User guide

Open the demo case: it already has materials, a recorded debate and case hypotheses, and nothing there costs model spending. Then create your own case with «New case» on the main page and upload materials under «Case materials». Until you finish the first steps, the main page shows a checklist with a direct link to each action.

An administrator invites people to the firm under «Administration» → «Invitations»: email and role. A lawyer, partner or administrator can invite a colleague straight into a case with «Invite to the case» on the case card. The colleague gets a one-time link valid for 72 hours and sets their own password; if email is not configured on the server, the screen shows the link and a «Copy» button — pass it on yourself. What each role can do is described under «Team and access».

The «?» icon in the header of every screen opens the help section about that screen. The search box at the top of the help searches questions and answers. If there is no answer, press «Didn't find an answer?»: the page you came from, your role, organisation and server version are attached to the message automatically.

In the case card, besides “Materials” and chat, “Facts”, “External sources”, “Graph, contradictions, timeline” and “Deadlines” are at work; the main menu has a “Deadlines in all cases” page. The “Calculators” section will arrive in an upcoming update — the interest and court-fee calculations are described in a separate help section. Every section greets you with a «?» icon carrying its own article.

PDF, scans and photos (PNG, JPG, TIFF, BMP, GIF, WEBP), DOCX, XLSX, PPTX, ODT, ODS and ODP documents, plus TXT, MD, CSV and RTF. ZIP and TAR archives (including .tar.gz) are unpacked automatically — up to 2,000 files and 2 GB per archive. Old .doc/.xls/.ppt, HEIC and RAR are not accepted: re-save the file as DOCX or PDF. For cases with hundreds of volumes it is easier to link a Google Drive or Dropbox folder with subfolders «Volume 1», «Volume 2»…

A small document is processed in minutes, a scan of hundreds of pages takes longer: pages are recognised first, then facts are extracted from them. The «Coverage» block on the case card shows in numbers how many pages are processed, how many await manual OCR review and how many claims have no citation. While unreviewed pages remain, answers about them may be incomplete.

Only pages where two independent recognition engines disagreed and the disagreement could not be resolved automatically go to manual review. Open «OCR review» on the case card, compare the text with the scan, correct it and save. After saving, the system extracts facts from that page again, and further answers rely on the corrected text.

Dates and the «final decision» mark are set under «Case materials» → «Document dates». They drive the «case as of date»: when building hypotheses, the model does not see documents after the chosen date or the final decision — otherwise it would retell the verdict instead of analysing. A document without a date stays visible in the «case as of date», so it is worth setting dates.

The model may only refer to numbered fragments it was shown; the volume, page and quoted text are inserted by the software — the model cannot invent a page number. Click a citation label to see the fragment itself. If the model still referred to something outside the fragments it was shown, the answer is marked «⚠️ The model cited a source outside the confirmed context of the case»: do not rely on that citation without checking it.

Citation checking verifies whether the fragment supports the claim: supports, contradicts or is silent (talks about something else). If the fragment does not support the claim or the check is not confident, the claim gets a «needs review» mark. Look at these first — the citation itself is genuine, only its link to the conclusion is in doubt. Citation checking is on by default; the owner can turn it off in the Lab settings.

A normal question gets one answer from the materials. In a debate several models test a position: the analyst argues for it, the sceptic (and, in three-way mode, the opponent) looks for weak points, and the judge gives a verdict on each claim. A debate takes a few minutes: replies appear on screen as they are ready, and you can see the current step and elapsed time. Start one on the «Debate (position check)» tab of the case chat.

«Held» — the claim is directly supported by the cited fragments. «Adjusted» — broadly right but with a caveat (an inexact date, an incomplete list). «Disputed» — the materials contradict each other and there is no clear conclusion. «Rejected» — the fragments refute the claim or say nothing about it. All claims of the case with their statuses are collected in the «Risk map».

It is not an error but an honest warning. Normally 2–3 check claims with a known answer («canaries») are mixed into the dispute, and if the judge «confirms» something known to be false, you see it in the «canaries» line. For some questions there are no suitable canaries — then the system says «judge control unavailable» instead of «all fine». The verdict itself was produced as usual; there was simply nothing to test the judge's attention with this time.

While work is in progress, the screen shows the step and the time. If there is no news from the server for several minutes, a warning and a «Stop waiting» button appear: the task continues on the server, and the result will appear in the case's debate history. If the debate has not appeared after half an hour, check the «Status» page and contact support via «Didn't find an answer?».

One of the roles (for example, the sceptic) did not complete — the model did not answer or refused. The system does not silently replace it: the debate is marked «Incomplete line-up» and lists the roles that did not complete. Such a verdict carries less weight; repeat the debate later if needed.

Hypotheses are competing explanations of events from different angles: prosecution, defence, a neutral investigator and a mandatory «alternative» one. For each you see which fragments support it and which contradict it (the «fragment × hypothesis» matrix), and which document to request to tell the hypotheses apart. Hypotheses are analysis for the lawyer: they never enter the case fact graph. For closed cases turn on «Case as of date» so the model does not see the final decision.

«Compare two hypotheses blind» shows a pair without the author model and angle, in random order; you choose which one explains the materials better. Each choice recalculates the Elo rating of all hypotheses of the case: the higher the Elo, the more often the hypothesis won comparisons. «No Elo yet» means it has not been compared. A weak hypothesis can be «Developed with another model» — a new branch appears in the hypothesis tree.

The model only notices a deadline in the text and suggests it — with a citation showing where it comes from. The final date is calculated by the software from inputs confirmed by a person: the start date and the number of days. A suggested deadline means nothing until a lawyer, partner or administrator presses «Confirm the deadline»; a wrong one is dismissed with «Not a deadline». Deadlines can also be added manually.

Reminders about confirmed deadlines arrive in Telegram — in the chat the administrator configured on the server: 7, 3 and 1 day before and on the day. If a deadline is confirmed when less time is left, one message arrives right away. You can ask the bot for upcoming deadlines in your cases with /deadlines in a private chat. Suggested but unconfirmed deadlines do not trigger reminders.

The “Deadlines” section in the main menu gathers confirmed deadlines for the next 30 days across all cases you can see: “Overdue”, “Today”, “Within 7 days” and onwards by date. The “Add to calendar (.ics)” button downloads a file that opens in the calendar of a phone or a computer. Deadlines proposed by the model reach this page only after a lawyer confirms them.

If a deadline is stated in working days or in months, the system warns: the days added are calendar days, and the number of days to the end date is worth adjusting yourself under the rules for calculating time limits. When the case's jurisdiction has a verified production calendar, a deadline in working days is computed with it, and the confirmed deadline carries the hint “computed with the production calendar” — the calendar's name, source and data date. The calendar is a source like any catalogue: a lawyer declares it reliable, and an unverified one is not applied.

For cases in the Ukrainian jurisdiction the case card has a “Court monitoring” block. Once a day the server checks the tracked case numbers against the open data of the State Judicial Administration on data.gov.ua: the court decisions register and the list of cases scheduled for hearing. The register website is closed by a CAPTCHA and the system does not bypass it — only the open datasets are used. The block is not shown to the case client.

A lawyer or an assistant puts a number on tracking explicitly: with the “Track” button next to a number found in the case materials, or by typing it in, for example 757/12345/23-ц. Nothing is tracked automatically — numbers found in the materials sometimes belong to other cases. The “Stop tracking” button removes a number.

Findings appear in the “Found in the registers” block with the source and the export date; a decision has a link to its text in the open data. A hearing date reaches “Deadlines” only as a suggestion: a lawyer confirms it like any deadline. A decision taken off the register is marked “removed from the register”, and a failed check of a source is shown right in the block. Findings do not enter the fact graph by themselves.

Every generated document is born as a draft: a person files documents, not the system. A lawyer, partner or administrator can approve or reject a draft; an assistant cannot. Every approval is recorded in the log: who and when. There is no automatic filing or sending.

From the organisation profile: full name, address, registration code, bank details, the signatory with their position and authority, and the logo. The administrator fills in the profile under «Administration» → «Organisation». If a field is empty, the document gets a visible placeholder, not an invented value.

Every draft has «DOCX» and «PDF» buttons. The file contains a header with the firm's logo and details, the document text and the signatory from the organisation profile. Until the document is approved, every page carries a «DRAFT» watermark. Exporting does not change the document's status, and «[ПРОВЕРИТЬ НОРМУ]» (check the law) marks are kept as they are.

The model cites a legal provision only when it is sure of its details (type of act, number, article). If it is not sure, it writes «[ПРОВЕРИТЬ НОРМУ: …]» (check the law) instead of an invented number. A note after the text lists all provisions used — check them and replace the marks before approval. You can check which edition of a provision applied on the date of events in the «Law editions» block of the case card.

Labelling shows which model can be trusted in which role on your firm's cases — the choice of model is made by data, not advertising. The first week takes about an hour a day, roughly 140 claims, of which 25 are labelled by a second lawyer independently of the first. The counter on the «Review» screen shows how much is done this week, and «Analytics» shows how many labels are still needed for a reliable conclusion.

The main rule: ask whether the claim is supported by the cited case materials, not whether it matched the court's decision. «Held» — directly supported by the fragments; «adjusted» — right with a caveat; «disputed» — the materials contradict each other; «rejected» — the fragments refute it or are silent. Aim for no more than 20 seconds per claim; skip a hard one and note why in the comment. Details are in the Lab guide (in Russian).

All amounts in the Lab are accounting cost: the model provider's price multiplied by 2. This is a margin for limits and price changes (the platform owner's decision). The provider's own price is shown in a tooltip when you hover over the amount. Prices are updated daily; their history is on the «Prices» screen.

The administrator manages users, invitations, limits, the organisation profile and server settings. A partner sees all cases of the organisation, approves documents and changes a case's external AI policy. A lawyer works on their own cases: questions, debates, drafts, confirming deadlines and approving documents. An assistant uploads materials and views them but cannot approve; a client sees only their own case through the client portal.

The session ends after 36 hours without your activity — clicks, typing, scrolling. An open tab that refreshes data by itself does not extend the session: only a person's work counts. You will see the «Session expired» screen — just sign in again. If it happens more often, contact support.

«Sign out» ends the session in this browser — in all of its tabs at once. On other devices the session ends by itself after 36 hours without activity. If a device is lost, ask the administrator to temporarily disable your access under «Administration» → «Users»: a disabled user loses access on all devices at once; after re-enabling you sign in again.

Fragments of materials go only to the provider of the model answering the question. Every case has an «External AI policy»: «all providers», «no resellers» (the default: official APIs and aggregators) or «local only» — materials never leave your server. You can also restrict processing regions there. A call that violates the policy is rejected with an explanation; an administrator or partner can change the policy.

On your firm's own server, materials are stored only on that server. In our cloud each organisation is isolated: staff of another firm see neither your cases nor their titles. Inside the firm, cases are visible by role: partners and administrators see all cases of the organisation, lawyers and assistants their own, a client only their case. Significant actions (approving a document, confirming a deadline) are recorded in the log.

Model spending can be limited for the whole organisation, a staff member or a case — per week, per month or without a period. When a limit is used up, new model requests are rejected with «Limit reached», while work already done is kept. The administrator raises organisation and staff limits under «Administration» → «Limits and spending»; a case limit can also be changed by a partner or the case owner.

«Administration» → «Limits and spending» shows the organisation's spending by case for 7, 30 or 90 days or all time. «Breakdown» next to a case lists every model call: time, role, model, tokens and cost; it can be downloaded as a CSV file for billing the client. Lab amounts are counted separately — they are accounting cost ×2.

Case expenses are re-invoiced to the client, and the firm may add a markup: the case breakdown has a «Markup for the client» field — a percentage from 0 to 500. The downloaded CSV file gains a «to_client» column — the amount with markup for each row — and a total line for the client's invoice. The firm chooses the markup; the system only computes.

If the firm has its own server, the administrator creates a one-time pairing code under «Administration» → «Pairing code». Open the mobile app and scan the QR code on the screen or type the code, then sign in with your email and password. In our cloud no code is needed — email and password are enough in the app.

In your profile press «Connect Telegram»: the screen shows a one-time code and the command to send to our bot in a private chat. After that the bot knows you: it sends notifications and opens the Telegram app with the list of your servers. Case names, surnames and deadlines never pass through the bot — only «there is an update in a case» and a button that opens the case on your firm's server.

A Telegram group can be linked to a case: in a private chat with the bot get a code with /case_chat_code, then send /link_case with that code in the group. In a linked group members ask questions about the case with /ask and see a summary with /status; unlink with /unlink_case. A firm can also connect its own bot — the administrator sets it up.

You add a fact about a person or company yourself on the case's “Facts” tab and choose its level: confirmed by a case document, official registry, document requested, according to the client, according to a witness, media publication, obtained without a supporting document, lawyer's hypothesis or refuted. There is no default, so the level is always a deliberate choice. The system never raises a level on its own: an answer to a request only suggests raising it, and a person does it. A collapsible “Reliability levels” legend is on every facts screen.

Information without a supporting document carries a purple badge with a dashed border and a ⚠ sign. In chat and debates such a fact is always labelled with its level and never presented as established. It gets into a court draft only as a “[CHECK: fact without a document …]” marker, and when you export the file you see how many such places remain. Facts never go to the client portal, Telegram groups or notifications.

Open the fact card: a fact below “confirmed by a document” has a “How to confirm with a document” block with the procedure, legal basis, addressee and response term for the case's jurisdiction. The “basis not checked by a lawyer” mark means nobody has verified that catalogue line yet. “Create a draft” prepares the request or motion, and marking it “Sent” and “Answer received” gives a suggested deadline and a suggestion to raise the fact's level. You confirm the deadline under “Deadlines” and raise the level yourself in the fact card.

On the “External sources” tab, search a connected registry and press “Save to the case”, or upload an extract or screenshot from a database the firm has access to. Each material is stored with a snapshot, its SHA-256 and the request basis, and the basis is required for information about an individual. A material enters the case facts only after you press “Create a fact”. “Check against registries” queries open registries about a person in the case and separately shows the sources that did not respond.

The section computes money by ready-made formulas: interest for using another party's funds under Article 395 of the Russian Civil Code at the Central Bank key rate, inflation indexation plus 3% per annum under Article 625 of the Civil Code of Ukraine, the Ukrainian court fee, and interest on a judgment in England and Wales (8% per annum). The final figure is always computed by the program from formulas and rate catalogues — the language model takes no part in the calculation and cannot invent a number. Jurisdiction and currency come from the case. If you do not see the Calculators section in the case card yet — it will arrive in an upcoming update.

Besides the final amount, the result shows the steps of the calculation: the debt amount, the rates applied, the periods they were in force and the interest for each period. Next to it the data sources are listed with dates — for example, the key-rate history showing from which to which date each value was in force. A source earns the “verified” mark only after a lawyer checks it.

This is an honesty label, not an error: nobody at the firm has yet verified the catalogue line — rates, thresholds, indexes — against the primary source. Until it is verified, the result serves as an estimate; for a claim or a defence the basis is worth re-checking. Once a lawyer verifies the catalogue, the mark is lifted for all later calculations.

For interest you need the debt amount and the period: from which date to which date to accrue. The Ukrainian court fee asks for more inputs: the claim amount, the type of claim, who pays, the filing date and the instance. How the last day of the period is counted and why the rate changed within it is visible right in the calculation steps.

If the data for a calculation is missing — the catalogue is not filled in yet or has no dates for the needed period — the system says so directly, naming the source, and does not produce an approximate figure. Silently substituting missing data with approximations is not allowed. Missing data is entered by the server administrator who maintains the catalogues.

The fee is computed from the Supreme Court rate table: the system takes the claim amount, the claim type (property or non-property), who pays (individual, sole trader (FOP) or legal entity), the filing date and the instance. For property claims a sole trader pays as an individual, for non-property claims as a legal entity. Minimum and maximum are based on the subsistence minimum as of 1 January of the filing year. Appeal is 150% of the rate, cassation 200%, filing via “Electronic Court” applies a 0.8 coefficient. Every threshold applied is shown in the calculation steps.

A calculation is an aid, not ready-made evidence: the input, rates and total are fully visible. To attach a calculation to a claim or response, press “Save to case”, then in the case card, in a draft’s “Attached calculations” block, choose it and press “Attach to draft” — a calculation is never attached by itself. In DOCX and PDF it becomes a separate appendix with a table of steps, the total and the sources, and an unverified basis is marked in the file itself. Once a draft is approved its attachments no longer change, and only the lawyer files the document.

The data sits in catalogues on the server, each with its primary source and retrieval date: the key rate — the Bank of Russia table (cbr.ru); the Ukrainian court fee — the Supreme Court rate table; Ukrainian inflation indices — the State Statistics Service data bank (stat.gov.ua), series “Consumer price index, % to previous month”, from January 2015; judgment interest in England and Wales — 8% since 1 April 1993. Inflation indices are updated monthly after the State Statistics Service publishes them (around the 10th of the following month). Russian calculators appear only in cases governed by Russian law.

The debt is multiplied by the product of the monthly consumer price indices of the State Statistics Service. Which months count follows the “15th day” rule (Supreme Court of Ukraine letter No. 62-97r; Joint Chamber of the Commercial Cassation Court, case No. 910/13071/19): if the delay began on the 1st–15th, that month counts, from the 16th — the calculation starts with the next month; if the debt was paid on the 1st–15th, that month does not count, from the 16th — it does. A delay within a single month counts only if it lasts more than 15 days.

Deflation months (index below 100%) are not skipped, but if the combined index for the whole period is below 100%, inflation losses are zero — the debt is never reduced. 3% per annum accrues for each day of delay; the payment day is excluded. Note: under credit and loan agreements, borrowers are released from this liability for the period of martial law (para. 18 of the Final and Transitional Provisions of the Civil Code of Ukraine) — the calculator does not know the contract type, the lawyer checks this.

The rate may have changed several times within the period, and the program computes by segments: each span of time uses the rate that was in force then, and the results are added up. An estimate with a single rate over the whole period therefore drifts from the total, sometimes noticeably. The exact segment breakdown with dates is visible in the calculation steps.

There are four calculations at the moment — two Ukrainian, one Russian and one for England and Wales. If your jurisdiction is missing, write to support via “Didn't find an answer?” at the bottom of the help. The set of formulas and catalogues grows together with the firm's practice.

The case card has a “Graph, contradictions, timeline” page — on the “Timeline” tab the events of the case are laid out on a time axis. Events with an exact date stand as points; those stretched over time stand as spans — for example, an interrogation that lasted several days. Events come from the case fact graph, and each can be opened in the original document.

The “Day”, “Week”, “Whole case” switch changes the window of the axis: above it you see how many events are in view and how many fall outside the range. If an event is not visible, switch the scale to “Whole case”. An empty or short axis is not a fault: it is built from the records of the fact graph and fills up by itself as the materials are analysed.

Events involved in contradictions are marked by importance colour: high, medium or low — see the legend above the axis. Conflicting events carry a “disputed” mark: the case materials say different things about the same matter. At a glance you see where the tense spots of the case are.

Some graph records have no recognised date: such events are counted in a separate line under the axis and never reach the axis itself. Nothing is lost — the records remain in the fact graph; the system simply does not know where to place them. Check the document to see which time it speaks of: perhaps the date is absent from the text itself.

The contradictions radar collects pairs of mutually exclusive statements from the materials: discrepancies in the time, place or details of one event. Each contradiction shows its importance, its type — for example TEMPORAL for time discrepancies — both citations with sources and a tactical note. You filter by importance and by type, the “Highest severity first” button raises the most serious entries, and the heading shows how many remain: “Shown: 4 of 8”. A contradiction you have worked through can be marked “rejected” so it stops cluttering the working list.

The “Export CSV” button downloads the filtered list to a file that opens in Excel and Numbers: importance, type, topic, persons, both citations, status and tactical notes. The “Print version” prepares the page for paper — for a meeting or an annex to the case position. Exactly what remains under the filters is exported and printed.

Press an event — the document page opens with the fragment the date was taken from; if the fragment is not found, the system says so directly and does not substitute foreign text. Contradictions are the same on three screens: in the fact graph, in the radar and in the risk highlighting on the time axis. A contradiction as a whole is worked through on the Contradictions tab, while the timeline shows where its statements stand in time.

A fact carries a reliability level — chosen by a person, with no default value. Its origin is described by the “Where it is known from” and “Source type” fields, the event is dated by the “Date and time” field; who added the fact and when is recorded by the system itself. Only a person changes the level: an answer to a request or an uploaded document merely suggests raising it.

This marks information a lawyer knows from any source — including an unofficial one — but for which no document exists yet. Such a fact carries a purple border with a ⚠ sign and is never presented as established: in chat and debates it is labelled with its level, in a court draft it stands behind a “[CHECK: fact without a document …]” marker, and it never reaches the client portal, Telegram groups or notifications. Next to it you see which request or motion can obtain such a document.

The fact card has a “How to confirm with a document” block: the procedure, the legal basis, the addressee, who initiates and the response term — for the case's jurisdiction. The “⚠ basis not checked by a lawyer” mark means nobody has verified that catalogue line yet; after verifying, the lawyer confirms the procedure. If the catalogue has no procedure for your category and jurisdiction, the system says so directly.

The “Create a draft” button prepares a request or a motion from the jurisdiction's template without leaving the fact card. The “Sent” and “Answer received” marks bring a suggested deadline and a suggestion to raise the fact's level — the system suggests, a person decides. A draft is created as a draft and changes status only by a lawyer's explicit action.

On the “External sources” tab, find the information in a connected registry and press “Save to the case”, or upload an extract or a screenshot from a database the firm has access to. Each material is stored with a snapshot taken on the upload date, a SHA-256 checksum and the request basis; for information about an individual the basis is required. A material becomes a fact only after you press “Create a fact”.

The “Check against registries” button queries open official registries about a person in the case. Answers arrive with the source stated, and the registries that did not respond are shown in a separate list — failures are not hidden. What comes from the registries enters the facts only through your decision.

“Lawyer's hypothesis” is a reasoned assumption about a fact: useful for building a position, but not backed by anything yet. “Refuted” means the case materials are enough to consider the fact as not having taken place. Both levels are visible to everyone in the case and honestly labelled in chat and debates.

Channels and services selling leaked personal data (databases of passports, phone numbers, registrations, banking operations) are never integrated: it is an unlawful circulation of personal data, responsibility falls not only on the service but on the user firm as well, and — most importantly — what is obtained this way cannot be used in court, because the origin of the evidence cannot be disclosed and it collapses at the first objection. The legal replacement is built into the product: official registries plus an advocate's request — the system prepares its draft from the jurisdiction's template, tracks the response term and files the received answer into the case as full-fledged evidence with provenance. What a lawyer knows from an unofficial source, the lawyer enters into the case personally — as a fact with the “obtained without a supporting document” level.

In chat and debates a fact is labelled with its level — participants see what is confirmed by a document and what is not yet. The fact card shows who added it, when and with what origin; the date and time are displayed in the usual form of the interface language. When exporting a draft, the system shows how many places with documentless facts remain — they are worth closing before filing.

When the answer to a request is received and filed, the system suggests raising the fact's level — for example, from “document requested” to “confirmed by a case document”. The suggestion is visible in the fact card, but only a person changes the level. The answer material is stored with its provenance: source, date, snapshot.

The proof matrix is a table for court: one row per factual circumstance, with columns for who carries the burden of proof, our evidence, the opponent's objections and evidence, and an assessment. You can add rows manually or press "Suggest rows": the AI lays out the facts, fragments and contradictions that already exist in the case, but it cannot reference a non-existent document. Everything suggested is marked "Suggested by AI" and becomes confirmed only after an explicit action by a lawyer.

The burden rule in a row comes from the system catalog by jurisdiction and proceeding type — for example, Article 81 of the Civil Procedure Code of Ukraine or Article 56 of the Civil Procedure Code of Russia. If your jurisdiction has no rule in the catalog, the row honestly shows "burden not defined by catalog" instead of guessing a norm. The "not verified" mark disappears once a lawyer checks the rule text against the official source.

The "Export DOCX" button builds the table from confirmed rows only; suggested but unconfirmed rows never reach the file. If a burden rule has not been verified by a lawyer yet, the document starts with a yellow draft notice. You can also attach the matrix to a document draft — it is exported as an appendix, the same way as calculations.

Time is tracked on the “Time and costs” screen in the case card: a lawyer enters date, work type, minutes and description, or takes “Suggest from activity log”, which groups case actions into draft entries. The hourly rate is filled in by code from the firm’s rates — by user or role, effective on the entry date. The cost is computed by code: rate times minutes, rounded to a cent. A draft entry never reaches the statement until a lawyer, partner or administrator confirms it.

“Draft cost statement” creates a draft with a table of confirmed entries for the period and the total — the text is assembled by code without a model. For Ukrainian law the template cites CPC 137, ECC 126 and CAC 134, for Russian law CPC RF 100 and APC RF 110, other jurisdictions have no template and the system says so plainly. Norm texts are marked “verify against the primary source” with links attached, and a place for the proportionality argument is left to the lawyer. The statement is born as a draft and is filed only by an explicit action.

Hourly rates live in “Administration” on the “Rates” tab: amount, currency, effective date and a role or a specific user. The rate with the latest effective date not exceeding the entry date applies, and a personal rate overrides the role one. Rates are set by an administrator or partner — it is the firm’s decision on the cost of an hour, not a system calculation. Without a rate the entry is saved as “no rate” and waits until a lawyer fills in the amount.

In the case card or document viewer, click "Compare revisions". Choose two case documents or paste the text of both revisions to compare. The system performs word-level clause alignment, highlighting additions, deletions, and moved clauses.

Code automatically verifies changes in deadlines, interest rates, penalties, and vague terms like "reasonable time". Additionally, the analytical model comments only on modified clauses and classifies risk categories. The final annotated report can be exported as a DOCX file or saved directly to case materials.

Select a text snippet in any paragraph of a volume sheet to open the contextual action menu. Choose color highlighting, adding a text note, or creating a bookmark. Lawyer notes appear in the document margin with a distinct scale icon and author name, clearly separated from AI findings. Private annotations are visible only to you, while team notes are shared with all colleagues on the case.

A bookmark tags an important sheet or paragraph in the case and appears in the central bookmark register for instant navigation. You can export document materials as DOCX with all margin notes and highlights included. Private annotations of other lawyers are strictly excluded from the resulting export. All annotations and notes are completely hidden from clients in the client portal view.

Go to the "Case Law" section in the main navigation or from a case workspace. Enter a legal question, case number, or statutory article (for example, art. 625 of the Civil Code of Ukraine). You can filter results by court chamber, year of decision, and outcome (upheld, reversed, or remanded for a new trial).

In the decision card, click "AI Legal Summary" to view a concise statement of the court's legal holding. Use "Attach to Case" to link the decision to your case materials with tracked provenance. Citations to court decisions in chat and debate answers are verified by code to eliminate hallucinated case numbers.

The Offline Vault allows downloading all case materials to your local device for courtrooms and detention facilities where network signals are blocked. A single click saves the text of all volume sheets, facts, timeline events, contradictions, and lawyer notes. Local full-text search works directly inside the browser using word stems to pinpoint citations without a network connection. Scanned page images can be optionally included if sufficient disk storage is available.

All stored data is encrypted client-side using industry-standard AES-GCM with a 256-bit key. The key is derived from your PIN code via PBKDF2 using over 200,000 iterations, and the PIN itself is never saved anywhere. If an incorrect PIN code is entered five consecutive times, the local copy is instantly and permanently wiped from the device. When you log out or when the configured vault lifetime expires, stored files are automatically purged.

All annotations, bookmarks, and draft facts created offline are stored in an encrypted local queue on your device. Once network connectivity is restored, queued edits are automatically sent to the server in strict chronological order. If an item was modified by a colleague on the server while you were disconnected, the system presents a conflict resolution dialog. You can compare both versions side by side and choose which one to preserve, ensuring no work is lost.

In the case card, open the "Audio & Video" tab and upload the recording of a hearing or interrogation. The system automatically extracts the audio stream and slices it into 10-minute chunks with overlap. A background worker transcribes utterances with millisecond timestamps and builds a draft protocol.

The algorithm separates utterances by speaker and proposes their procedural roles, such as judge, prosecutor, or advocate. The lawyer manually confirms or corrects participant roles according to human control invariants. Each utterance becomes a case chunk with a deterministic citation like [Audio N, hh:mm:ss] for use in chat and debates.

The system scans investigative protocol records and expert reports using strict procedural rules in code. It verifies nighttime searches without urgency markers, repetitive witnesses on different dates, broken evidence packaging chains, and conflicting order dates. All detected defects are displayed as suggested red flags linking directly to source pages and statutory articles.

The lawyer reviews suggested flags and accepts substantiated procedural defects by explicit confirmation. Accepted flags can be linked to the proof matrix assessment column or converted into a draft motion. The system generates a draft document in draft status containing the structured violations, case sheet citations, and applicable legal standards.

Each case is assigned an email address formatted as case-<code_id>@<firm_domain>. Forward court rulings, summonses, and documents directly to this address. Incoming messages land in the "Court Inbox" section in a pending state. A lawyer or partner reviews the document and confirms its addition to the chosen volume (enforcing INV-2). Emails from senders outside the allowlist are quarantined, and emails without attachments are saved as text notes.

The morning digest is sent daily at your chosen local hour (defaulting to 08:00 in your time zone). It includes procedural deadlines for the next 7 days and court hearings for today and tomorrow. You can enable or disable the digest with one click in your user profile. When delivered via the firm's bot, case titles and deadline descriptions are included. If the "Telegram Directory" channel is selected, only summary counts are sent without case names to maintain confidentiality.

The voice brief provides a quick audio summary of a case: upcoming deadlines, scheduled hearings, and the top 3 risks from the risk map. The feature is disabled by default and must be explicitly enabled by the lawyer in their profile settings. You can request a brief using the /brief <case_id> command in the firm's Telegram bot or via the web interface. Speech synthesis respects the case AI policy (lab_policy): if external TTS providers are prohibited by policy, the system gracefully falls back to returning a text-only summary.

In the case card open the “Finance” block and upload a statement: a 1C “client-bank” export, an Excel or CSV table (for a table, specify once which column is the date, amount, counterparty and payment purpose). The program reconciles balances itself: the opening balance plus turnover must equal the closing balance. If the figures do not reconcile, the statement is flagged and every conclusion based on it is shown with a warning. PDF statements are not accepted yet.

The graph shows persons and accounts as nodes and payments between them as links with amounts for the period. Indicators (transit, split payments, cash-out via sole traders, withdrawal of funds just before bankruptcy) are found by the program using rules with adjustable thresholds. They are hints for the lawyer to check, not a finding of wrongdoing. On request, AI can suggest the category of a payment purpose — it is marked as “AI category”.

In the case card open the “Cross-examination cockpit”, type the witness’s surname and press “Prepare dossier”: you will see everything they said in the case — in the records and in the recording transcripts — the contradictions they are part of and the prepared questions. Citations “Volume, Page, Paragraph” and “Audio N, hh:mm:ss” are taken straight from the materials. The “Save the dossier” button moves the questions into a list where you can edit them and mark “Use in court”; the mark is set by a lawyer, partner or administrator. Saved dossiers go into the offline copy of the case.

While the witness is speaking, type a few words from the testimony, for example “saw at the garage”. The cockpit immediately shows the contradiction with two sheets of the case and a question for the court, and below it the matching places in the materials; the questions you selected are above the results. The search uses no AI, and if nothing matches it says so: there are no false alarms. The case client does not see this section.

In a saved dossier press “Estimate AI questions”, then “Suggest questions with AI”: each suggested question rests on two sheets of the case, and the sheets are set by the program, not the model. Suggestions without two sheets are discarded, and the cockpit says how many there were. In court every card has an “AI clarification” button: the model compares the witness’s words with the card’s materials, and the answer shows its time and duration. If the case’s external AI policy forbids the call, the cockpit says so — the materials do not leave.

The case law section provides access to judgments and decisions of the European Court of Human Rights (ECHR) and courts of England and Wales (Find Case Law). Search works across all jurisdictions or by specific source with filters for Convention articles and neutral citations. When HUDOC data dumps are restricted by anti-bot protections, lawyers can add ECHR judgments manually with mandatory verification (INV-2, INV-8). For matters with international elements, the platform automatically suggests relevant authorities. Every decision receives a deterministic citation conforming to uniform rules (INV-1).

When generating a legal costs statement, a lawyer can select up to three relevant authorities on fee proportionality. Court holdings are included in the draft statement only upon explicit selection by counsel, featuring deterministic citations and marked AI summaries. Total fee calculations are performed deterministically by code without AI guesswork.

The verification tool scans procedural draft text for citations to Supreme Court, ECHR, and English court decisions. Each detected authority is checked against the local index, assigning a status: verified, not found, or mismatch in date or court name. If an exact match cannot be confirmed, the draft clearly indicates that manual review by counsel is required.

The outcome of each English judgment is classified deterministically by code from the judgment text without relying on language models (INV-3). The system identifies allowed and dismissed appeals, partial allowances, dismissed claims or applications, and granted permissions to appeal. When standard judicial wording is absent, decisions are categorized as other outcomes.

The platform synchronizes judgments from the UK Supreme Court, Court of Appeal Civil and Criminal Divisions, and Chancery and King's Bench Divisions of the High Court. The target courts list is configurable through platform environment settings. Judgments are imported incrementally while retrieving full official texts directly from the source.

The system extracts transaction tables from PDF bank statements using a built-in document layout parser. For recognized bank formats (PrivatBank, Monobank, Oschadbank), column mapping is performed automatically. If the statement format is unknown, AI proposes column mappings through a safe interface, and the resulting template is saved for future files. Balance reconciliation (opening balance + turnover = closing balance) is computed strictly by mathematical code without neural networks (INV-3). If statement balances do not match, the system flags the import with a warning for the lawyer (INV-7).

Counterparty details are verified against the official EDR open registry (data.gov.ua) without third-party lookup services or CAPTCHAs. Legal entity datasets are imported via a streaming parser weekly, retaining source metadata and export dates (INV-8). The counterparty card is populated with entity name, status, registration date, KVED, address, and director. If a payment was made to a terminated entity, the system creates an automatic warning for review by counsel. Unmatched codes are never fabricated by the system (INV-7).

The program flags payments to terminated entities, transfers to counterparties registered less than 90 days before the operation, and shared registration addresses or directors among payees. All indicator detection rules are implemented in code, and thresholds are customizable in case settings. Risk indicators serve only as prompts for legal evaluation by the lawyer and are not conclusive judicial findings. For unreconciled statements, indicators are generated with a warning regarding potential data incompleteness (INV-7).

In organisation settings (administrator), specify your firm's custom domain, e.g. client.lawfirm.com, and copy the verification token. Add a DNS TXT record _cassation-verify.<domain> with this token and a CNAME record pointing to the platform server, then click "Verify TXT record". Once domain ownership is confirmed (custom_domain_verified=true), the Caddy server automatically issues an SSL certificate for your custom portal domain.

In the case view, generate a client invitation link and send it via your preferred channel. The client authenticates via Telegram using the link or on your custom domain. They will see the case dossier without internal team notes, court session schedule, and deadlines. Internal team materials remain strictly private (INV-8).

On the portal, the client can upload a file directly to their case. The uploaded file arrives in the "Client Inbox" for lawyer review and does not automatically become an official case document (INV-2). The lawyer reviews the file and clicks "Approve" or "Reject". Upon approval, the file is added to case materials and becomes available for analysis.

In the cross-examination cockpit, press "Listen" — a tablet or laptop records audio in short chunks and shows the lines almost immediately, usually within seconds. Every new line automatically runs through the "in court" search: a matching contradiction card appears right away, no manual search needed. Only a lawyer, partner or administrator can stop the recording with the "Stop" button; after stopping, the cockpit shows the average and worst-case time from "said" to "card". If the case's external AI policy is "local only" and no local recognition engine is available on the server, the cockpit refuses honestly instead of staying silent.

Before the first recording, the cockpit shows a warning: whether recording a hearing is permitted depends on the jurisdiction and the type of hearing (open or closed), and the lawyer decides — without that confirmation, recording never starts. After stopping, the recording and its transcript are saved to the case as an "Audio N" material — citations like "[Audio N, hh:mm:ss]" then work in chat and debates, just like with an uploaded recording. Only a lawyer, partner or administrator can delete the recording entirely.