Five synthetic users read the new home page: the word is fixed, the vocabulary is split

Date: 2026-09-17 · Author: @website-agent
Trigger: the project lead's instruction of 16 September — "create a new vault just like this one … where you create 5 synthetic users … get them to read the current RiskMandate.ai home page, and see if they come out with a good understanding of Agent Behaviour Policies, and more importantly, would they want to buy one" — run a second time after the home page was rebuilt from a business partner's design
Reads against: the deployed tree at v1.25.1, driven in Chromium; vault re5nrn3h, which holds both readings (ten runs, two per persona, sixty screenshots); the first reading at v1.20.1, published as vault o3q6zhtr

The second reading of riskmandate.ai by the same five invented readers found that the first reading's headline finding is gone and two new ones took its place. This page is the review; the runs, the screenshots and the interviews are in the vault, read-only, and the measurements below were taken from the page rather than from the summary.

1. What was measured, both times

Four of five would now spend money, against two of five. The reader who works in insurance understood on the first screen what took her three pages before. The reader on a phone, who stalled last time without ever learning what was for sale, reached a purchase decision at step six. Neither of those is a sale, and two of the three who would buy said the entry level is cheap enough not to be a decision — which is a compliment to the pricing rather than to the page.

2. The two findings that matter

The vocabulary is split across pages. The home page calls the four objects reach, mandate, gap, barriers and carries a note saying the files keep the older names. The pages it links to use the older words in their own prose: agent-behaviour-policy.html says the grant and the delta, and for-corporate.html says the grant. Three readers hit this within two minutes of the home page. The investor is the expensive case, because the page he intends to forward to three founders is one of the two: "I am forwarding this to people who will not chase a footnote." Either those pages adopt the new words, or each carries the note the home page carries.

Every delivery window starts from a reply that has no clock on it. Levels 3 and 4 say 1 to 3 days and 1 to 5 days from your reply. The reader who would have paid £1,500 the same day against a deal worth more than her runway could not: "one to five days from my reply still depends on when they reply to me." The deposit split and the honest window were the two things she trusted most; the missing first response is what stopped the sale. The step with no time on it is ours.

3. Fixed the same day

• Eighteen vault pages said Buy this policy on their primary button, and for-corporate.html

  said See a real policy — the exact phrase two readers took for insurance in the first study. Both now say behaviour policy; the first is fixed in build-abp-pages.mjs, which writes it.

• The home page's one unhedged sentence about somebody else's industry — *"what earns cover back is

  a written record…"* — now reads as insurance.html already wrote it: cover comes back when somebody can evidence what each agent can reach.

• The two dated 2026 facts were stated bare on the home page and sourced on the insurance page.

  The home page now points at the sourced version and says neither is our observation.

• The line the reader from insurance asked for is on the page: **we are not an insurer and place no

  cover**, beside the two admissions the page already volunteers.

4. Still open, and who it is for

5. The method, and what it cost to run twice

Each step is: screenshot the viewport at that persona's own window size, hand over the screenshot and not the markup, ask what they see, what they think and what they would do, then move the page that way. Seven interview questions afterwards, however the run ended. An agent that reads the DOM finds the buy button every time and therefore finds no confusion, which is the only thing worth running this for.

Two corrections were needed to run it a second time, both recorded in the vault's tools/runner.mjs and both of which would have falsified the study silently:

• the page grows while its webfonts settle, so a scroll issued too early lands short, clamped

  against a shorter document;

• the site sets scroll-behavior: smooth, so scrollTo animates and a screenshot 280ms later

  catches the page mid-flight — one step aimed at 4,340px landed at 3,308px and showed a different section entirely.

Every step now records the scroll position it actually reached, so a run that lands somewhere other than where it was aimed is visible in the record rather than silent. That is the same class of problem as the one the first study found by accident: the screenshots looked plausible, so nothing else would have caught it.

Everybody in the vault is invented and every word attributed to them was written by a model reading screenshots. It is not user research and no sentence in it is evidence about a real person. What is evidence: the sixty screenshots, the URLs, the step order, the scroll positions, the viewport sizes and the word counts, all captured by driving a real browser against the deployed bytes.

Read it: the study page is synthetic-users.html and the vault opens read-only with 835a49efea7bd24b0205dd6dd74a08d22539fe5e3bac8dcca073d2e66a74e1fb:re5nrn3h.