Changelog.
Latest releases & updates.
Every feature, fix, and improvement shipped to LexVio — in reverse chronological order.
- New
The mobile app checks whether it is current (in testing)
On launch and on return to the foreground the app asks the server whether its build is still supported — separately for Android and iOS, so an App Store review that lags Play never locks iPhones out. A normal release is a dismissible nudge; a build the server can no longer serve shows an update-by date with a grace period; only past that date is it retired. Your session survives the update — no new sign-in code.
- Improved
A pause is part of the meeting record
Pause and Resume were already on every surface; each pause is now kept with its position. The transcript shows “— Paused 45 min —” where it fell, the minutes account for it, the exported minutes say “paused 12 min” beside the recording time, and a meeting stopped while paused ends where the pause began. A meeting left paused for four hours, or recording with no audio for six, is closed for you and finalised with what it has.
- Fixed
Blog photos and covers load
Uploaded images were handed out as private-storage links and failed to load. They are now served by the API with a year of caching; existing posts are corrected on read, so there is nothing to re-upload.
- New
Your byline, your call
Your author profile appears on your published posts by default, and a switch on your profile hides it; an editor can also withhold it, and the profile page says which of the two applies. My posts gained a status stepper, View live, Share, Withdraw on a submitted post and Edit-and-resubmit on a rejected one. The editor gained a live preview with word count and reading time, Ctrl/Cmd+S, an unsaved-changes guard, Insert image and a pre-submit checklist.
- Improved
Nexus knows the product it lives in
Ask Nexus about your meetings, today's tasks, a contract's lifecycle or your blog posts and it answers from your own records, with chips that open the meeting, task or post it cited. Its help guide was rewritten around Meetings, Todo, Contract lifecycle, Blog, Uploads and Account, every line checked against what the product actually does.
- Improved
A five-hour recording reaches its minutes
Uploads up to 300 MB — about five hours of audio. Long recordings are split and transcribed in order; speaker labelling runs in sequence so “Speaker 2” stays the same person across the whole meeting; minutes for anything over about an hour are drafted section by section and then reduced to one set. Nothing is truncated silently any more, and a long finalisation is no longer mistaken for a stalled one.
- Fixed
Retry runs, minutes generate on paid plans, no chunk is transcribed twice
Retry within an hour of a failed finalisation now actually runs, and a stuck meeting is swept every five minutes. Minutes generation no longer refused an organisation on a paid plan after its transcript had landed. The last chunk of a live recording is transcribed once.
- New
Share minutes as PDF or Word
On the phone app (in testing), Share minutes sends a PDF or a Word file to any app or saves it to the device: cover, participants, summary, numbered decisions, action items with owners and dates, next steps. Speaker labels are marked inferred, times render in your own zone, and the export is written to the matter timeline. The same export is on the desk API; the desk button is on the roadmap.
- New
Analysis progress, by step name
While a contract is being analysed the report page says which step it is on — reading the document, deciding the jurisdiction, indexing, analysing, preparing the report — and names the step if it fails. No percentage: the pipeline cannot measure one honestly, and a made-up number is worse than none.
- Fixed
Tester findings on the phone (in testing)
Creating a task tells you whose Today it lands on; a task can link a vault document or upload a new one on the spot; typed text no longer hides behind the keyboard in Vio, sign-in and the code screen; files the app cached — exports, drafts, recordings — are wiped on sign-out.
- Fixed
Global plans were being rate-limited like anonymous traffic
Every organisation on a global plan had been capped at ten requests a minute — the limit for a caller with no account — which a phone's start-up alone exceeds. Each plan now carries its own limit, matching its tier.
- Fixed
Account deletion has a real cooling-off
Requesting deletion no longer locks you out on the spot. Your account stays active for 30 days, the pending request shows on every device, and you can cancel it in that window — which the copy always promised.
- Improved
AI usage priced in USD, as you are charged
Billing → Usage shows AI cost in US dollars, priced at the moment of each call rather than converted later. A period with an unpriced request shows no figure rather than a partial sum, and the cycle-end date reads your commercial invoices — a monthly subscriber was being told the cycle ended with the plan.
- Fixed
The phone shows this product's privacy policy
The phone's Privacy screen had rendered a different edition's policy and legal basis, because the public route reads the site origin and a phone sends none. It now reads your organisation's edition, as the blog feed already did.
- Fixed
Health score without a category that does not apply
The organisation health score had averaged in a constant for a category this product does not include. It now scores contracts, compliance and litigation risk, and the desk renders the categories the server actually sends.
- Improved
A locked screen names what is missing
Every plan gate now answers with the feature it is gating, so the desk and the phone open the right locked screen and upgrade link instead of guessing from the menu. Two menu items whose surfaces are not part of this product no longer appear on either.
- Platform
Where you last signed in is recorded
Your account notes whether you last signed in from the desk or the phone, and which app build, so support can help with the right one.
- New
Todo — one task list for the whole practice
A task is what a person decided to do by a date: standalone or linked to a matter, a vault document or an obligation, with an assignee, a due date or time, priority, notes and a snooze that hides it without moving the date. My tasks, matter, team, follow-up and category views; a Today that follows your own time zone; a reminder digest at your 08:00. Tell Vio “Ask Anna to verify the indemnity clause by Friday” and it drafts the task, resolves the person and the matter, and files nothing until you confirm. On every plan, on the desk and the phone.
- New
Meeting Intelligence — record, transcript, minutes, tasks
Record a meeting from the phone or the desk, or upload a recording. The transcript builds while you talk, in ~45-second chunks; speakers are labelled Speaker 1, 2… and renameable — inferred, never identified, and every screen says so. Minutes come as summary, decisions, discussion, action items, unresolved questions, commitments and next steps; action items become tasks with one tap; everything lands in the matter timeline. Vio can read minutes and search meetings, and never writes one.
- New
Contract lifecycle — a contract record that can change
Statuses from draft through executed, expiring, expired and archived, each move audited. Edit the extracted metadata — a corrected expiry is marked as yours. The notice deadline is derived from expiry and notice period, flagged at 30, 7 and 1 days with a task filed at 30; auto-renewal is read from the contract's own terms — yes or unknown, never assumed. Record a renewal decision, upload a new version onto the chain, archive from the vault. Deal Desk approvals now notify the approver and the requester and move the contract; workflow steps become tasks.
- Improved
One accept list for every upload
Every door — Contracts, the phone's chunked upload, Vio attachments, bulk ZIP — accepts the same five types: PDF, DOCX, TXT, JPEG and PNG. Photos and scans are OCR'd. A legacy .doc is refused with a message rather than mis-read as text.
- New
Every desk surface has a phone screen (in testing)
The screens the phone used to hand off to the web now open on the phone: workflows, bulk jobs, vault views, redline sessions, negotiation, DD rooms, drafting, the knowledge base, M&A and procurement deals, team, billing and support. Where the phone cannot do a thing — payments, some exports — the screen says so in words, never as a dead tap.
- Fixed
Redline generation after the model upgrade
Generating a redline failed for part of 11 September after the move to Sonnet 5: the new model's reply can open with a reasoning block, which the pinned SDK read as an empty answer. Fixed the same day. The streaming copilot was unaffected.
- Security
Prompt-injection guard on all three model vendors
Instructions hidden inside an uploaded document were already defanged before reaching Anthropic; the OpenAI and Google legs of Deep Audit now get the same guard from one shared implementation. Every answer records which vendor and model produced it.
- Improved
The right model for each request
Each AI call declares what it needs — how fast, how much is at stake, whether the output is checked — and the router picks the tier from that: Haiku 4.5 for short structured calls, Sonnet 5 by default, Opus 5 for long reviews, Fable 5.1 for deep research or a retry after a low-confidence answer. Latency and tier are recorded on every call.
- Security
The sub-processor register says what reaches whom
The public register now states that OpenAI receives contract sections under Deep Audit (not only embeddings) and lists Google, whichever way the deployment default is set. Over-disclosure is the correct failure direction for a compliance representation.
- Platform
Current-generation models, one price table
Sonnet 5, Opus 5 and Fable 5.1 from Anthropic; GPT-6 Astra for the OpenAI leg of Deep Audit. Every call is now priced by the model that actually answered from one table for all three vendors — Haiku had been over-counted threefold — and a model the table does not know is priced at its vendor's ceiling, never at a cheaper default.
- New
Profile, devices and quiet hours on the desk
Settings → Profile edits your name, profession, job title, phone and preferred jurisdiction — the desk could not change your own name before. Settings → Devices lists signed-in phones and signs one out, session and push together. Quiet hours are editable; a window needs both ends.
- Improved
Upload past 50 MB from the browser
A file over 50 MB goes up in chunks with per-chunk retry, up to the API's 500 MB. A 120-page scanned bundle is no longer refused before it reaches the server.
- Fixed
A large upload decides its jurisdiction from the document
A chunked upload was analysed under your organisation's default jurisdiction whatever the governing-law clause said. It is now decided from the document like every other upload, and takes the same analysis options — party role, goal, industry, matter, Deep Audit.
- Improved
An evidence pack is frozen on every analysis
Every analysis now freezes a pack of what the product said, so a re-analysis cannot rewrite the earlier position. Download gives you the exact bytes the SHA-256 digest was taken over, and the digest is copyable — run shasum -a 256 and compare, with no access to LexVio needed.
- Platform
Retention and backup jobs exist — both off by default
A seven-year audit-log retention job now exists (retention had been indefinite); it refuses any window under the statutory floor. A nightly database backup job — pg_dump, SHA-256, encrypted S3, an integrity check, a retention prune — now exists too. Both ship off until an operator enables them, and our SOC 2 control record shows the backup control as a gap, not a pass, until that happens.
- Fixed
Currency is the contract's own
Every contract had been stamped with a default currency it never named, which then appeared on exposure figures, due-diligence roll-ups and exported PDFs. A contract now carries the currency it states or none; a report says “currency not stated” rather than inventing a symbol, and roll-ups keep an explicit unknown bucket.
- Fixed
Quiet hours follow your organisation's time zone
Notification quiet hours had been evaluated in one fixed zone for everyone, so a 22:00–07:00 window could be silencing you in the middle of your working day. The zone is now your organisation's.
- Fixed
The blog list shows who wrote a post, and how long it takes to read
The list said “LexVio Editorial” and “1 min” under every post. It now shows the author's card where the author is visible, and a reading time computed from the whole article.
- Security
Webhook deliveries scoped to the subscribing organisation
Intelligence API event deliveries are scoped through the key that owns the subscription; a subscription with no organisation receives nothing rather than everything. The same pass made the public template library and the privacy policy answer for the site you are on, and an Intelligence API key of this edition can no longer be pointed at a dataset that is not part of it.
- Fixed
Intelligence API webhooks create again; copilot matter search degrades gracefully
Creating a webhook subscription had returned a server error since that morning's edition work; fixed, with tests that go through the real route. Copilot's matter search now falls back to keyword search when embeddings are unavailable instead of failing.
- Improved
Nineteen surfaces stop assuming a jurisdiction
The copilot's system prompt, its modes, personas and sub-agents and its self-check, the redliner, contract compare, the negotiation coach, role summaries, Deep Audit's synthesiser, the contract risk simulation, DD rooms, M&A diligence, the clause benchmark, obligations and intake now reason from the law your contract names, or from general commercial-contract principles when it names none. The redliner no longer supplies a default governing law when the contract is silent — silence is not consent to any law.
- Fixed
A contract under Delaware, Ontario or DIFC law reaches its own rules
Retrieval was dropping the state or territory from a governing-law code, so a contract the resolver had correctly placed under a sub-national law reached nothing. The clause benchmark with no comparable clauses now says so instead of pointing you at another market's practice, and your own uploaded clauses are never filtered out of your own search by jurisdiction.
- Fixed
Your data rights are not a plan feature
Exporting or erasing what the assistant remembers about you works on every plan — a paying customer had been asked to upgrade to exercise a GDPR Article 15 or 17 right. Creating and editing remembered facts stays a paid feature; reading, exporting and erasing them are rights.
- Fixed
Every email carries this product's branding and links
The verification email, team invites, payment receipts and dunning notices now render with this product's chrome and point at this product's sign-in. A team invite had pointed at a different sign-in host, which would have created your colleague's account on the wrong product; a receipt had failed to send at all.
- Improved
The plan card lists only what this product includes
Three features that had been listed on every global plan while none of their endpoints existed for this product — AI research, a drafting critic and negotiation coaching — were removed from the card. Nothing you could reach has changed.
- Security
Air-gap covers every vendor, and the register lists them all
An air-gapped deployment now refuses Anthropic, Deepgram, LlamaParse and Twilio on the same hard gate, not on whether a key happens to be present. Deepgram and Twilio are named in the privacy policy and the sub-processor register; Deepgram's entry declines to name a processing region the integration does not select.
- New
Statutory claims are verified against nine official sources
Australia (Federal Register of Legislation), Hong Kong (Department of Justice datasets) and Ontario (e-Laws) join Great Britain, Ireland, the EU, California, Germany and the Netherlands — the Dutch adapter confirms a citation exists but never marks it verified, because there is no official English text to compare wording against. A claim is verified only when the provision text is retrieved from the official source and supports it; an outage can never produce “contradicted”. DIFC is a written refusal: its terms of use forbid what an adapter does.
- Security
54 dependency advisories cleared, 14 held with written reasons
Pillow, PyJWT, cryptography, python-multipart, lxml, aiosmtplib and json-repair updated, ranked by how reachable each is from an upload or a session token. Of the 14 held, nine cleared the same day with the move to FastAPI 0.141 and Starlette 1.6; the five that remain are developer tooling and a renderer held per its own note, none on a request path.
- Platform
Every deploy is gated on the test suite
The full suite runs before anything reaches production, and a security scan — dependency audit, static analysis, secret detection — reports on every pull request.
- Improved
The welcome email states your joining offer, in your currency
The offer's window, token grant and end date lead the welcome email, read from the terms your organisation actually holds, and the plan table in it is priced in USD. Google sign-ups receive it too — they had received no mail at all.
- Improved
The plan wall says the true thing
A plan wall now tells the difference between an allowance you never had and one you used up — with the numbers and the reset date — and links to this product's upgrade page. New accounts are no longer told they have used all of a zero allowance.
- New
Write for the blog from the app
My posts, an editor with submit-for-review, a public author profile and a reviewer's queue. Authors never publish — an editor does, choosing the edition each post appears on. Reviewers and authors are notified at each step.
- Fixed
The free scanner keeps to what it promises
Three scans a day and 25 a month per caller, grounded and scored like the signed-in pipeline. Its upgrade line reads the joining offer that sign-up will actually apply, and says nothing when there is none.
- Fixed
A token grant from support no longer breaks the first analysis
An organisation granted extra AI tokens by support hit an error on its first analysis; fixed. Unhandled errors now come back as a proper response the app can show, instead of the browser's bare “Failed to fetch”.
- Improved
Checkout collects the tax facts and asks for consent
Stripe checkout now takes your billing address and a validated VAT or tax ID — the location evidence and reverse-charge basis the rules ask for — and shows the 14-day withdrawal acknowledgement at the pay button, recorded with your consent.
- Platform
global.lexvio.ai is the canonical host
Invites, sign-in links and adapter identities point at global.lexvio.ai; an invite to a global organisation can no longer land on a different sign-in host. Consent honours Do Not Track and Global Privacy Control, and Accept grants analytics only. Billing → Invoices shows commercial invoices with their currency and totals.
- Improved
The GB and IE adapters read the real statute book
Titles like “Law of Property (Miscellaneous Provisions) Act 1989” now parse on the GB adapter, and it reaches legislation.gov.uk reliably. The Irish adapter no longer calls pre-1922 law fabricated — the Statute of Frauds (Ireland) 1695 is real, in-force law the index simply does not cover — and defers instead.
- New
global.lexvio.ai is live
LexVio's global edition: a separate product on the same platform. Plans are priced in USD and charged through Stripe; invoices are commercial invoices in their own series with no computed tax; currency is a property of your organisation, and a converted amount is never quoted. Shared content — clause standards, playbook defaults, statutes — reaches your organisation only when it is tagged for your jurisdiction; an unknown jurisdiction sees none rather than someone else's.
- Security
An account signs in only on its own product
Signing in to the other edition's site is refused with a pointer to the right one — after the password check, so the refusal cannot be used to probe for accounts. Phone and API sign-ins are untouched.
- New
A joining offer, kept
A live offer sets your trial window and a monthly token grant at sign-up, and its terms are copied to your organisation, so a later change to the offer cannot change what you were promised. One live offer per edition at a time.
- Fixed
A retried payment cannot produce a second invoice
A payment gateway that retries a webhook is answered once: the event is claimed, processed and only then marked done, so one payment yields one invoice. The per-day AI cap counts in-flight work so concurrent analyses cannot overshoot it, and reminder emails are never re-sent after a crash mid-sweep.
- Improved
Scores a short contract can survive, and pages on every span
Missing clauses no longer swamp the score: the k-th omission of a severity costs less than the one before, so a six-clause NDA is not scored like a catastrophic contract. Every grounded span carries its page for a PDF; none is invented for DOCX or pasted text. Rate a finding and your verdict survives a re-analysis.
- New
Every finding carries its evidence
The explanation behind a finding cites the source span it rests on — with its page number where the source has pages — and the statute references it resolved; a suggested fix is marked unverified, because a claim about what should be in a contract is one the contract cannot evidence. A clause the model quoted that is not in your document is no longer silently dropped: it is reported as contradicted, excluded from the score, and never rendered as advice. One verdict vocabulary — verified, unverified, contradicted, ambiguous, not checked — across citations, clauses and statutes.