What's built, and what isn't.

Builder, editor and publish work; structured data verified in the generated pages.

Verified in the generated sites' HTML.

Fixed 22 September and verified on development: a summary of the business, its products and services for AI assistants. Reaches production at the next release.

Each site has its own sitemap, but a platform-wide sitemap address takes priority and answers instead, with the wrong pages. It also lists pages only, not blog posts or products.

Real Cloudflare integration; no customer domain created yet.

Answers visitors from the business's live timetable, prices, staff, events, shop and policies as cards; captures a lead when it can't answer; rate-limited per visitor.

Sign-in, shop orders and invoices, packs, subscriptions, downloads, returns, support tickets, profile and security all work.

Bookings are view-only today. A cancel-and-detail module exists but isn't connected; service invoices arrive by emailed link, not in the portal.

Answers from the customer's own bookings, orders, downloads and memberships. Read-only: it can't yet move a booking for them.

Send, verified inbound and a drag-and-drop builder all work.

Sending credential absent and unsubscribe routes unmounted. Deliberately not switched on until both are fixed.

Real Twilio integration and verified webhook; the missing callback settings landed 22 September and reach production at release.

Sync is built: details, hours and closures push to Google automatically on change, and services and products push once registered. New reviews are pulled into the Inbox every 15 minutes, with replies sent back. But connecting a Google account can't complete yet, because the Google app credential isn't configured, so none of it has run against Google. Posts, photos and insights call the wrong Google API, and the review-reply button on the profile page is a stub.

Products go to Google as service items, not a product catalogue. There's no Q&A, no working Merchant Center feed and no GEO visibility tracking.

White-label Stripe Connect accounts, a 7-step onboarding inside Simma, ID verification and an audited KYC trail. Live-mode keys not yet in production; they go in before the first paying customer.

Shop, bookings, invoices, payment links and first membership charges go to the business's own account with Simma's fee taken. Refunds, disputes, payouts and Xero/QuickBooks sync work. Built and tested in Stripe test mode; switched to live keys before the first customer.

The v10 draft prices (the card's actual cost plus Simma's fee, from 2% + 20p on Free down to 0.4% + 5p on Scale, with payouts by plan) are not what's charged yet. The code uses older rates (about 1.5–3% plus a fixed fee), and takes no fee on in-person sales, recurring memberships or deposit balances. One rate source is the fix.

The code exists; none of it is on a schedule yet, so today it only runs when triggered by hand.

Not built. Not claimed on the Money page.

A future direction on Stripe's embedded-finance products, country by country. Nothing built. With the other ideas for the platform on Where Simma goes next.

Resolved on every page, with a precedence test for each pair of layers. The industry, specialism and niche layers are built; industries are added as configuration, starting with instructors. 23 countries. A café is tested to receive none of a dance school's words, flows or menus.

The machinery runs, but franchise settings, and menus by specialism, niche or country, have no content yet. Website starting templates exist for dance schools only. The customer business-type picker offers four types and is admin-only.

Read from the code, 11 October 2026: head office keeps a product catalogue with a price band per product, and locations price inside it (anything outside is refused); a brand site's location finder leads to each location's own site; every paid booking or order records the franchisor's royalty, and a cancellation reverses it (a record of what's owed, not a split of the money). Not yet: one network shown working end to end, a central catalogue of services, and content in the franchise settings. Franchise networks →

Six languages are declared; French is partly translated. A business is one kind today, so a studio with a shop can't be both.

One definition of headcount, counted live and enforced by tests. By design it changes emphasis (marketing mode, setup, suggestions, AI context) and never hides a feature.

Consent is recorded with its lawful basis; marketing and AI consent are separate; a preference centre; a cookie banner in UK GDPR, EU GDPR or California opt-out mode.

Every action is written to one audit log, with whether a person or an agent took it.

Requests tracked against the 30-day deadline, with identity checks. The export says when it's incomplete rather than hiding a gap, and states how many safeguarding records it withheld, never their content. Staff pages that failed to load were fixed, and a team admin can download an export as JSON or CSV (27–28 September; in the next release).

Erasure requests are logged and the deadline runs, but nothing yet deletes a person's records. Built before the first business whose own customers could ask.

Delete, anonymise and archive rules and legal holds are built; nothing runs them on a schedule yet.

Not built for US health records, and not claimed. Booking and customer forms warn people not to enter health information.

Live inside Simma's development work system, not yet in the customer product. Each job is graded simple, standard or complex and offered only to AI agents on the matching model tier (Haiku, Sonnet, Opus), with escalation when a tier fails. Measured on matched jobs: Haiku used 73% of Sonnet's tokens and passed 2 of 3; Sonnet passed 3 of 3.

First live test on a real Simma decision, 27 September 2026: agreed with the current classifier 30 of 30 times, 0.3 s median. The Judgment service is now built, with 15 decisions registered; each is still answered by the current AI model. Jev can be switched on per decision, alone or alongside the model for comparison, with a tool that measures accuracy, confidence, speed and cost. Switched off for every decision today.

The feed engine and 7 publishers (waitlists, orders, stock, prospects, overdue invoices, content to approve, overdue tasks) are built. A step waiting on the owner now gets its own card, and a fix stops a private task's title reaching colleagues (29 September). The new feed is behind a switch that's still off, so Home shows the older intranet feed with memory cards.

An idea card can be approved from Home; the plan, its steps and who approved them are recorded. A "Waiting on you" page now lists every step and idea waiting for a person, and content can be approved from it, at the approval role; no agent can approve for a person (28 September–1 October; in the next release).

Agents are named identities with a list of permitted tools; every task step records whether a person or an agent did it; every agent action writes to the audit log.

Switched off on 14 September because suggestions appeared in the owner's own to-do list. It returns once agents have their own list.

Roles, limits and approval rules are designed; the reasoning engine they run on is live. As named, customer-facing agents they are the next build, and they appear on this site as vision.