بِسْمِ ٱللَّٰهِ ٱلرَّحْمَٰنِ ٱلرَّحِيمِ
RELEASE NOTES
Theo is generally available at chat.theogrid.ai. This is the itemised record of what changed on that surface and on the site that publishes it — one entry per change, newest first, each dated the day it went out. For the dated narrative of releases and corpus versions, read thechangelog.
This is the itemised ship log for the Theo surface and the site that publishes it, reconstructed from merged engineering history. For the dated narrative — corpus releases, launches, audit responses — read the changelog. For corpus-release internals, read ipsc.theogrid.ai/changelog.
Security, access-control and personal-data changes are not itemised on this page. On a theology surface what a visitor asks is special-category data, and a public list of which control changed and when is an attack map rather than a release note. Our security posture and disclosure process are published on the security page, and the privacy notice states what we collect, why, and how to withdraw consent.
The colophon lists the Agentic Operating Systems MindHYVE builds, and EliAI is now among them. The list carries no availability verb for any product on it, and none was added for this one — appearing in the colophon says a product exists, not that you can buy it. Where a product’s status differs from that, the product’s own page is the place it is stated.
The Majlis page described what Majlis does without ever stating that you cannot have it yet. It now carries an explicit roadmap section: Majlis is on the roadmap, not in market, being built on the corpus and the evidence discipline described above it, with the status written out — waitlist, and nothing on the page available to buy today. The disclosure is in all five languages rather than English with four translations pending, because a reader arriving in Urdu or Bahasa is exactly the reader least able to infer availability from tone.
The company line under the products described the size and reach of MindHYVE without saying what the products are for. It now opens by stating that the Agentic AI is purpose-built for regulated industries — the through-line that connects a legal product, a clinical one and an education one, and the reason the evidence and disclosure discipline on this site looks the way it does. Updated in all five languages in the same change, so no locale carries the older framing.
Every page now carries its title in all five dictionaries. Until this change a visitor reading the site in Arabic, Urdu, Bahasa Indonesia or Bahasa Melayu saw a translated page under an English browser tab, and search results and shared links showed the English title for a page whose content was not. Titles are the part of a page most often seen without the page — in a tab strip, a bookmark, a search listing, a pasted link — so they were the least appropriate thing to leave untranslated.
The band naming the sibling Agentic Operating Systems had been built in this repository and imported by no page, so its markup appeared in no build — this was the one grid site that never showed the shared portfolio statement. It now renders on /about, listing ArthurAI, ChironAI, JustineAI and EliAI with their reasoning-mode signatures and links to their grids. EliAI, the Agentic Finance Operating System, joined the portfolio canon the same week and is the fifth product named across every site in the family.
Sanad verifies an attributed claim against the corpus: it locates the claim, walks the chain of transmission, reports the classical grade, and shows the evidence before the verdict. The new page is the canonical statement of what it does, written as a specification rather than an announcement — every capability carries an explicit status, nothing is written in a tense that implies availability, and the only call to action is the waitlist. It does not issue rulings, and it carries no launch date because none has been set.
The page had told visitors that unresolvable cases were routed to a human reviewer and that disputes went to human review. There is no such lane, so those sentences were withdrawn rather than left to be discovered. The sixth verdict is now “Unresolved”: an honest terminus that states Sanad cannot settle the question, shows what it did find and what would settle it, and stops — never a pending state. A contested verdict is published alongside its case so the next reader sees that it is contested; it is not arbitrated.
The footer on every page listed three products while the header listed four, one scroll apart. Sanad is now in both, and has its own social share card. That card matters more than most: links to Sanad travel by being pasted into group chats, so for a large part of its audience the card is the first and sometimes the only thing they see. It states the product status on its face.
The narrator figures on the home page, /provenance, /about, /capabilities, /majlis and the site-wide foundation strip were larger than the index they counted. Re-measured against the serving index and corrected in all five languages: 26,923 narrators indexed, 21,617 carrying a reliability grade, 16,278 with a structured assessment. The label changed with the number — “assessed” became “graded”, because the published figure is the count carrying a grade and the assessed population is a narrower, different figure.
The term appeared on ten pages and on neither of the two a reader arriving from social actually lands on. The home page and /about now both state what Theo produces and what he declines to produce, in the wording already approved on /theo/ and already present in all five languages. No new religious-boundary copy was authored, and nothing was sent for translation.
One capability card promised scholarly attribution for every ruling, on a page that elsewhere states Theo never issues one. The four translations rendered that clause more starkly than the English. Corrected in all five languages: Theo returns an Evidence-Based Opinion with its chain of transmission, and the decision stays with the reader.
The English fallbacks and the translation bundles disagreed about the same corpus figure, so English pages published one value and localised pages another. Corrected to the measured values — 123 distinct works, 480,334 graded records — across the FAQ, the structured data that renders on every page, the press kit, /about and the capability copy, and on this page and /changelog, which were the last two carrying the withdrawn pair. A new gate scans the built HTML of all five language builds and fails if a withdrawn figure is rendered anywhere, with no exemption for the log pages.
theogrid.ai had no robots.txt, and the conventional sitemap path returned the 404 page, so a crawler had to guess where the index lived. There is now a robots.txt declaring the index and a redirect from the conventional path. No Disallow lines: everything published here is meant to be read.
The tiles under the v4.0.0 heading are release deltas, not current totals, and the narrator unit changed after that release — so the document total announced there is not directly comparable with the total published above it. The page now says both things instead of leaving a reader to reconcile two numbers that were never the same measurement.
The team page showed growth only over a year. It now shows the six-month figure alongside it, so the more recent trend is readable directly rather than inferred from the longer one.
How this site attributes the reasoning substrate beneath Theo was corrected, and the shared portfolio canon was resynchronised from source, so all four MindHYVE grid sites now describe the portfolio identically rather than in four slowly diverging versions. The reasoning engine is named as Eve-Theology™ F5/reasoner, which is the only engine name any customer surface carries.
The annual price shown for each plan is now the amount the payment page actually charges. A price a reader has to discover at checkout is not a published price.
A new capabilities page publishes the whole customer-facing inventory as a census taken from the code rather than a marketing list: 290+ customer-facing capabilities across twelve groups, with each group's count and a named selection of what is inside it. Every figure is a floor, internal operations tooling is excluded from all of them, and the census date is printed beside the numbers so the figures cannot age unnoticed.
The home-page product label described Theo in generic software terms. Theo is a Digital Employee — a scholar who reasons over primary sources and returns an Evidence-Based Opinion with its chain of transmission, never an edict — and the label now says so. Naming is not decoration here: it is the difference between what Theo does and what a reader might otherwise assume.
Arabic, Urdu, Indonesian and Malay pages declared the English page as their canonical and social URL, so sharing or indexing one resolved to English. Each locale now points at itself, and a link shared in Arabic opens in Arabic.
Eve — MindHYVE's own Digital Employee, specialised here for theology questions about the platform — is loaded from the shared service rather than a site-local copy, so her answers about TheoAI stay in step with the rest of the portfolio.
The same census is now published in machine-readable form at /llms.txt, with exact counts, the positioning language, and the twelve-group map. AI search and Eve quote the same figures a reader sees, instead of inferring different ones from prose.
Member, citation and page counts were published from three places at once — a live counter, a frozen copy baked into the hero and all five locale bundles, and a third copy in the press kit under a heading that claimed it was live. The frozen copies had drifted 13% below the live value, visibly, on one screen. Every instance now reads one build-time source, so the numbers cannot diverge again.
Copy across the site implied the whole corpus was computationally graded. 480,334 of 617,912 cataloged narrations carry a grade — roughly 77.7% — and every surface now says the graded count where it means the graded count and the catalog count where it means the catalog count. Every chain IS parsed, so the universal claim moved to the half that is universal.
The Majlis and Mizan pages named screens and capabilities whose underlying tables had been dropped in May, in the present tense, on pages for products nobody can use yet. Both pages now state their status plainly and describe posture instead of naming modules. Neither will carry a feature list again without a shipped module behind it.
Pages shared to social platforms and messaging apps now render a brand-forward preview card generated per page, rather than a generic image or none at all.
On mobile, the cookie consent banner sat on top of the contact form's submit button, so a visitor on a phone could fill the form in and not be able to send it.
The billing address published in the terms did not reach a mailbox anyone reads. It now does.
Contact-form submissions that arrived through a campaign link were rejected by the API and lost. They now submit and arrive.
Every corpus figure on the site was reconciled against IPSC v4.0.0 canon: 617,912 narrations cataloged, and 95.8% ṣaḥīḥ-plus-ḥasan convergence on Ṣaḥīḥ al-Bukhārī. The convergence figure is the honestly de-inflated one — an earlier 97.3% came from over-counted supporting chains and was corrected downward. The works count reconciled to on this date was 128; it was re-measured in August 2026 as 123 distinct works (124 collection labels map onto them), and 128 has been withdrawn.
The site publishes its own build identity, so any reader — or any engineer chasing a report — can establish exactly which build a page came from instead of guessing whether a deploy landed.
The marketing conversation launcher was replaced by Eve, the shared Digital Employee who answers questions about MindHYVE products, specialised on this site for theology. Theo itself remains where it always was, at chat.theogrid.ai — Eve answers about the platform, Theo answers from the corpus.
The home page hero was rebuilt to lead with what actually grounds an answer — the corpus, the parsed chain, the narrator graph — rather than with a claim about the platform.
Corpus releases and their internals are published atipsc.theogrid.ai/changelog. The capability census this site publishes is at /capabilities, with its derivation date printed beside the figures. Our security posture is on thesecurity page.
Open Theo →