What shipped in WC.
Platform release notes for JustineAI™ WC — changes to the product itself, not to this website. In market. For changes to the site, see the site changelog; for the other in-market edition, see the PI release notes.
Security, access-control, and confidential-data handling changes are not itemised publicly. Justine works on privileged matter files and medical records, and for that kind of system a public list of which control changed and when is an attack map rather than a release note. Customers receive the full detail through their engagement, and our security posture and disclosure process are documented in the Trust Center.
- Early September 2026
Client money on its own ledger, and a three-way reconciliation
Trust accounting is now a surface rather than a spreadsheet somebody keeps: the firm’s trust accounts, their balances and the full transaction ledger in one place, with each matter’s sub-ledger on the matter file itself. Reconciliation is three-way — the bank statement balance, the book balance and the sum of the per-matter sub-ledgers set against one another, with the outstanding items that explain a difference listed rather than netted away, and completed reconciliations kept and reopenable. Two things this deliberately is not: it is a records surface and not an audit, certifying nothing against any state bar rule; and the statement balance is typed in by a person, because no bank is read and there is no bank feed. A reconciliation completed over a difference nobody explained is recorded as forced rather than clean, so the distinction survives in the record.
- Early September 2026
The injured worker can see their own case
An invited client can now sign in to a portal scoped to their matter and see where it stands, the injury type, the dates that matter and the firm acting for them — plus any document the firm has explicitly released, one document at a time rather than by opening the file. What a client is never shown is enforced on the server, not hidden in the page: case value, settlement ranges, the rating figure and the limitations date are never sent. There is no self-registration; access starts with an invitation the firm issues, and where an email address appears on more than one matter the portal refuses to guess. Said plainly because it is half-finished: issuing that invitation and releasing a document are API operations today with no screen in the workspace, and the product does not email the invitation — the firm passes the link on itself.
- Early September 2026
Deposition testimony, addressable by page and line
Transcripts come into the matter and every passage afterwards is quotable as page:line. This takes excerpts — roughly forty pages at a time — rather than full volumes, and plain text rather than PDF, so a PDF transcript still has to be converted first. The more useful half is what happens when it cannot read one: the parser refuses instead of guessing, and the refusal is a report saying what it read, where it stopped and which reason applied, in line numbers with no testimony quoted back. A refused transcript stores nothing at all, including the pages that read cleanly, so a partial ingest cannot later be mistaken for a whole one.
- Early September 2026
Search across the book, and across one matter
Search runs in two scopes. Across the firm it reads the firm’s documents and its matter records; on a single matter it also reads that matter’s deposition testimony and case notes. The scopes read different corpora on purpose and the page says which are in play, because one search reading everything is the thing people will assume otherwise. Every hit carries where it came from — testimony to page and line, a document result to the document, a matter-record hit to the record it matched on rather than to anything inside the file. Two limits render beside the results rather than in a footnote: document search is firm-wide and cannot be narrowed to one matter, so a document ranking below the strongest matches will not appear on that matter; and a document that has been uploaded but not yet analysed is not searchable, so those files are counted and named instead of being quietly absent.
- Early September 2026
Bring an existing book of matters in
A firm arriving with matters already open can import them from a CSV or TSV export, columns matched by header name, the file read and checked before anything is written. The honest description of what it then does: it creates matters one request at a time, and it is not a transaction. A file that fails on row 174 leaves the first 173 as real matters in the book — there is nothing to roll back — and the run belongs to the tab you started it in. Re-running the corrected file is the recovery path, and the importer recognises the rows that already landed. Up to 2,000 rows at a time.
- Early September 2026
Records requests, document blockers, and one alert you cannot mute
Records requests are tracked on the matter with their status and deadline, and the QME packet now shows what is blocking it — records that cannot go in because service was never recorded or the page count is missing. Two things are stated rather than implied: the past-due flag is derived when the page is drawn, because nothing on the server marks a request overdue on its own; and Justine can report how many records are blocked but cannot yet name each one. Separately, the notification centre gained a tier that cannot be switched off — a statutory deadline warning reaches you past a muted preference, a global unsubscribe and quiet hours. Exactly one kind of notice carries that status, and the override beats a preference and never a consent: with no SMS consent on file the warning arrives in the app and by email rather than by text.
- Late August 2026
Nine named agents, and an approval inbox they cannot get past
Justine now works the book through nine named agents — a docket sentinel, inbound mail, a benefit reconciler, a QME packet assembler, correspondence, a standing red team, book of business, client communication and fee leakage. What they may do is not a matter of prompt wording: one authorization step classifies every action, and while reading, computing, drafting, saving and telling someone inside the firm are things an agent may perform, anything that sends outward, files outward or touches money can only be queued for an attorney. That queue is the new Approvals screen, alongside the record of what Justine declined to do. Agents start in shadow — proposing and logging, executing nothing — and five of the nine are promoted to live today.
- Late August 2026
The §4610 utilization-review clocks
The UR clocks are now computed with the rest of the docket: the five-business-day decision, the 72-hour expedited clock, the outer 14-day cap, the physician-notification and written-notice clocks, and the §4610.5 independent-medical-review clock that follows a denial — each carrying the statute or regulation it comes from. Business days and court days are counted as genuinely distinct from calendar days against the judicial-holiday calendar. A computed date remains a docketing aid you confirm, not a certification.
- Late August 2026
Kite alongside the Combined Values Chart — and the delta between them
A rating now returns both totals: the Combined Values Chart figure and the Kite additive figure, each region having run its own full adjustment chain first, with the difference reported in disability points and — where the average weekly wage is known — in dollars. Justine selects neither. Kite is persuasive panel authority available only where substantial medical evidence shows the disabilities do not overlap, and that is a finding for the attorney and the judge. The delta is the argument, so the delta is what she puts in front of you.
- Late August 2026
Where an apportionment opinion does not carry its burden
When the physician’s stated reasoning is supplied with the rating, Justine screens it against the substantial-evidence standard and reports where a bare percentage or an unexplained split falls short, along with what it is worth. The screen is report-only: it never quietly restates the rating, and where the answer is genuinely indeterminate it says so rather than resolving it in either direction.
- Late August 2026
Drafts you can open, edit, version and export on your own letterhead
A draft is now a document rather than a transcript: open it, read it, edit it in place, and save the edit as a new version with the original preserved. Export as Word or PDF rendered on the firm’s own letterhead, with the firm’s not-legal-advice notice applied by the renderer rather than written by the model — and a plain warning when no letterhead has been set, instead of invented stationery. There is deliberately no Send button on the editor; outward correspondence goes through the approval inbox, and only if the firm has switched sending on.
- Late August 2026
A Spanish letter is checked for being in Spanish
Which letters may be bilingual is derived from who receives the document, so a filing does not quietly become bilingual and a client letter can. When Spanish is asked for, the composed prose is checked to confirm it actually came back in Spanish, and the job fails loudly rather than handing over an English letter labelled Spanish. This is a check on the language and not on the quality of the translation: a person who reads Spanish should still read it before it goes to the client.
- Late August 2026
Firm settings that reach the work
A number of firm and personal settings were recorded but never read by anything that mattered. Capability toggles, the default letter language, the record-retention period, the default service method and the notification switches are now genuinely load-bearing — the toggle changes the behaviour, and a test asserts the behaviour rather than the wiring. One setting was deliberately refused rather than connected: a firm cannot switch off court-day counting for a statutory computation, because a preference that shortens a jurisdictional deadline is not a preference.
- Late August 2026
The deadline calendar
Every jurisdictional clock across the book now has a calendar view alongside the list, so a month with three overlapping statutory dates reads as a month rather than as a queue.
- August 2026
How impairments combine, drawn
The Combined Values Chart is now shown as a visual: how each impairment folds into the combined figure, in order, rather than a single number at the end of the arithmetic.
- August 2026
The disputed band between competing apportionment opinions
Where two medical-legal opinions apportion differently, the range between them is displayed as a disputed band instead of one figure being silently chosen. What is contested reads as contested.
- August 2026
Georgia benefit corrections
Two Georgia computations were wrong and are corrected. A claimant earning under $50 a week was rated against the weekly minimum as though it were a floor on the award, which overpaid the figure; and the death-benefit caps under § 34-9-265 were not loaded from the sourced schedule, leaving a 2024 date-of-injury claim computed about $50,000 light. Both now compute from the sourced statute with the section cited.
- August 2026
Statute-of-limitations periods that are not whole years
A limitations period expressed in fractional years was truncated to the whole year, ending the computed period up to six months early. Periods are now computed to the month.
- August 2026
The rating engine declines to guess
A date of injury later than the newest published rate row used to extrapolate off the end of the table; it now declines to compute and says why. The same applies to deadline and gap analysis on a jurisdiction the engine is not calibrated for — it reports that it cannot answer rather than answering with California’s rules.
- August 2026
Georgia and Texas guidance re-grounded on primary sources
The Georgia and Texas guidance was re-checked line by line against the statutes and official schedules themselves, and corrected where it had drifted. Texas rates on the AMA Guides 4th and Georgia on the 5th; each state’s answer now cites its own controlling section.
- August 2026
Documents are classified against the right state
An incoming workers’-comp document is classified against the state whose law governs the claim, rather than being read as a California filing by default.
- August 2026
Indemnity reconciled into a benefit ledger
Indemnity paid, owed and projected are reconciled into a single ledger per claim, so what has been paid and what is outstanding can be read off one surface instead of being assembled by hand.
- August 2026
Every surface says who you are working with
Justine is a Digital Employee, and every surface where she produces work now says so — the disclosure is part of the surface, not a setting. The attorney reviews, attests and signs; rating and settlement figures are projections the attorney verifies.
- July 2026
Georgia and Texas benefits computed in the app
Georgia PPD / TTD / TPD and death benefits, and Texas TIBs / IIBs / LIBs with the 90-day impairment-rating finality deadline, are computed and displayed alongside California — three states, each on its own law rather than California’s applied elsewhere.
- July 2026
Which state governs
Where more than one state could take jurisdiction, a governing-jurisdiction panel sets out the competing bases and which one governs, with the reasoning shown. It is a research aid the attorney verifies, not a determination.
- July 2026
California rating on the official basis, labelled as an estimate
The California rating captures the inputs the official schedule actually uses — occupation group, age, impairment number and variant — and the Value Engine now reads the verified server-side rating rather than computing its own. Permanent-disability dollars are labelled as estimates everywhere they appear, and a $0 rate override is ignored rather than treated as a rate.
- July 2026
A separate clock for each injury claim
Each injury claim on a matter carries its own Labor Code § 5405 clock on an injury-claims card, because they do not run together, and a case-facts card holds the facts those clocks depend on where a human can check them.
- July 2026
The parallel work is visible
Valuing a case, planning a medical-legal QME strategy and red-teaming a case each run as a visible panel of parallel specialist reasoning rather than a single opaque answer, and a capability library sets out what Justine can be asked for.
- July 2026
A guided tour and the manual in the app
A first-run tour walks the workspace and the owner’s manual is available inside the app. The tour was later corrected so its card cannot sit off-screen on a phone and cannot swallow a click meant for the page underneath.
JustineAI™ is decision support bound by the ABA Rules. The attorney reviews, attests and signs every output; rating and settlement figures are projections the attorney verifies.