Take the moves from the concepts

The one thing that would outweigh the whole panel. Five invented readers cannot tell you what a stranger does, and the weakness is not the sample size — it is that the sample and the thing being sampled have the same author. A bigger panel scales the problem rather than the evidence, and is recorded as won't-do on the review so nobody reads the silence as a plan.

One person who has never seen the store, shown the same four folds at the same two sizes and asked the same six questions.

Raised twice, independently. The outside review flagged that “licence to operate” can be read as regulatory permission unless something beside it says otherwise; the panel found “powerful agents, clear boundaries” was the only headline every reader understood on sight.

Neither settles it, and nobody misread the live headline. It is also the name of a document inside the vault, so changing it on the homepage alone splits the vocabulary. A ruling, not a change.

A live defect, found from outside. This store prices the pack at £10 and riskmandate.ai's homepage still carries “1 · £5 The pack, downloaded”. The price moved on a memo from the project lead and the product site was not moved with it.

A buyer who reads both properties — which is the whole point of them being one flow — sees two prices for one thing. It is the other team's file to change, so it goes to them as a correction rather than a suggestion.

The one section of the chosen design that is not built. The v3 pages carry a mission line under the heading why RiskMandate exists that uses a word this site bars absolutely, on every page, with no allowlist and no carve-out for a quotation — because a positioning phrase is a claim, but a priced checkout is an offer, and every page here carries a price.

Their own notes page records that the project lead asked for that line. So this is a collision between two things somebody is right about, and it is not the builder's to settle. Three ways out: the line stays off store.sgit.ai and lives on riskmandate.ai instead; or it is rewritten to say the same thing in words a priced page may use; or the rule is changed, which is a bigger decision than one section.

Recorded at /admin/ rather than silently dropped.

Five of the fifteen published shapes overshoot their own grant by exactly one. Each shape publishes four totals: the size of the grant, how much of it was wanted, how much was not asked for, and how much has nothing real in the way of it. For ten shapes wanted + not-asked is exactly the grant. For claude-code-cli, claude-code-cli-confirmations-off, claude-desktop, github-actions and scheduled-job it is the grant plus one, every time.

Off by one on five shapes and exact on ten is a convention, not noise — most likely one capability counted on both sides of a split somewhere upstream. The picker says so on those five rather than smoothing it over, and a check refuses a note on a shape that does add up. What it cannot do is say why, and only riskmandate.ai can.

Found building the policy picker: the panel's capability strip was being painted green for wanted and amber for unbounded, which for these five drew a partition of a grant that is not one.

The picker offers a filter it cannot run. The capability vocabulary is promoted into this repository — 23 named behaviours, verb.object.reach — and every shape publishes how many of them its grant contains. The join between the two is not published: nothing says which behaviours are in a particular grant.

So By behaviour on the picker lists all 23, and choosing one narrows nothing and says why. That is the same answer the design team reached independently on the same gap. It is the honest answer and it is not a good one: the reader who filters by read.credential.host is exactly the reader worth having.

One field per shape upstream — the list of capability ids in its grant — turns the filter on, colours the capability strip in the panel, and makes the fifteen shapes comparable on something other than a total.

Six published vaults, and they begin 3.4 screens down. The band starts at 3,098 pixels of a 4,395-pixel front page at 1280, and at 6,677 of 9,406 on a phone — 7.9 screens.

Two panels have now named it as the thing that decides them, both times the readers with the largest budgets: on 16 September it was why the live page beat three concepts that had deleted it. A build check put it back on the new front page the day that page became the store, and it went back at the bottom.

Under the four facts, above the four levels. A reader who has seen six real documents before seeing a price reads the price differently. Raised by the five-reader journey of 16 September at /admin/reviews/.

521 pixels of filter bar, eighteen chips, and the first of the fifteen policy shapes at 1,220 pixels on an 844-pixel viewport — a screen and a half of ways to narrow fifteen things before one of them is visible. Eight of the eleven category chips have a count of one.

Collapse the whole bar behind one control that says how many shapes are showing, the way the behaviour menu already collapses. A chip that filters fifteen things down to one is a link to that thing wearing a filter's clothes. Raised by the five-reader journey of 16 September at /admin/reviews/.

The picker's search has no field, because this site has none anywhere. It reads the keyboard: press /, type, and the list narrows. On a phone with no hardware keyboard that is nothing at all, and the dialog says so and points at the chips.

The design handback solved it with an on-screen keyboard drawn out of buttons. That was refused when the picker was built — twenty-six buttons to filter fifteen things is worse than the fifteen things — and the refusal is still defensible at fifteen. It stops being defensible at fifty.

The ruling is not about search, it is about how large the catalogue is going to get. If it stays at fifteen, the chips are the filter and the search control should not appear on a touch device at all. Raised by the five-reader journey of 16 September at /admin/reviews/.

From 01 Marketplace, corrected by the one reader who checks things. Four short statements between the hero and the first product section: fifteen application templates, four levels, six vaults open to read now, the top two signed by a person. Every one of the four is already true here and already in the ledger.

Each item links to its claim. The concept's version has four ticks and not one of them is a link, which is the difference between a trust strip and a claim — and the reason this store can ship the idea and the concept could not.

Superseded by the design that is the store from v0.3.16, and done there. This unit was written as an edit to the previous design; the new one carries it as built.

The live card sets the price at 28px inline with the delivery estimate, at body weight. Every concept sets it at 42px as the second-largest thing in the card, and every reader shown that version used the prices to compare.

A type-scale change, not a redesign. It is the measurable difference between a card somebody compares and a card somebody reads.

Superseded by the design that is the store from v0.3.16, and done there. This unit was written as an edit to the previous design; the new one carries it as built.

Identical order in every card so the eye can run down a column: level and delivery on one 12px line, the name, one line of what it is, the price, one sentence, one action.

The claim chips have to go somewhere. They are load-bearing — they are how a reader checks that “delivered by a person” is a thing somebody stood behind, and they are one of the two elements that won the two readers who would spend the most. This is a move, not a deletion.

Superseded by the design that is the store from v0.3.16, and done there. This unit was written as an edit to the previous design; the new one carries it as built.

From 02 Guided, which lost the panel four-one and won this outright. One pill per application, each a link, with a sixth reading “something else — from £500”.

A visitor whose agent is not in the list currently reaches the end of a directory. The sixth chip sends them to the level that exists precisely because their deployment is not the template.

Superseded by the design that is the store from v0.3.16, and done there. This unit was written as an edit to the previous design; the new one carries it as built.

The review's first recommendation and the right one. Three destinations and the order stay in the header; eleven evidence and buying-mechanics destinations move to the footer; eight already have a home in the console.

Two things do not move. The version badge is how a reader knows which build they are quoting, and the model-assistance strip is required above the main element by this repository's own rules with a check that refuses a page without it. They can get quieter. They cannot go to the footer.

Superseded by the design that is the store from v0.3.16, and done there. This unit was written as an edit to the previous design; the new one carries it as built.

It is two lines on a phone before the headline starts, and the headline starts 358px down. One line with the rest behind the link is the same disclosure at a third of the height.

It stays above main. It is one of the two elements that won the two readers who would sign off £1,500, and a build check holds the floor, so the failure mode is loud rather than silent.

Superseded by the design that is the store from v0.3.16, and done there. This unit was written as an edit to the previous design; the new one carries it as built.

Ruled on 16 September, by the project lead, out of the v3 design round. Pack, Vault, Tailored, Reviewed — a short name at heading size with the precise line underneath, which is the anatomy the design brief proposed rather than a choice between the two.

They live in data/offers.yml as short_name and short_sub beside the price, and check_next_is_the_offer_data holds both / pages to them: a name that drifts from that file fails the release, which is the point of a name a buyer repeats to their board.

The next store is a whole store. Ten pages at /, all rendered from the files the current store is already built from: the homepage, the product page, the policy picker, your order, checkout, what lands afterwards, what arrives at every level, who this is for, the claim ledger and who does the work.

The order is the live store's order: same local-storage key, same schema, same line shape, same reference alphabet, same SKUs, so an order built in this round is still there on the store that sells today, and back. Two build checks hold that and hold every page to the file it renders from.

The one screen from the pack deliberately not built is the generic template. It is a reading column already exercised by the ledger and the reviewer page; re-rendering an existing content page at a second URL to demonstrate it would put the same paragraphs in two places. Recorded at /admin/.