Changelog

What shipped, when.

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

This log covers the marketing site. Platform release notes — clinical app and API changes — are published per edition at CDS release notes and OM release notes.

  1. September 23, 2026CDS

    The patient engagement page now shows the patient’s screen

    • The entry below this one recorded that patient engagement was the one CDS page with no capture of the product, because the screenshot that would have gone there was of a clinician surface and putting it on a patient page would have mislabelled it. That gap is now closed properly. The page carries the shipped pre-visit interview — the intake progressing through its phases, an answer being written back where the patient can see it, and the notice telling the patient they are talking to AI — on the same synthetic headache record the schematic above it already used, so the illustration and the screenshot now describe one thing instead of two.
    • It is captured from the product running locally rather than from the demonstration environment, which is a deliberate difference from the other captures and is recorded next to them. The patient portal is not a clinician surface, the demonstration environment has no patient account on it to photograph, and pointing a camera at a live clinical system to produce a marketing image is not a habit worth acquiring even when the record in it is invented. Everything in the frame originates in this repository; the capture refuses to run at all if the product asks for anything the script has not been told how to answer, so there is no route by which real information could reach the picture.
    • The first two attempts were thrown away rather than shipped — one carried the development toolbars that only appear when the product is run in development mode, the other an orange DEVELOPMENT stamp across the corner. Neither would have changed a single automated check, because none of them can see inside an image. Both were caught by looking at it.
  2. September 23, 2026CDS

    The CDS pages rebuilt around how the product is bought, and a capability that was about to be lost

    • Measured on the built export before anything changed: /cds/ ran 2,004 words in the body with 125 links, while the home page had already been cut to 656. It was the last long product page on the site, and it was the one a prospect is most likely to land on. It now runs the same shape as the rest of the site — the claim, the proof, the product, who it is for, the price, the ask — at 978 body words with 110 links. The facts, the status verbs and the design language are unchanged; the register is what changed.
    • The block that had to move was the twenty-cell capability grid, and moving it turned up a real defect. /cds/capabilities/ is the page whose whole claim is that it is the complete inventory, and it did not carry everything the grid did. Symptom intelligence and pre-existing condition intelligence were named on the product page and nowhere on the inventory page, and because the inventory page and the machine-readable corpus at /cds/llms.txt render from the same source, they were missing from what an AI crawler reads too. Cutting first would have deleted two shipped capabilities rather than relocating them. The cells are now one array that only the inventory page renders, the two missing capabilities were added to the corpus, and a check reads the built pages and fails if any of them goes missing again.
    • Every specialty page now shows the product rather than describing it. The repository ships ten real captures of the clinical surfaces, taken against a synthetic demonstration record with no patient information, and until today only one of them appeared outside the gallery — the eight occupational-medicine pages had theirs and the clinical-decision-support pages did not. Clinical intelligence, radiology, labs, documentation and prescribing each carry the capture of the screen they describe, labelled as synthetic. Patient engagement did not, on this date: there was no capture of the patient portal, and a mislabelled screenshot on a clinical page is worse than none. The entry above this one closes that gap.
    • Every page in the family now offers a conversation at the top as well as the bottom. Four of them led with a link to more of our own documentation as the loudest next step, and the release-notes page offered nothing at all — which is a strange answer to give someone who has just read to the bottom of it to find out whether the product is still being built.
    • Both edition pages now carry their product film, high up, straight after the claim. The player is a facade: the poster image is served from this site, and nothing at all is requested from the video host until you deliberately press play — which is what lets a video sit on a site that gives visitors a real opt-out. Getting it to work needed a change to the site’s content-security policy, because a video frame was previously refused outright and the only symptom was a blank rectangle on a page that still loaded perfectly. That has now caught this site out three times, so the fix was proved by driving a real browser and pressing the button rather than by reading a header, and the proof is re-run on demand with the policy deliberately broken to confirm it can still tell the difference.
  3. September 23, 2026OM

    The OM pages rebuilt around what a practice buys, and two claims corrected

    • Measured before anything changed: across the nine OM pages there were 183 mentions of AI, five framings of what it takes off a physician’s plate, and zero quantified outcomes or proof points. Revenue cycle — the thing an occupational-medicine practice owner raises first — was twenty-one of three hundred and twenty-three features, sixth of eleven groups, behind medical-legal. The product carries nineteen billing services and the whole California money path. The site was organised the way the product is built rather than the way it is bought.
    • A new revenue-cycle page leads on the entitlement almost nobody collects. California makes late payment self-executing: Labor Code §4603.2 increases a late treatment bill by fifteen per cent with interest at the civil-judgment rate retroactive to receipt, and §4622 adds ten per cent plus seven on a medical-legal bill. Collecting it means knowing the exact date a properly documented bill was received and counting calendar days on every charge. The page then walks the rest of the path — charge and fee calculation, explanation of review read line by line and never auto-posted, second review at ninety days with the five-day mail rule, independent bill review at thirty. The statutory intervals are drawn to scale, so a reader can see that the window to dispute is wider than the window the carrier had to pay and that independent bill review is the narrow one.
    • A new page sets out how the reasoning is prevented from fabricating. Eight guarantees, each naming the function in the product that implements it: an entry that cannot reach a source the model was given is dropped rather than flagged, a second pass runs with the opposite instruction and tries to refute the first, page citations are verified back against the page, legal authorities come from a closed registry so an invented case citation is not a risk the system runs, a date that cannot be read is left blank, sources that disagree are surfaced as a dispute rather than silently reconciled, and a truncated document says so. For a product with no customer references yet, that is the only kind of proof worth offering, and it is checkable rather than asserted.
    • Two claims were wrong and are now corrected. The exam page listed PHQ-9 and GAD-7 among the screening instruments; neither exists anywhere in the product, and the psychiatric template captures the Labor Code §3208.3 compensability elements instead — predominant cause, the violent-act standard, the six-month bar. Separately, two captions written during this rebuild described the product as pre-launch while the rest of the site correctly states it is generally available. This site has had that exact contradiction before, in the machine-readable corpus rather than the page copy; the captions now make a statement about the data in the screenshot and say nothing about availability, which belongs in one place and only one.
    • Three capability families that shipped in September had appeared nowhere: psychiatric examination with the §3208.3 test, cardiopulmonary, and toxic exposure with the statutory exposure routes. The full set of twenty official DWC forms is now listed, including the whole Disability Evaluation Unit package — DEU 102 in particular, which routes a treating physician’s report under 8 CCR §10160.1 and is the more common filing for a treating practice. The causation page now carries the §3212 safety-member presumptions and the trap inside them: §5402(b)(2) and §4663(e) name two different sets of sections, and §3212.15 is in the first and not the second, so a post-traumatic stress case takes the shortened compensability clock and still owes a full apportionment determination.
    • Nine of the eleven OM pages now show the product itself. Twenty de-identified screenshots were already in the repository and unused; the two pages without one are the capability index and the release-notes list, which a screenshot does not improve. Every caption states that the data is synthetic and that no real patient or practice appears anywhere on this site.
  4. September 21, 2026OM

    Every published OM count re-derived from the code, and four capabilities described for the first time

    • All three counted lenses were sixteen days behind the product, and every one of them moved up: user-facing features 299 to 323, customer-facing AI capabilities 78 to 91, background automations 24 to 26. The floors follow to 320+, 88+, 20+ and a 430+ total. Nothing here is a new claim about the product — each figure is the output of a command run against the ChironAI OM repository at one commit, recorded beside the number it produced, and each moved only because the arithmetic crossed a rounding step.
    • No gate on this site could have caught it, and that is the more useful half of the story. The blocking feature-census check re-reads a registry file — a record of a previous derivation. It agreed with the published page because both were written in the same act, and it stayed green for the whole sixteen days. A figure computed from a record of work done will always agree with itself; only re-running the census against the product repository can disagree, and that is what was done here. ChironAI OM’s own drift watch had been reporting the gap for weeks, on the other side of a repository boundary where nobody reading this site would see it.
    • Seventeen counted features belonged to no published group. Ten route families that shipped after September 5 — among them the payer registry, the three secure-link delivery surfaces and clinic credentialing — had no entry in the grouping rules, so each was counted in the headline total and invisible in every group beneath it. All ten are mapped and the thirteen group counts now sum to the census exactly. The CDS count was measured the same way on the same day and did not move, which is what makes the OM movement a fact about the product rather than about the instrument.
    • Four capabilities the product has and the site had never described are now described, as outcomes rather than as screens. A practice-wide impairment register, so a rating waiting on a physician is found by reading one list rather than by opening cases until you hit it. A payer registry that keeps one entry per carrier, claims administrator and employer however many ways staff have spelled it, and sets the Labor Code §4603.2(b) governmental-payer window once on the employer instead of once per case. Four practice registers built on the server and handed over as files — including the §4603.2 pay clock and receivables ageing, so a practice can answer what it is owed at ninety days and give an auditor the register rather than a screenshot. And medical-legal receivables aged by bracket beside the Labor Code §4622 second-review deadlines still open, so a collectable balance is not lost because a ninety-day window closed unwatched.
    • Three surfaces that had no marketing at all are also now covered: reports, requests for authorization and records productions served by an expiring, revocable link rather than emailed. They were invisible to the coverage gate because it excludes dynamic routes as “marketed via their parent listing page”, and these three have no parent. They are the only screens in the product that a person outside the practice — an attorney, a claims examiner, an outside paralegal — ever sees.
    • The machine-readable capability corpus at /llms.txt is generated from the same source as the capabilities page, so all of the above reached it in the same change. Two census dates in its header were hardcoded while a third was a reference; all three are references now. This pass moved all three lenses on one day, which is exactly the case where a hardcoded date would have been refreshed for one lens while silently vouching for two stale ones in the same sentence.
  5. September 14, 2026Platform-wide

    The team page now derives every count it prints, and the headcount is a September figure

    • Three people were removed from the published leadership roster. Every number beside that list is computed from the roster itself rather than typed next to it, so the page moved on its own: 18 of 202 listed publicly, 184 operators across the extended workforce, Newport Beach 12 and Islamabad 5. The two counts that were still hand-maintained — the region comments and the header total — were already stale before this change, reading 22 against a list of 21.
    • The LinkedIn headcount was refreshed to its September pull: 202 people, 6-month growth 32%, year-on-year 153%, two-year 1,022%, median tenure 0.9 years. This is the first time the total has gone DOWN, and it is not the three removals — that arithmetic gives 209, not 202. The two events are unrelated and the source file now says so, so that nobody reading the history later infers a cause that was not there.
    • The date printed beside that figure was wrong, and had been for months. This page rendered “LinkedIn (May 2026)” with the month hardcoded, so a May date sat next to an August number and would have sat next to a September one. It now renders the same field the number comes from, which is what the other portfolio sites already did.
  6. September 12, 2026Platform-wide

    EliAI joins the family — five shipped Agentic Operating Systems, not four

    • The home page, the About pages and the site footer described ChironAI as one of four Agentic AI Operating Systems engineered by MindHYVE.ai. It is one of five. EliAI, the finance Digital Employee, is now named alongside Arthur, Justine, Theo and Chiron everywhere that count appears.
    • The family page gained EliAI’s division of labour on the same terms as its siblings: what the Agentic AI does — read the whole file, retrieve the comparable cohort, value the position, cite every figure to the comparable that produced it, name disagreement between specialists rather than averaging it, re-mark on every event, and decline to price what the cohort will not carry — and what the human does, which is set the mandate, commit the capital, sign the funding, negotiate the lien, accept the mark, and carry fiduciary responsibility for the book.
  7. September 5, 2026Platform-wide

    What counts as one feature is now written down — and both counts are re-derived from it

    • Both capability pages published a user-facing-feature count under the words “counted from the codebase”: 600+ for OM, 220+ for CDS. Neither reproduced. OM’s figure was literally the raw route-decorator census — 608 endpoints published as 608 features — and CDS’s 224 rested on a census document that does not exist. That was never an inflation problem. It was a unit problem: nobody had written down what one user-facing feature IS, so no two counts could agree and no gate could defend either number.
    • The definition now exists, next to the figures it governs, written as data so a script can apply it and two people counting independently land on the same number. One feature is one object-and-action pair a customer can invoke, derived in six mechanical steps: enumerate every route; collapse create/read/update/delete/list into managing one thing; collapse transports, versions, selectors, lifecycle transitions and hyphenated resource families; drop operator-only and machine surfaces; require reachability at BOTH ends — registered on the server and actually invoked by the client; and drop the surfaces the separately-published AI figure already counts. Step one reproduces the old published numbers exactly, which is how the instrument is calibrated rather than merely trusted.
    • Applied, it gives 299 for OM and 126 for CDS. OM moves 600+ to 290+ and its total 700+ to 400+; CDS moves 220+ to 120+ and its total 270+ to 170+. Each total moved by exactly its own features delta and by nothing else — the AI and background-automation figures, corrected in their own passes on August 30 and September 4, were not touched. These are UNIT corrections, not shrinkage: 608 and 299 count different things, and nothing was removed from either product.
    • The reachability rule paid for itself again. Thirty collapsed surfaces in OM and seventeen in CDS are declared and reachable on the server but invoked by nothing in the client — among them OM’s multi-clinic-group administration and its four register exports, and CDS’s four blob endpoints and both DICOM upload starts. That is the same “present but inert” shape the AI pass found in August, and it is the reason a count is not allowed to be a count of declarations.
    • The per-group counts beside each heading were hand-set literals calibrated against the old headline; left alone they would have summed to more than the whole on the same page. Every counted feature is now assigned to exactly one group by rule, and the group counts are census output. A blocking gate re-reads the census against every published figure and fails in either direction — exact, floor, display string, total arithmetic and each group — and its self-test asserts both that seven seeded errors are caught and that a neutered detector misses all seven.
  8. September 4, 2026Platform-wide

    The background-automation floors come down on both editions

    • Both capability pages publish their counts under the words “counted from the codebase” and describe every figure as a floor rather than a rounded-up headline. Two of those floors did not hold, and an unsupportable floor contradicts the method the page states for itself. OM moves 25+ to 20+ and CDS moves 15+ to 9+.
    • OM’s 28 did not reproduce from any lens. The cron count was an unanchored search that matched a decorator written in prose inside a comment, so 19 was really 18; one of the seven “queue modules” imports a single type and is the handler for a queue already counted, so 7 was really 6; the weekly prompt-injection evaluation was counted here as capability while the same product’s own registry excludes it as assurance rather than something a customer asks for; and the durable job-trace substrate is the pipe the queues report through, not one of the things flowing down it. The mechanical count is 24, and a hand pass under the customer-facing rules the AI lens already uses reads 21 — so the floor is set at 20+, low enough to survive the audit that has not been done yet.
    • CDS’s 15 reproduced exactly, which was the tell: nine in-process scheduled jobs plus six worker entrypoints, two counts added without either being checked. All nine scheduled jobs run. None of the six workers do — the workers directory is referenced by no pipeline and the infrastructure declares no Function App at all. One has no producer, two publish to queues that do not exist, two duplicate scheduled jobs on identical schedules, and the last was deliberately rewired to send inline because it could not reach its own secret. Email is still a live capability; it just is not a background automation any more. All fifteen are now itemised and individually classified behind a blocking gate in the product repository, so the figure is defended rather than recalled.
    • Each headline total moved by exactly its own component delta and by nothing else: OM 714 to 710, which still clears its published 700+ so that floor does not move; CDS 279 to 273, which does not clear 275+, so that one moves to 270+. The features and AI figures on both pages were left alone. OM’s 600+ features figure remains the largest claim on either page that no command reproduces; the honest remedy there is a re-census that writes down its unit, not a smaller number, and it is named rather than quietly adjusted.
  9. August 30, 2026OM

    The AI capability floor comes down, 160+ to 75+

    • Itemising OM’s AI layer as data made the number re-derivable. It did not make it customer-facing, and nobody had asked the second question. Of the 165 anchored entries, 87 are the Foundry client, retry and backoff, embedding calls, OCR, prompt-version constants, streaming-versus-blocking transports of one analysis, queue job types and controller surfaces. 78 are things a customer gets. 160+ was an overclaim for two days and the floor is now 75+, with the total moving 800+ to 700+.
    • Every registry entry now carries a customerFacing decision and, where it is excluded, a one-line reason — judged against the same four rules used for CDS, so the two products are comparable. The build refuses an entry that has not been classified, so the next capability cannot be added without someone deciding on the record which side of the line it is on.
    • The reachability rule found ten entries with no production call path, four of which would have been published as capability had the classification been done from their names. The starkest: a section-summariser whose prompt is imported only by a test, in a code path that returns “no summary” unconditionally. Nothing about the product changed this week. The instrument did.
  10. August 29, 2026Platform-wide

    Two published figures corrected against their own instruments

    • The MindHYVE family card said ArthurAI has four editions in market. It has five. The card was written before the fifth shipped and nothing recomputed it, which is the shape of error a hand-maintained figure produces even when every individual page is right.
    • The exact behind OM’s “160+ AI capabilities” moved from 166 to 165 — and it went DOWN because the instrument got stricter, not because the product lost anything. The anchored registry that backs the figure grew two rules: it now rejects anchors that resolve in any file and therefore prove nothing, and it rejects two entries citing the same evidence. Sixty vacuous anchors were re-anchored and one double-counted prompt asset was removed. The published floor is unchanged.
  11. August 28, 2026OM

    The AI capability figure became re-derivable, and the floor moved to 160+

    • OM’s AI capability count had been the one published figure no command could reproduce. It came from a prose itemisation in a full read — named behaviours like ex-parte taint propagation — and reading the code back gave 1,748, or 14, or 54, or 98 depending on the lens, none of them the published number. It is now itemised as data, every entry anchored to a file and a pattern that must exist, behind a blocking gate whose self-test perturbs all of them to prove the matcher is live. The number is the length of that list rather than an assertion about it, and the floor moves 130+ to 160+.
    • The OM release-notes page was missing the August releases. They are published.
  12. August 27, 2026CDS

    The capabilities page showed a human half of what the machine feed carried

    • /cds/capabilities rendered 37 named capabilities from a page-local narrative arranged as a clinician’s day, while /cds/llms.txt rendered 82 from the capability index. Same product, same site, two inventories — and the one a human reads was less than half the one an AI crawler ingests, under a headline claiming 275+. The full inventory now renders above the workflow walkthrough from the same constant the machine feed uses. The narrative stays; it is the better way to meet the product, but it is no longer the only thing a visitor can see.
    • Five capabilities were added that had been shipping unmentioned: automated-decision preferences under California consumer-rights law, in-product help that checks the service before answering, monthly usage visibility, subscription and billing administration, and the clinician credential profile.
    • Two routes were deliberately not marketed, both because reading the code contradicted what the route name suggested. One is compiled out of production entirely and is not a customer capability at all; the other is operator payment configuration rather than a customer surface.
  13. August 27, 2026OM

    The site caught up with the product: four figures and a whole subsystem

    • Forty-five feature commits had landed since the last full read, including the DWC reporting workflow, the QME med-legal surfaces and the entire records-disclosure subsystem. The published floors had stopped being conservative and started being wrong in the understating direction. Features move 570+ to 600+, and the records-disclosure surfaces are published as a thirteenth capability group — a document library, production packets with Bates numbering, an honest manifest of superseded and duplicate pages, a withheld log, and productions that can be voided without destroying the record that they happened.
    • Seventeen shipped capabilities that no page mentioned were added to the index.
    • One page was publishing two different numbers for background automations, because two surfaces each restated the figure instead of reading it. Both now read the same constant.
    • A confidence band claimed a re-derivation date that four of its numbers had not earned. The date now describes only what was actually re-derived on it.
  14. August 27, 2026Platform-wide

    Corpus and disclosure corrections

    • The machine-readable corpus told every model that a generally-available product was pre-launch. An engineering file had become the thing contradicting the site, and a crawler reading it would have propagated the contradiction rather than the correction.
    • The MindHYVE family card carried IPSC figures from the March v3.0 build and published them as current.
    • Foundation-model disclosure scoping was synced from the ArthurAI canon, so the rule about where a model name may appear is stated identically across the family rather than drifting per site.
    • Independent coverage is now cited beside the California Northstate claim rather than the claim standing alone, because a wire release and independent editorial are different kinds of evidence and a reader should be able to tell which one is being offered.
  15. August 22, 2026OM

    Four California clinical pilot evaluations published

    • Four California practices treating workers’-compensation patients have executed Clinical Pilot Evaluation Packages for ChironAI OM — thirty-day, no-fee evaluations in live clinical operation, each with a HIPAA business-associate agreement in force before any patient record is entered. None of it appeared anywhere on the site. They are published on the deployments page as a third category, kept distinct from institutional partnerships because the warrant is different: the institutional partners are named under a press release issued against a signed MOU, while these four are named solely because each gave the written publicity consent their agreement requires.
    • The content-security policy allowlist dropped a host this site does not use.
  16. August 20, 2026Platform-wide

    Edition status reconciled, partnership language throughout

    • OM is stated as generally available since August 12, 2026, and the edition page now carries that date the way the CDS page always has. The status had been published three different ways — the nav said one thing, the site metadata another, and body copy on the gallery a third — because two separate fields held the same fact. The nav badge is now derived from a single source, so the three cannot drift apart again.
    • The footer now reads “In strategic partnership across three regions,” matching the Trust Center and the three-regions page rather than asserting a harder verb than they do. Dated press-release descriptions of specific engagements are unchanged; they name their own instrument.
    • The OM capability corpus no longer describes the product as pre-launch — a line that had become the engineering file contradicting the site. It now states, as the CDS corpus already did, that it describes capability only and that availability is stated on the site.
    • This log gained edition tags and a filter, and the twelve entries below it were backfilled from the shipping history between June and August.
  17. August 19, 2026OM

    OM landing page rebuilt

    • The /om landing page was rebuilt around the five jobs a California workers’-comp visit actually asks of one physician, bringing the edition to the same depth as the CDS lane.
  18. August 19, 2026Platform-wide

    Guardrail hardening

    • A pull request that carries commits from a do-not-merge pull request now fails the guardrails gate rather than quietly inheriting them.
    • The canonical copyright notice is checked by the carry gate.
  19. August 17, 2026Platform-wide

    Instrument-accurate claims, full sitemap, Eve widget

    • Deployment claims on /about/three-regions and /about/founding rewritten to name the instrument behind each engagement rather than flattening MOUs, alliances and collaborations into one verb.
    • Foundation-model names across /insights and /resources rewritten to the customer-facing Eve-Healthcare F5/reasoner label.
    • Sitemap coverage restored — twenty-six pages were missing from it, so they were not being offered to crawlers at all.
    • The Eve Digital Employee widget was moved onto a shared host of its own — one Eve across every MindHYVE site. That host was chat.eve.mindhyve.ai at the time of this entry; it was later renamed api.eve.mindhyve.ai, and the old name has since been retired and no longer resolves.
    • A single NUL byte could switch the guardrails check off. It no longer can.
  20. August 16, 2026OM

    OM capability corpus published

    • The OM capabilities page now publishes the full customer-facing capability count, re-derived from the live codebase rather than restated, and the figures are read from one source so the page, the machine-readable corpus and Eve cannot diverge.
    • Counts are published as floors against exact census figures. A floor survives diligence; an exact number invites a line-by-line audit of every item.
    • A machine-readable corpus is served at /om/llms.txt for AI crawlers, with a groups-at-a-glance list and a canonical page URL.
  21. August 16, 2026CDS

    CDS capability corpus published

    • A machine-readable CDS capability corpus is served at /cds/llms.txt, matching the OM corpus in structure.
    • The corpus states plainly that it describes capability only and makes no deployment, adoption or customer-count claim — availability is stated on the site, not asserted in the corpus.
  22. August 12–14, 2026Platform-wide

    Naming discipline and a mobile layout canary

    • Third-party foundation-model names scoped to approved surfaces only, and language overstating the generality of the system removed from customer copy.
    • The brand gate is now route-aware and understands which surfaces may name Microsoft, Phi and Azure.
    • A mobile layout canary was added and immediately found four live defects on narrow viewports. All four are closed and the ledger is empty.
  23. July 31, 2026Platform-wide

    Consent model, share cards, narrow viewports

    • Analytics moved to a CCPA opt-out consent model.
    • Brand-forward social share cards generated per page.
    • The editorial rule no longer overflows on narrow viewports.
    • A format check was added to PR validation so unformatted edits cannot reach a deploy.
  24. July 29, 2026Platform-wide

    Privacy and legal accuracy

    • Contact-form input is redacted from session replay. Free-text on a healthcare contact form can contain health information, and session replay is not a place for it.
    • Legal pages corrected to describe what we actually do: the terms no longer promise a fourteen-day refund we do not give, and the billing contact is an address that can receive mail.
    • First-touch attribution captured on contact leads.
    • A PR-validation pipeline was added so branch protection could be enabled at all.
  25. July 19, 2026Platform-wide

    Lead handling made durable

    • Transactional email standardised on one provider, removing a second delivery path that could fail independently.
    • Dead-letter capture added, so a lead that fails every delivery attempt is retained rather than lost.
    • Contact and waitlist notifications moved to a branded renderer, and leads land against discrete fields rather than a free-text blob.
  26. July 3–6, 2026Platform-wide

    Agentic operating system repositioning, product gallery

    • The home, how-it-works, CDS and OM lanes were rewritten to lead with the agentic clinical operating system rather than a feature list — Chiron drives the encounter and shows its reasoning; the physician reviews and attests.
    • Residual lab-trending feature claims purged from the site: they described something the product does not do.
    • A public product gallery was added for both editions, on synthetic demonstration data and labelled as such.
  27. July 2, 2026Platform-wide

    Trust Center build-out

    • Trust Center gained SOC 2 Trust Services Criteria controls, a subprocessor list, incident response and availability sections.
    • The frontier-model data-handling statement was reworded to match the controls actually in place, removing an unbacked de-identification claim.
    • A secret-scanning gate was added to the production deploy pipeline.
  28. June 21–27, 2026Platform-wide

    Agentic showcases, brand canon corrections

    • Reasoning showcases published for both editions, so the agentic behaviour is shown rather than described.
    • Canonical Digital Employee spelling corrected across brand canon.
    • Substrate hero treatment on the home page.
    • Lead-API key comparison made timing-safe, and reliance on a long-lived directory secret reduced.
  29. June 19, 2026Platform-wide

    Bias architecture on the record, Africa region, edition parity

    • Trust Center now carries a “how we handle bias” section 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.”
    • Africa region corrected: the region now hubs from Nairobi. Kadisco General Hospital — our customer — remains in Addis Ababa, Ethiopia, and its deployment content is unchanged.
    • CDS and OM edition framing equalized across the site so both editions read with the same depth and structure.
    • 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.
  30. June 18, 2026Platform-wide

    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.
  31. June 17, 2026Platform-wide

    Frontier-component abstraction + format gate

    • Third-party frontier model names abstracted on customer surfaces; Microsoft, Phi, and Azure remain named, and customer copy uses the Eve-Healthcare F5/reasoner label.
    • Added a Prettier format:check stage to CI so unformatted edits can no longer reach a deploy.
  32. June 6–8, 2026Platform-wide

    Deploy hardening

    • Unknown routes now return a real 404 (removed the Static Web App navigation fallback that was producing soft-404s).
    • Single-hop trailing-slash redirects; hardened a handful of broken links and asset hrefs.
    • Contact routes for legal, privacy, and security moved to @mindhyve.ai role addresses.
  33. May 30, 2026Platform-wide

    Team, legal, Eve-Genesis

    • Added a /team page and retired /about/leadership; team data synced.
    • Standardized commerce policy pages under /legal/* (terms, privacy, refunds, cancellation, acceptable use, promotions).
    • Eve-Genesis page gained the cognitive-fingerprint and riddle-origin reveal.
  34. May 20, 2026Platform-wide

    Share-card coverage fix

    • An audit found 43 of 44 pages were shipping imageless share cards — only the apex emitted an og:image. A centralized social-metadata helper restored 100% Open Graph / Twitter card coverage across every page.
  35. May 13, 2026Platform-wide

    Initial production deploy

    • chirongrid.ai launched on Azure Static Web Apps behind Azure Front Door.
    • Two editions live: CDS — Clinical Decision Support (GA) — and OM — Occupational Medicine — built first for California workers’-compensation physicians.
    • Information architecture: Editions dropdown, six-column footer, Trust Center, Technology (Eve-Healthcare F5/reasoner), and the For-physicians lane.
Next stepOne conversation

See it run an encounter that looks like yours.

Tell us the workflow you want to see and who needs to be in the room. A member of our clinical and engineering team replies within two business days to arrange it — a real person, not a sequence.

Talk to our team →

What happens after you send it

  1. 01. The form reaches a monitored inbox. No auto-sequence.
  2. 02. We reply within two business days to book a time that suits you.
  3. 03. Forty-five minutes: your workflow, the reasoning trace, procurement questions.

Prefer to talk now? +1 (949) 200-8668, or email hello@mindhyve.ai.