Changelog

What shipped, when.

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

This log covers the site. For the platform itself — what shipped across the editions and ArthurAI™ Studio — see the platform release notes. Deeper detail than either page carries comes through your engagement.

  1. September 18, 2026

    The Studio pages now carry the whole product, and a 69-question FAQ

    • A census of the ArthurAI-Studio codebase against all 102 live pages found the site understating the product in ten places and overstating it in three. Both directions are fixed together, because a complete capability list that still sells something undeliverable is worse than an incomplete one — it closes the deal support has to unwind. Ten capabilities ship today, on the FREE plan, and appeared only in release notes: four simulated learners who take a course before anyone pays for it, a pedagogical lint with a reviewing panel, a freshness pass that notices when the world moved under a claim, one-tap translation into eight named languages, seatless reviewer links, version history, two-factor authentication, self-service data export and erasure, and a visible usage meter.
    • Three claims were withdrawn. The pricing matrix dropped its SSO row on September 11 and the FAQ sentence selling SSO on the Business plan survived on the same page — a buyer reads the prose. That same answer also gave Team review-and-approval, which every plan including Free already has, and verifiable credentials and LTI 1.3 grade passback, which only Compliance Autopilot can reach. And a “four compounding advantages” lede introduced a list that has always rendered two cards: fork-and-remix and the curriculum marketplace are internal staff surfaces, with no marketplace endpoint under the customer API at all.
    • The capability figure now derives instead of being retyped. /studio published “35+” as a literal while the capability registry recorded 48 for Studio alone and the estate floor had already moved to 40+ — three numbers for one product, none of them reading the others. A gate added days earlier pinned the figure to the CARDS on the capabilities page, which is circular: the number published about the product was computed from how much marketing copy had been written about it, so adding a card grew the claim. It now reads the code census, and the gate keeps every check it made plus one it did not: a floor above the thing it floors is not a floor.
    • A comprehensive FAQ answers the sixty-nine questions that decide a purchase — what happens at the monthly token budget, whether an SME can review without a seat, which export format sits on which plan, what the free plan really includes, who at MindHYVE can see your courses. Every answer was written from the entitlement catalogue and the route gates rather than from the pricing page, because a capability named on the wrong plan is not a rounding error but a refusal the customer meets after deciding to trust us. Five things are deliberately not claimed, including any Studio-scoped SOC 2, FERPA or WCAG certification: the Trust Center does not scope Studio, and the FAQ says so rather than letting a platform attestation read as a Studio one.
    • Two gates now hold the line. Withdrawn claims cannot return to the Studio pages, because all three of the corrections above were applied in one place and left standing in another. And no two pages may emit the same FAQ question: twelve pages here publish FAQPage structured data, duplicated Q&A does not earn two rich results but risks earning none, and writing sixty-nine questions produced exactly one collision — invisible to every other check, found only by diffing the sets by hand.
  2. September 11, 2026

    The Studio pages now describe the Studio you can buy

    • An audit read every capability claim on /studio and /studio/capabilities against the code that would have to implement it. The structural finding was that Studio is two products in one container image — a customer SaaS on Postgres and a staff authoring console on Cosmos — and roughly twenty of the forty-seven capability cards described the second one. The customer deployment carries no Cosmos endpoint, so every route behind those cards returns a 500 there. This is the same error as the “twenty-plus tools” claim corrected yesterday, generalised: the capabilities page’s own header named its source files, and every one of them was staff-only.
    • Single sign-on has been removed from the pricing matrix rather than softened. There is no SSO feature flag, no SAML anywhere in the estate, and nothing to deliver against — the word survived as a tier blurb string, which the public pricing API was also serving. Four more rows were wrong in both directions: verifiable credentials and LTI 1.3 grade passback are Enterprise-only and were sold on Team; hosted learner links and approval governance work on every plan and were sold as paid upgrades. Two limits that actually govern a plan — the monthly authoring budget and the workspace count — were disclosed nowhere and now are.
    • Compliance Autopilot had four claims that did not hold, and it is the section read by regulated buyers. The report is not “sealed”: the PDF generator itself refuses that word, in a comment explaining that an unkeyed digest nobody stores gives an auditor no reference value to compare against. Studio does not email training links to employees — an operator copies each one. And recertification cannot happen before a credential lapses, so the promise that people “retake the training from the link” before expiry described something the product refuses. The dashboard, the report, the credential and its public verification link are all real; the sentences around them were not.
    • Several claims were fixed in the product instead of the copy, which is the better outcome where it is available. “Exactly as authored” is one: the setting was real, stored, and silently dropped by the export, so every course that chose verbatim arrived at the edition adapted. And “How Arthur authored this” was rendering a provenance field no customer-path writer ever set — the panel was built, shipped, screenshotted and written about while reading an empty key.
    • The published capability figure falls from 50+ to 35+. That is the honest direction: the number was derived from the cards on the page, and the cards were describing a product the reader could not sign into. A figure that only ever moves upward is not a measurement.
    • Worth recording how this happened, because the mechanism is the reusable part. Three of the false claims trace to code comments asserting behaviour no code implements — a comment in the learning-paths service describing prerequisite logic that was never written, and a test docblock describing a field nothing produced. The marketing copy appears to have been written from those comments rather than from the code. A green test over an impossible fixture kept one of them believed for months.
  3. September 11, 2026

    Studio’s pages catch up with Studio — and two claims that were not true come down

    • The Free plan both had and did not have export, about four hundred pixels apart on the same screen: the plan line sold “export a portable copy”, and the feature matrix directly below it priced export out of Free. The matrix was right — the product gates it — so the plan line is what changed. Free authors complete courses and publishes them to a shareable hosted link, which is what it actually does.
    • The capabilities page credited Arthur with “twenty-plus permission-scoped tools”. Measured in the product rather than remembered: the Arthur a customer talks to has five — web search, reading your course library, opening an outline, searching your own history, and the help library. The twenty-plus registry is the internal staff console, which no plan grants and which the customer deployment cannot even authenticate against. A number that describes a surface the buyer cannot reach is not a small overstatement; it is a different product.
    • A new section, The glass box, covers four things that shipped over the last month and had no page at all. Arthur streams his actual reasoning while he works — and it is now kept, so a course authored last month can still tell you why it is sequenced the way it is, and a finished answer still carries the thinking behind it. He researches the live web and cites the sources the search returned rather than URLs recalled from memory. Every tool call is visible as it happens. And a turn now runs as a job on the server, so closing the tab or switching apps on a phone no longer loses the answer.
    • Both gates on that work are stated in the copy rather than left out. Web search in chat is on paid plans. Live research while authoring is opt-in, per course, from Creator up — off by default, because a course silently grounded in whatever a search returned is a course whose author cannot say where its facts came from.
    • Studio’s anchored AI/agentic count goes from 21 to 24, taking the exact figure across both halves from 41 to 44, and the published floor moves 35+ to 40+. The floor has deliberately not tracked every increment — it stayed put at 39 and at 41 — but 35+ against 44 understates by nine, which is the point at which a conservative number starts inviting the question of what else it is not tracking. The customer-facing total follows the same arithmetic to 290+.
    • Three entries came OUT of that registry in the same pass, because the code they anchored to had been deleted. That is the registry’s own rule and it is the half worth noticing: a published floor that only ever absorbs new work, and never gives back what stopped being true, is not a floor.
    • Smaller corrections, all found by reading the pages against the product: the Dawn Directive sentence rendered “a 18-course” — fixed with a helper rather than a hard-coded “an”, since the number comes from a shared constant; the improvement loop now states the precondition the capabilities page already did, that it reads learner signal from an ArthurAI™ edition; Open Badges 3.0 is hedged the same way in all four places it appears rather than three; the Creator plan no longer says “small teams” beside a one-seat entitlement; and the capabilities page stopped enumerating five pipeline stages in prose that disagreed with the five stages listed immediately below it.
  4. September 10, 2026

    Studio’s workflow gains a step: Arthur’s reviewers now argue about the course before you see it

    • The How-it-works section on /studio described four steps — brief, build, approve, deliver. There is now a fifth between build and approve, because the product gained one. Before a course is handed over, a panel of specialist reviewers audits it for alignment, readability and structure, and then disagrees: reviewers read each other’s findings, dispute the severities they think are over- or under-called, and a lead reviewer rules on whatever the panel cannot settle. The whole deliberation is visible while it happens.
    • This is a change in kind, not in volume, which is why it earns a step rather than a bullet. Course health already existed, but it was two model calls fired in parallel isolation whose lists were concatenated — nothing reconciled them, and a severity was whatever the single reviewer that noticed it happened to say. It also only ran when someone thought to press a button, so a newly generated course was handed over with its problems intact and nothing anywhere saying so.
    • The site copy claims Arthur fixes what it can of its own output before handing over, so it is worth being precise about what “fixes” means: the score reported afterwards is measured, not assumed. Subtracting the findings that were revised would report a better course whether or not the revision actually worked, so the revised lessons are re-read and only findings confirmed resolved are dropped. A revision that did not land keeps its finding and the score does not move.
    • Studio’s AI/agentic count goes from 19 to 21 with the two new entries — reviewer cross-examination, and review at creation time with verification — taking the exact customer-facing total across both halves from 39 to 41. The published floor stays 35+. It was true at 39 and it is true at 41; a headline that tracks every increment is a headline nobody can check, and this one has already moved three times in a day.
    • One thing the count did NOT gain: course health was rewritten into the reviewer panel rather than registered alongside it. It is the same behaviour, deliberating — registering both would have published one capability twice.
  5. September 10, 2026

    The portfolio band names five, the parent company says what it does, and a shared course link finally previews

    • The portfolio now lists five sibling Agentic Operating Systems rather than four, on the home page, the footer and the MindHYVE family page. The grid was re-laid at the same time: three columns sat four cards evenly and left a fifth alone on its own row, so it is two-up then four-up. That column count is the only place in the component that knows how many products there are, which is the thing to re-derive when the portfolio grows again.
    • The portfolio band also carries a statement of what the parent company does, in the same words chirongrid.ai uses. Of the four product sites this was the only one that said nothing about MindHYVE.ai at all. It sits in the band rather than the hero deliberately — the hero belongs to ArthurAI, and a product page that opens with its parent buries the product.
    • Sharing a link to a course now previews as a course. The share card these links produce carried the ArthurAI logo and nothing else — the same image for every course anyone had ever shared — while the title and description beside it were already specific to the course. Only the picture said nothing, and the picture is the half people look at. It is a proper card now, generated from the same template as every other card on this site.
  6. September 4, 2026

    The AI capability figure goes back up, because the second half of the count was missing

    • Earlier today the AI/agentic figure moved from 50+ to 25+ when the estimate behind it was replaced by an itemised, anchored count. That entry is below, and it named in writing the deployed surface still uncounted — the workers that generate course curricula, lesson content and practice questions, the narration path, and more. Those have now been counted. The figure is 35+, against an exact 39: 20 in the learning platform and 19 in ArthurAI Studio.
    • The rule did not change to suit the number, either time. An entry counts only if a buyer would name it, its loss would be felt as a missing behaviour rather than as slower software, it is not a selector or a variant beneath something already counted, it is reachable in production, and it actually calls a model. What changed is that the learning platform’s registry had been scoped to one of its two deployed units and said so. Covering both took it from 33 anchored entries to 123 — of which 20 pass all five rules.
    • Counting properly also removed things. A capability that generates a learning profile from performance data is registered on a queue nothing sends to, and where its model call belongs there is a placeholder string. The entire spoken-question path reads a setting the service never defines, so transcription fails everywhere, not just in production. Three of the ten AI controls shown to administrators are read by no code at all. And “ask about an uploaded file” was overstated: images are genuinely read, documents are not — the tutor receives only the filename. That one had been published; it is corrected rather than quietly kept.
    • Eleven further behaviours are real, shipping and useful, and are not AI. Arthur’s course suggestions say so in their own source comment — a heuristic, not a model. At-risk detection is two thresholds; practice grading is a comparison; read-along timing is arithmetic over character counts. All of them belong in the features figure, and none of them is published as AI.
    • The customer-facing total follows the arithmetic from 275+ to 285+. It has no independent life — it is features plus AI plus automations, and only the AI term moved. The features half still has not been re-derived and is still the August 16 census. The AI half has now been wrong in both directions before settling, which is the argument for giving the features half the same treatment rather than trusting it.
    • Both halves are now behind build gates that refuse an unclassified entry, an exclusion with no stated reason, an anchor so generic it proves nothing, and two entries resting on the same evidence — each of which was deliberately broken to confirm the gate fails. Re-derive with `node scripts/checks/ai-capabilities-anchored.mjs` in ArthurAI-Backend and in ArthurAI-Studio.
  7. September 4, 2026

    The AI capability figure comes down, because it is now counted rather than estimated

    • The published AI/agentic figure moves from 50+ to 25+. Nothing about the product changed — the instrument did. On August 29 the figure rested on 32 anchored capabilities in the learning platform backend plus “26 services that touch an AI client” in ArthurAI Studio, and that note said in writing that the number must not move until Studio’s half was itemised the same way. Studio’s half is now itemised: 79 entries, every one anchored to a named symbol in deployed code, behind a build gate that fails if an anchor stops resolving. Counting services that touch an AI client turned out not to be counting capability.
    • The rule applied is the one ChironAI already uses, so the two products can be compared: an entry counts only if a buyer would name it, its loss would be felt as a missing behaviour rather than as slower software, it is not a selector or a variant beneath something already counted, it is reachable in production — and, added for Arthur, it actually calls a model. That gives 19 customer-facing capabilities in Studio and 7 in the learning platform backend: 26, published as a floor at 25+.
    • The fifth rule earned its place. Five Studio behaviours that read as agentic from their names make no model call at all. “Arthur Noticed”, the proactive review notices, is a pure rule function — its own source comment calls it a pure-code nudge — and the weekly digest is a database rollup, credential grading is deterministic scoring, and the improvement ranking is arithmetic over an imported signal. All five ship, all five are useful, and every one of them would have been published as AI on the strength of its name.
    • Two claims about Studio were describing a surface a customer cannot reach. The “22-tool authoring agent” is the internal staff console; the Arthur a Studio customer talks to carries four tools, reads their own courses before advising, and answers how-to questions from the product’s help library. That copy now describes what a customer gets. Twenty-five further Studio capabilities are staff-only and are excluded for the same reason — the customer deployment is not configured to serve them, and no plan grants access to the console that is.
    • The customer-facing total follows the arithmetic down from 300+ to 275+, since this file defines it as features plus AI plus automations. The features half has NOT been re-derived and is still the August 16 census — the same census whose AI half proved to be counting a repository that has never been deployed. It should get the same treatment before anyone quotes it in diligence, and the staleness gate has been pointed at the older of the two dates so it keeps saying so.
    • Re-derive either half with a command rather than a memory: `node scripts/checks/ai-capabilities-anchored.mjs` in ArthurAI-Studio and in ArthurAI-Backend. An accurate smaller number beats an indefensible larger one.
  8. August 29, 2026

    The AI capability figure is re-derivable, and the Dawn Directive count reads from one source

    • arthurgrid.ai publishes “50+ AI capabilities”. Thirty-three of them were counted from ArthurAI-V2-UnifiedSystem, a tree that is not deployed anywhere — verified three ways: no running container uses any image it builds, its image names do not exist in the registry, and its build project has no pipeline definitions at all. The figure was true and its evidence was not, so the evidence was replaced rather than the number lowered: the deployed backend now has 32 AI capabilities itemised and machine-verified behind a blocking gate, and ArthurAI Studio’s deployed backend carries 26 more. The right answer to weak evidence for a true claim is better evidence, not a quieter claim.
    • The Dawn Directive™ course and hour counts are now read from a single constant rather than restated in Studio page copy, so the programme figures cannot disagree between pages.
  9. August 27, 2026

    Foundation-model names removed from the home page, coverage cited beside the claims

    • Foundation-model names had leaked onto the ArthurAI home page. Naming the model underneath is barred on customer surfaces — the customer-facing label is the Eve engine, not the vendor model — and the home page is the most-read surface on the site. Removed, and the branding gate that should have caught it was widened so the same class of leak fails the build rather than reaching production.
    • Independent coverage is now cited beside the partnership claims rather than the claims standing alone. A wire release and independent editorial are different kinds of evidence, and a reader should be able to tell which one is being offered.
    • The MindHYVE family card carried IPSC figures from the March v3.0 build and published them as current.
  10. August 20, 2026

    Edition status reconciled, portfolio statement corrected

    • Edition status is now stated one way. Five editions carried “In market” while FLE carried “Generally available”, and /capabilities renders that field — so the same page showed six generally-available editions labelled two different ways. All six are generally available: CLE, SLE, VLE, ULE, FLE and PLE, alongside ArthurAI Studio. GLE, HLE and ALE remain on the roadmap. This completes the correction started on August 12, which noted “five, of six” still to go.
    • The shared portfolio statement said “four editions in production — ten across the roadmap”. That conflated editions with Digital Employees — the same sentence then says “four … Operating Systems” — and understated the portfolio, since ArthurAI alone has seven generally available. It now reads twelve generally available across four Operating Systems, with the arithmetic recorded in the canon file so it can be checked rather than trusted. Authored here and copied verbatim to the other three grid sites.
    • The /about/founding timeline’s “Today” entry was out of date on three of four products — it described four ArthurAI editions in market and JustineAI in field trial. It now matches the current status of each.
    • Two build gates added: one fails when page copy restates an edition status instead of reading it from the source of truth, the other warns when the published capability census or this changelog goes stale. The second one is why this backfill exists.
    • Added the canon:hash script. The canon drift check had always failed with “regenerate the manifest with pnpm run canon:hash” — and that script existed in none of the four grid repositories, so the one instruction the failure gave could not be followed.
  11. August 17–19, 2026

    Claim accuracy pass, capability census on the page, guardrail hardening

    • Every named-partner verb and published figure was checked against its source, and deploy claims on the home page and /about/deployments were rewritten to name the instrument behind each engagement rather than flattening different agreements into one word.
    • The /capabilities page now carries the capability census counts, read from a single source so the page, the machine-readable corpus and Eve cannot diverge.
    • A self-description as an “assistant” was removed from the ULE faculty-of-record copy. Arthur is a Digital Employee that reasons and drafts for an educator to review.
    • The Eve Digital Employee widget now loads from chat.eve.mindhyve.ai — one Eve across every MindHYVE site.
    • Guardrails: a pull request that carries commits from a do-not-merge pull request now fails; the canonical copyright notice is checked; and a single NUL byte could previously switch the guardrails check off, which it no longer can.
  12. August 16, 2026

    A WAF in front of the site, machine-readable capability corpus

    • A web application firewall now sits in front of the arthurgrid.ai Front Door, and its desired-state definition was reconciled with the live prevention posture so the two cannot silently diverge.
    • /llms.txt gained a Capabilities section carrying the exact census counts, so an AI crawler reads the same numbers a buyer sees.
  13. August 11–12, 2026

    FLE generally available, processing-entity rewrite, route-aware brand gate

    • ArthurAI FLE published as generally available rather than “in market” — the first of the six.
    • The /trust pages were rewritten to describe the processing entity accurately rather than as a SKU, and a press correction was landed alongside it.
    • The brand gate became route-aware, so it understands which surfaces may name a specific third-party platform component and which may not — customer copy uses the Eve-Education™ F5/reasoner label.
  14. July 31, 2026

    Share cards, contact-API storage posture

    • Brand-forward social share cards generated across the whole site.
    • The build now fails if the contact API’s storage is left publicly reachable — a posture guard rather than a one-time fix.
  15. June 19, 2026

    Bias architecture on the record, CIAI relationship corrected, one-typeface system

    • Responsible AI page expanded with a “how we handle bias” section grounded in the portfolio canon: 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.”
    • CIAI relationship corrected sitewide: CIAI is an independent customer of ArthurAI™ VLE, not a MindHYVE subsidiary, sister company, or family member. The Dawn Directive page, the home page, and the “Dawn Directive as vertical infrastructure” insight were all updated to remove the subsidiary/sister/family framing.
    • 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 sites.
    • 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.”
  16. May 17, 2026

    Trust Center backbone, How-it-works depth, audience hubs, edition deeps

    • Trust Center expanded with 6 new pages: Security FAQ (30-question FAQPage with Schema.org markup), Subprocessors (counsel-grade list of every third party that touches institutional data), Responsible AI (pre-LLM and post-LLM guardrails, logging discipline, human-in-the-loop posture), Accessibility Conformance (WCAG 2.1 AA posture, VPAT availability), Data Processing Agreement summary, Incident Response (severity classification, notification cadence, runbook).
    • How-it-works trio added: the Learning Cognitive Profile deep dive with academic lineage (Felder-Silverman, Kolb, Sweller, Kirschner) and explicit calibration against the strong-form learning-styles hypothesis; the 6-step lesson flow anatomy; the AI tutor architecture; assessment architecture.
    • Edition deeps shipped: SLE safety-and-disclosure with parent communication kit and state-AI-disclosure-law landscape; ULE faculty-of-record posture with accreditation alignment; CLE SOC 2 readiness narrative.
    • New audience hubs: For Parents (plain-language coverage of what the platform does and what rights parents have), For Procurement (single landing page aggregating DPA, security FAQ, subprocessors, accessibility conformance, audit rights), For Press (boilerplate, fact sheet, founder bio, brand assets, story angles).
    • Brand and positioning: Why ArthurAI (the ten posture statements that define what the platform refuses to be), Student Data Rights (plain-language FERPA / COPPA / state-law / GDPR coverage), K-12 Funding Sources (Title I / III / IV-A / IDEA / ESSER / state / philanthropic landscape).
    • Footer cleaned: 9 dead-link footer entries from the original IA replaced with live pages. The footer now links to every published page.
  17. May 16, 2026

    Capability pages and audience-lane narratives

    • 4 capability pages shipped, one per shipping edition: SLE, ULE, VLE, CLE, each with 8 capability cards calibrated to the audience and grounded in the V1.7 product spec.
    • 4 audience-lane narratives: Your day with SLE (K-12 teacher), Your semester with ULE (university faculty), Your program with VLE (vocational instructor), Your quarter with CLE (corporate L&D leader).
    • First-month onboarding playbook for educators new to the platform.
    • Per-edition partner evidence cards added to /sle, /ule, /vle, /cle drawing from the verified institutional-partner roster.
  18. May 17, 2026 (earlier)

    Brand system, font swap, icon refinement

    • Body font swapped to IBM Plex Sans (replacing Inter) for editorial character; restricted to 400 and 500 weights to preserve the editorial register.
    • Icon mark tight-cropped: created icon-mark.svg with viewBox bounded to the medallion (preserving the original icon.svg with safe-zone padding for lockup compositions). Icon at 56px in the nav reads at ~2.5x effective vs the original full-padded 40px.
    • Wordmark color system aligned to the v6.1 SVG colors: charcoal ‘Arthur’, gold ‘AI’, taxonomic-grey ‘TM’ superscript.
  19. May 17, 2026 (earlier still)

    Initial production deploy

    • Site provisioned on Azure: rg-arthurai-marketing-prod / swa-arthurai-marketing-prod (Standard tier) / appi-arthurai-marketing-prod (Application Insights) / law-arthurai-marketing-prod (Log Analytics).
    • Azure Front Door Premium fronting the Static Web App: afd-arthurai-marketing-prod with WAF, custom domains arthurgrid.ai (apex) and www.arthurgrid.ai (www) created and validated.
    • AFD endpoint live at ep-arthurai-marketing-prod-b7eqdwfwegg7fub2.z02.azurefd.net. DNS cutover from Wix to AFD pending founder go/no-go.
    • Initial 27-page IA shipped: home, 5 edition pages (sle / ule / vle / cle / ple), 4 trust pages (compliance, security, data-handling, disclosures), 4 technology pages, 3 about pages, contact, pricing, how-it-works, for-educators, resources, 3 legal pages.