Changelog

What shipped, when.

A public log of what we've added, expanded, or corrected on the JustineAI™ marketing site. The site evolves alongside the platform; the changelog makes the cadence visible.

Platform-side release notes — changes to the product rather than to this site — are published per edition: PI and WC. Security, access-control and confidential-data handling changes are not itemised on either page; your engagement carries that detail in full.

  1. September 9, 2026

    A fifth sibling, a not-live claim that had stopped being true, and where you came from

    • EliAI joins the family. The portfolio statement, the footer, the landing family band and /about/mindhyve-family now name five Agentic Operating Systems rather than four, and link to eligrid.ai — which had been linking to all four siblings while none of them linked back. The wording is copied verbatim from the shared canon rather than typed here, and the same change landed on the Arthur, Chiron and Theo sites in the same pass, because nothing in the estate detects one site stating a different portfolio from the others.
    • Push notifications came off the “what isn’t live” list for Workers’ Comp — 19 entries down to 18. The site was telling attorneys there was no push channel; the shipped WC bundle has a real subscription path and the setting is rendered unconditionally in Notifications, next to the deadline reminders. The claim was checked against the live product before it was deleted, because the other explanation for a gate like that going red is that the gate is wrong, and deleting the claim to make it pass is how a not-live list quietly turns into fiction.
    • Signup links now carry first-touch attribution to the product portals. When a firm meets us through a referral or a campaign, disappears for a week and comes back by typing our name into a search box, the channel that actually introduced us is the one recorded — not the search that closed it. No visitor id, no fingerprint, no cookie, no third-party pixel is minted: the fields carried across are the campaign parameters an advertiser put in the link, the referring site, and when you first arrived.
  2. September 5, 2026

    Six WC surfaces the site had not caught up with — and two not-live claims that had gone stale

    • The WC capability catalog was re-audited against the product repositories’ master and, this time, against what the live surfaces actually report — the WC app and its API each publish the commit they are serving, and both matched the commit that was read. “Reachable by an attorney today” is the bar this page claims to hold itself to, so it is now a checked statement rather than an assumption about deployment.
    • Six surfaces added: trust accounting with a three-way reconciliation and a per-matter sub-ledger; a matter-scoped portal the injured worker signs in to; search across the firm and across a single matter; deposition transcripts addressable by page and line; importing an existing book of matters from a CSV export; and records requests with document blockers, alongside a notification tier that cannot be muted.
    • Every one of them is published with its limit attached, in the entry rather than in a footnote: search never reads all four corpora in one query and cannot narrow document search to a matter; deposition ingest takes roughly forty-page excerpts of plain text, not full volumes or PDFs; case import creates matters one at a time and is not a transaction, so a file that fails partway leaves the earlier rows already created; and trust accounting is a records surface, not an audit — no bank is read, and a reconciliation forced over an unexplained difference is recorded as forced.
    • Two entries in the “what isn’t live” list had quietly become false, which is the direction of drift nobody looks for. The site was still telling attorneys that the injured worker gets a letter and not a login, and still counting page-and-line deposition citation as finished code with no screen. Both had shipped. A not-live list that is never re-read stops being honesty and becomes an understatement of the product.
    • The client portal is published as what it is — live at the client’s end, unfinished at the firm’s. An invited client can sign in; issuing that invitation and releasing a document are API operations with no screen in the workspace yet, and the product does not email the invitation. That half is named in the not-live list rather than rounded up into the capability.
    • No published figure moved. The headline counts are floors — 240+ capabilities, 135+ features, 90+ AI and agentic — and new surfaces leave a floor true, so nothing was restated. An exact re-census was not attempted here and no count was inferred from six additions.
    • The catalog’s group count is now derived from the groups themselves. It had been the word “nine”, typed into the headline beside an item total that was computed — one fact with two sources, in the sentence that introduces the page.
  3. August 29, 2026

    Capability re-census · a WC catalog · three figures corrected down

    • The capability census was re-derived from the live PI and WC codebases, against origin/master in all three product repositories rather than a working checkout — two of the three were sitting on branches behind master, and every count taken from them would have been an undercount. The previous census was taken on 2026-08-16, before the largest build window the product has had.
    • New /wc/capabilities — the full Workers’ Comp catalog, in nine groups. WC already had an overview, a how-it-works and release notes, and no page that inventoried what had been built behind them; PI had one. The page counts only what an attorney can reach — a route they can click, or something they can ask Justine for — and names, at the foot, the finished and tested work that has no screen yet and is therefore counted as nothing.
    • Headline counts moved up: 240+ customer-facing capabilities (was 230+), 135+ user-facing features, 90+ AI/agentic capabilities (was 80+), and 10 background automations (was 7). A fourteenth capability group was added for the agent fleet and its approval rails — nine named agents, one authorization step that lets an agent read, compute, draft, save and notify inside the firm but only ever QUEUE anything that sends outward, files outward or touches money.
    • Three published figures were wrong in our own favour and have been corrected down or scoped, rather than left standing. Statute-of-limitations coverage was published as 52 jurisdictions; the table holds 51 (50 states and DC) and always did. Compliance and security controls were published as 24; the enumerable inventory lists 14, and 24 could not be reproduced from any artifact, so the number published is now 14. The California rule corpus is stated as 399 rather than “~400”, and the statute-of-limitations table is no longer described in a way that implies it shares the rule corpus’s author-to-reviewer verification, because it does not.
    • One figure was corrected in the other direction, because it undersold the product: roughly four in five of Justine’s AI capabilities are deterministic legal and actuarial engines — statute math, rule lookup, date computation — not model-composed output. The site had been saying “roughly half”. 27 of the 35 reasoner-callable tools never reach a foundation model.
    • Correction to the entry of July 11, 2026 below: it described the public intake portal as live. The intake portal’s server side is real, but the public page and the endpoint it calls do not currently meet, so a firm’s generated portal link does not resolve end to end. The claim has been withdrawn from that entry and is not made anywhere else on the site until it does.
    • WC and PI release notes brought current for the August window: the approval inbox and the record of what Justine declined; the §4610 utilization-review clocks; Kite reported alongside the Combined Values Chart with neither selected; the substantial-evidence screen on apportionment; drafts that can be opened, edited, versioned and exported on the firm’s own letterhead; Spanish letters verified to be in Spanish; and — on PI — the fix for an interrupted reasoning stream that used to report itself as a completed answer.
  4. August 27, 2026

    Disclosure scoping synced, and a corpus gate that pinned the wrong number

    • The rule governing where a foundation-model name may appear is now synced from the shared canon rather than kept as a local override, so it cannot drift per site. The customer-facing label is the Eve engine; the model underneath is not named on customer surfaces.
    • The IPSC works count is 123, not 128 — and the gate meant to protect that figure had been pinning the wrong one, so it was enforcing the error rather than catching it. A check that agrees with a mistake is worse than no check, because it converts a typo into a defended position.
  5. August 21, 2026

    Six-month headcount growth published

    • The team figure now shows growth over six months rather than a single number, and the published band was widened to hold it. A point figure invites the question of when it was true; a band with a trend answers it.
  6. August 20, 2026

    Claim gates, per-edition release notes, and five figure corrections

    • Edition status and published figures are now gated in CI. Both had been hand-maintained across many pages, which is the arrangement that produced the five corrections below.
    • Per-edition release notes were added, so what shipped in an edition is stated where a buyer of that edition looks.
    • SOC 2 copy asserted more than the claims registry supports: it described the report as covering our controls, in the present tense, and referred to the attestation as already held. The registry lists it as pending. The copy is now future tense throughout. Claiming an attestation you do not yet hold is the single most expensive sentence to get wrong on a legal-technology site, and the gate that enforces this refuses the phrasing outright rather than trusting a reviewer to catch it — including, correctly, in this entry’s first draft.
    • Headcount is 212, propagated from the canonical source rather than restated per page. The publicly-listed count was a literal that disagreed with the roster it was supposed to describe. The IPSC card quoted a corpus two versions old, and a gate now pins it.
    • The portfolio statement was synced from the corrected ArthurAI copy.
  7. August 16–17, 2026

    Machine-readable corpus, all-edition capability showcase, model SKUs removed

    • A machine-readable capability corpus was published, so an AI crawler reads the same inventory a person does rather than a thinner summary of it.
    • /capabilities now shows every edition, census-exact, instead of describing a subset in prose.
    • Foundation-model SKUs were taken off the landing and Trust surfaces. Naming the model underneath is barred on customer surfaces, and the Trust Center is precisely where a reader is entitled to assume the copy has been checked.
    • Cross-product deployment language was softened to engagements, matching the instrument each claim actually rests on.
  8. July 28–31, 2026

    Lead attribution, consent model, and share cards

    • First-touch attribution was added so a lead can be tied to the campaign that produced it.
    • The consent banner had been covering the contact form on mobile — blocking lead capture outright on the surface the site exists to drive. Fixed, and analytics moved to a CCPA opt-out model.
    • Brand-forward social share cards are now generated per page, so a link shared anywhere unfurls as the page rather than as a generic site card.
    • An editorial rule was overflowing its container on mobile.
  9. July 19–27, 2026

    Workers’ compensation: multi-state grounding and a statute demonstration

    • The WC edition surfaces now show the multi-state grounding and a statute demonstration, so the claim that Justine reasons against jurisdiction-specific law is shown rather than asserted.
    • A promise of a self-serve trial start was removed. The trial is assisted, and saying otherwise sets an expectation the product does not meet at the moment a buyer is most attentive.
  10. July 11, 2026

    Workers’ Comp in market · capability reconciliation

    • Workers’ Comp is now in market for applicant-side California practice. New /wc product page in the burgundy WC identity: the PD Value Engine (WPI → FEC → occupation → age; strict-AMA vs. rebutted), apportionment reasoning (Escobedo · Almaraz-Guzman · Ogilvie), AME/PQME analysis, MMI/P&S, UR → IMR tracking, C&R-vs-Stips settlement with net-to-client, the WCMSA flag (analysis, not a dollar calculator), and bilingual EN/ES client letters + EN/ES voice. Rating and settlement figures are projections the attorney verifies.
    • Under-sold live features surfaced: client SMS with an attorney-alert number-privacy relay, document RAG (“ask over your documents”), and — on WC — bilingual English/Spanish client letters and voice. A new Client communications section on /pi and /wc.
    • Roadmap items are now clearly labeled as coming, never implied live: Filevine integration, a client case-portal, and phone-intake / telephony. The turnkey approved live integrations are Microsoft 365 and Dropbox; DocuSign is available as a bring-your-own connection (firms supply their own integration key to view their envelopes, not send). Native e-signature (ESIGN/UETA) is the default. (Corrected 2026-08-29: this entry originally described the public /intake portal as live. See the entry above.)
    • “Predictive analytics” softened to “outcome statistics” — heuristic firm and venue statistics, not a trained ML model.
    • Edition roster updated across the site: PI and WC (applicant-side, California) in market; MT, MM, IB, EL on the roadmap.
  11. June 19, 2026

    Bias architecture on the record, edition framing

    • New bias-architecture section on the Trust page, in the portfolio voice: you cannot train bias out of a model — you separate the reasoning from the knowledge from the jurisdiction, so the bias becomes something you can read, audit, and govern. We describe population-free reasoning plus runtime steering; we never claim the system is “unbiased.”
    • Six-edition roster framing equalized — PI in market; MT, WC, MM, IB, and EL on the roadmap, each with a status badge so the footer ships no dead links.
    • Shared portfolio canon refreshed: the cross-site “bias is architectural” sentence added, and the portfolio statement updated to “four editions in production — ten across the roadmap.”
    • Added this public changelog and wired it into the footer.
  12. June 18, 2026

    Single-typeface system

    • Typography unified to a single typeface — IBM Plex Sans across display, body, and italic — matching the MindHYVE Brand Bundle canon now shared across all five MindHYVE marketing sites.
  13. June 17, 2026

    Frontier-component abstraction

    • Third-party frontier model names abstracted on customer surfaces; Microsoft and Azure remain named as the platform, and customer copy uses the Eve-Legal F5/reasoner label.
  14. June 6–8, 2026

    Deploy + asset hardening

    • Unknown routes now return a real 404 status instead of a soft-404.
    • Replaced 404ing favicons with PNGs from the JustineAI™ medallion; corrected broken asset and essay links.
    • Contact emails routed to @mindhyve.ai role addresses.
  15. May 30, 2026

    Consent, team, legal, Eve-Genesis

    • Added a GDPR/CCPA consent banner; Microsoft Clarity is gated on consent.
    • Added a /team page and retired /about/leadership; the leadership block now points to the team roster.
    • Stripe activation commerce policies standardized under /legal/*.
    • Eve-Genesis page gained the cognitive-fingerprint and riddle-origin reveal.
  16. May 19, 2026

    Initial production deploy

    • justinegrid.ai launched as an editorial site: /pi (Personal Injury, in market), /technology (Eve-Legal F5/reasoner), /about, /trust, /contact, and /legal.
    • Six-edition roster — PI in market; WC, MM, MT, IB, and EL on the roadmap — plus the ROI calculator and the Founding Firm program.