved style profile now shapes the reports, slide decks and figures Elnora generates, setting colours, type, spacing and logos once for the group. A draft made for an internal review is already presentable when it reaches a collaborator or a funder.
You can now name a numbered figure in an uploaded paper and have it lifted into the write-up, and Elnora renders its own molecular views, secondary structures and plots beside it. Published data and new results end up compared on one page.
Three specialists share the work: a researcher handles literature synthesis, a data analyst takes results, and a critic reviews the design, ranking each point it raises and proposing a concrete fix. Weak controls surface while the design can still be changed.
Every answer closes with a source list naming the specialists Elnora consulted and the literature identifier behind each claim. A reader can follow a single statement back to the paper it rests on before that line goes into a manuscript.
Questions about what Elnora can run are answered from a live search of its own skill and tool catalog. Public accessions from GEO and SRA feed straight into a run, and single-cell quality control and variant-calling pipelines execute in the session.
Questions about what Elnora can run are answered from a live search of its skill and tool catalog, and responses drawing on general knowledge carry that label. A scientist reading the answer can tell which claims trace to a catalogued capability.
Public accessions from GEO and SRA now feed straight into a run, which means single-cell quality control and variant-calling pipelines execute in the conversation that drafted the protocol. Results sit next to the method they were produced under.
A skill loads together with the tools it calls, and both run inside a single session. Gene names in a regulation claim are matched to standard identifiers in Ensembl and UniProt before Elnora searches PubMed and Europe PMC for supporting evidence.
A skill now loads together with the tools it calls, and both run inside a single session. Each tool call and the result it returned is written into the session transcript, which makes the path from question to protocol step readable after the fact.
Gene names in a regulation claim are matched to standard identifiers in Ensembl and UniProt before Elnora searches PubMed and Europe PMC for supporting evidence. Every result carries the organism it was observed in, which a scientist needs when judging how far a finding transfers.
Elnora is connected to Open Targets for target and disease association scoring, and to Addgene for vector search with annotated sequences. A cloning request returns a plan that names its vector and the reason behind the choice, and pending invitations can be resent from the member list.
Elnora is now connected to Addgene, so a vector search draws on catalog entries and saves the annotated sequence of the one it picks. The search covers vectors a lab can order, so a cloning plan rests on material that ships.
A cloning request now returns a plan that names the vector to use and the reason behind the choice, with the annotated sequence attached. The reasoning is written out, which means a bench scientist can judge the pick before ordering anything.
The member list now carries a resend action for invitations that are pending or past their window. The fresh invitation keeps the role that was chosen and opens a new window to accept, which means a colleague stalled at onboarding rejoins in one step.
Elnora is now connected to Open Targets, so a target and a disease can be scored for the strength of their association, with the drugs known to act on that target and the trial stage each reached. A shortlist can be triaged before any bench work starts.
Buffer choices are checked against the salts they are paired with as a protocol is written, and research questions go to the source best suited to each one. Long runs stream protocol text into the run view, and an organization can be terminated, restored or purged on a schedule.
Buffer choices are now checked against the salts they are paired with, and a compatible form is selected as the protocol is written. A pairing that would precipitate is visible in the draft, before anything is weighed out.
Research questions go to the source best suited to each one, and answers name the papers or web pages behind them. A claim in a draft protocol can be traced back to its citation.
Long runs now stream protocol text into the run view as it is produced, alongside the reasoning and the code behind each step. Whether a run is heading the right way is clear before it finishes.
An organization can now be terminated, restored, or purged on a schedule. Termination removes members and opens an export window before the data is purged, which means a lab winding down has a defined period to retrieve its records.
Protocol edits save as separate versions, and undo and redo step a method back and forward through them. A single run moves from your data to a finished protocol, with analysis, literature search and drafting happening in the same pass.
A single run now moves from your data to a finished protocol: Elnora analyses the results, searches the literature and drafts the method, with Knowledge Base files in reach throughout. The draft cites your results and their sources, and each step can be traced back to its evidence.
Protocol edits now save as separate versions, and undo and redo step a method back and forward through them. Each version records its author and the version it came from, which means a method that changed over weeks can be read as an ordered trail of decisions.
A protocol carries several chat threads at once, each listed in the protocol view and reopening with its full history. Submissions sent through the API are extracted into a draft that opens in the editor and sits in the submitter's own protocol list.
A protocol now carries several chat threads at once, each listed in the protocol view and reopening with its full history. A troubleshooting question and a drafting conversation stay separate, and the reasoning behind each stays readable weeks later.
Submissions sent through the API are extracted into a draft that opens in the protocol editor and sits in the submitter's own protocol list. Each extracted step stays editable, and the draft becomes the working version a scientist corrects and builds the method on.
Any agent reply in the protocol chat takes a thumbs up or down, by mouse or by keyboard. Protocol views and their dialogs lay out for tablet screens and half-width desktop windows, keeping the reading order of the full screen.
Any agent reply in the protocol chat takes a thumbs up or down, by mouse or by keyboard. The rating stays attached to the reply it judges, so the next scientist reading the thread sees which answers held up at the bench.
Protocol views and their dialogs now lay out for tablet screens and half-width desktop windows, with the same reading order as the full screen. A protocol drafted at the desk can be read and followed at the bench.
Protocol edits now save as timestamped versions, and any two of them open side by side with each change on its own row. Imported documents and pasted text arrive already formatted in the editor's own markdown.
Protocol edits now save as timestamped versions, with the most recent marked current. The method run at the bench stays on record, which means a change can be traced to the day it was made and an earlier version restored for a repeat run.
You can now open two versions of a protocol side by side, with each change on its own row and both panes scrolling together. A changed volume or incubation time reads as one highlighted difference, and a revised method can be checked before it goes to the bench.
Imports of documents and pasted text arrive in the editor's own markdown, with lists, numbered steps and section breaks already formatted. The import shares one version history with every later edit, which keeps an imported method and its revisions together as a single record.
