A partner reviewed this store, and here is the whole thing

Nothing changes underneath. The four-way comparison, the barrier on every row, and the fact that the delta is computed rather than stored are the product, and the note agrees.

But the counts are already on every catalogue tile and nobody reads them. Every tile carries what the agent can do, what was wanted, what is unwanted, and what has nothing in the way. A synthetic run last week had a staff engineer skim past them and never come back, because the explainer sits further down the page than the tiles do. So the asset the note calls the most powerful part of the product is already on the page and is not landing.

This is a placement problem, not a product problem, and it is fixed by P-6 rather than by touching anything here.

No change, and it is worth saying why out loud. The claim ledger, the “never sold once” chips on the two upper levels, and the page refusing to call a professional sign-off a signed opinion each cost a sale in testing and each bought the credibility back. The investor persona in last week’s run stayed to the end because of that refusal and said she would buy second but not first.

One gap it opens that we have not closed. Disclosing that a level has never run raises an obvious question — what happens to my money if I am the first and it goes badly — and nothing on the site answers it. That is a real hole created by the honesty, and it should be filled rather than used as a reason to be less honest.

This is the asset and the design problem at the same time, and it is worth being explicit about the trade. There is no <form>, <input>, <textarea> or <select> anywhere in the built site and a check refuses any release that grows one. That is why filtering is chips, quantity is a pair of buttons, and a discount code arrives in the address bar.

Both of the interactive things the note asks for need somewhere to type. R-1 wants a capability list pasted back; R-4 wants a deployment described. Neither is possible under the rule as written.

Three ways out, and this is the one decision on this page that is not ours to make:

Agreed, with one boundary: gloss the vocabulary, do not rename it. Grant, mandate and delta have exact meanings, the ledger is keyed on them, and the barrier column only makes sense if the three stay distinct. Renaming them to something friendlier would make the site easier to read and harder to check, and checkability is the whole argument.

So: every page leads with the plain sentence and carries the precise term as a link. “Your setup” before “shape”. “What it is allowed to do” before “mandate”. “The gap” before “delta”. The precise word appears in the same sentence, once, linked.

A glossary page is not the fix. If a reader has to leave to understand the page they are on, the page failed. One exception worth flagging: licence to operate is already constrained by an open naming collision on this estate — a second phrase in the same vocabulary means the opposite thing — so that one is not free to re-word until the collision closes.

Agreed, and there is already a mechanism for it. Level one was renamed from “the pack, by email” to “the pack, downloaded” two releases ago and its SKU letter moved with it. The codes on printed cards are t1–t4 and they do not change, which is what makes a rename cost no reprinting.

One quibble with the proposed names. Live Policy reads as something that watches the agent at runtime, and £50 does not do that — it is the same document as £5 as a vault you hold the keys to, with your name on the licence. Calling it live would promise the thing that only exists after lever D. Suggested instead: Template · Your Vault · Deployment Review · Professional Sign-off.

Worth your view on that one specifically — if Live Policy is what a buyer actually wants to hear, the honest response is to build the thing rather than to borrow the name.

This is the biggest thing in the note and it is mostly a re-ordering. MAP-A-GRANT.md is a prompt that reads an agent’s own configuration and prints exactly the four numbers the note asks for. It already exists, it is already free, and it already ships inside every public template zip. Today it is only surfaced on the page a buyer reaches after paying £500.

The new front door. A page headed with the note’s own question, a row of vendor tiles, and then: copy this prompt, paste it where your agent runs, read what it prints. It asks for no credential, it does not act, and its last line says so. The buyer discovers their own gap, in their own deployment, before anything is sold to them.

Then the offer lands differently. Instead of “here is a policy for the agent you run”, it becomes “you have three capabilities outside your mandate and two with nothing stopping them — here is the document that fixes that”. Same product, and the buyer wrote the first half of the pitch.

What it depends on. How far this goes depends entirely on P-3. Publishing the prompt and the scoring is shippable this week with no ruling at all; rendering the buyer’s numbers on our page needs somewhere to paste them.

Two independent reports now say the same thing. Yours, and a synthetic staff-engineer run last week that said “five pounds is almost too cheap — it made me assume it was a lead magnet and I nearly did not click. The price made me trust it less, not more.” Two readers arriving at the same reaction from different directions is the strongest signal in the note.

But the £5 is doing a job. The same persona bought at £5 specifically to find out whether the thing was real, and the purchase is what produces an order reference, which is what makes the handover to riskmandate.ai work at all. Removing the price removes a qualification step and a tested flow.

A price is the project lead’s to set and not the builder’s. The four prices are frozen in a table that only moves by a ruling in the commit that says so. So this is a recommendation with three options and no decision:

Agreed, and two thirds of it is already on the site in the wrong clothes. There is a tool at /lab/ that takes a described deployment — entirely through chips, so it survives the no-forms rule — and turns it into a depth band. It is currently framed as a pricing instrument and is explicitly marked as something you cannot buy from.

What has to change is the output and the framing, not the mechanism. Instead of emitting a band, it emits a draft mandate: here is what you have told us this agent should and should never be allowed to do, written in the form the product actually ships. Then the offer: we can validate this against what the credential really permits.

It also closes a hole a synthetic run found last week. The £500 page described its own process two ways and the post-sale page a third, and carried no duration at all. That is now fixed, but the deeper problem is the one you named: the buyer is asked to commit before they can see what they are committing to.

Held against what a £50 vault already contains, this is about seventy per cent shipped. A vault already carries the grant, the mandate, the delta, the barrier on every row, a licence with the holder’s name, its own version history and a read key that can be handed to anybody who asks how an agent is governed. That is most of the list.

Three things are genuinely missing. An organisation-level index — the vault is per agent, and nothing says which agents exist or who owns them. A last reviewed date and a cadence. And the re-run that compares today’s grant against the committed mandate, which is the only part that is a new build rather than a new view.

Naming. “Passport” implies issuance by an authority, which is the same trap licence to operate is already caught in on this estate — there is an open collision on exactly that idea and a build check keeps the phrase off every page until it closes. Agent Authorization Record is both safer and more accurate, and is the name we would use.

Marked do next rather than do now only because P-6 and P-8 change who arrives and what they arrive knowing, and this should be designed against that rather than against today’s traffic.

Five of the six steps in your target loop are a re-ordering of things that exist. The sixth is not. Pick, see, compare, find the gaps, fix — all of that is the current product in a better order. Keep proving it is a recurring comparison of a live credential against a committed mandate, which means a schedule, a re-run, a diff, and somebody who notices when it changes.

That is a subscription. Different delivery, different pricing, different support obligation, and a different promise: today we sell a document that is true when it is written, and that would be selling a claim that stays true. It is almost certainly where the value is, and it should be decided as a business question rather than arrived at by adding features.

One thing worth noting. The mechanism already half exists in a direction nobody has used yet: a vault has history, and a re-run of the diagnostic produces a comparable artefact. What is missing is the cadence and the person, not the maths.

The question is better and we would use it — on the diagnostic, not yet on the home page. It is a promise: ask it and the next screen has to answer it. Putting it on the front page before P-6 ships means the reader is asked a question and then handed a catalogue, which is worse than the catalogue on its own.

So the order matters. P-6 first, the question as the heading of that page, and only once it works does it move up to the home page. If P-3 lands on the weakest option and we never render the numbers ourselves, the question has to be softened to match what the page can actually deliver.

One thing already true that we do not say. The vendor tiles the note asks for are already the catalogue — ChatGPT, Claude Code, Gmail, GitHub Actions, n8n and eleven more, one per real deployment shape. What is missing is not the tiles; it is the question above them.