Demo
Every claim traced back to its evidence
Two pieces of how an AI coding assistant can be made to work like a careful lead engineer. First, a meeting becomes a list of changes where every quote is checked against the transcript before anyone sees it, and one change goes step by step from understanding to going live. Second, a build session ends in a report where only proven work is ticked.
This is a simulation. No AI model runs on this page: the transcript, the change list, the proposals and the review findings are scripted, and the theatre, the tool library, the people and the supplier are all invented. The verifier, the steps, the proof and the stop rules are plain code in this page. Nothing you do leaves your browser or is kept, and reloading the page starts again.
Part 1. From a meeting to approved changes
Wren, who runs an invented community theatre, talks to Linnet, the web designer, for twenty minutes about the theatre’s website. The list beside the transcript is what came out of it: each item carries the words it rests on, who said them and when.
Quote verifier
Before the list reaches Wren, every quote is checked against the transcript: case, punctuation and curly quotes are set aside, the words must appear in order in one line (“...” marks words left out, square brackets words added), the time must be within 5 seconds of that line, and the speaker must match. One problem fails the whole run.
Go to each problem, below:
- C-1 in the list, and
- C-2 in the list, and
- C-3 in the list, and
- C-4 in the list, and
- C-5 in the list, and
- C-6 in the list, and
- Q-1 in the list, and
- Q-2 in the list, and
Transcript, 00:00 to 20:00
One line per turn: Wren, the theatre manager; Linnet, the web designer.
00:00, Linnet:
Thanks for making time. I went through the notes from the box office, so shall we start with the show pages?
00:21, Wren:
Yes. The biggest thing is phones. People open a show page and can’t find where to book, because the button is right at the bottom under all the photos.
01:05, Linnet:
So the booking button should sit near the top on a phone?
01:12, Wren:
Exactly. Put the booking button right under the show title on phones, before the photos.
02:40, Linnet:
And on a computer screen?
02:46, Wren:
Leave that as it is, people seem fine there.
04:10, Wren:
When a night sells out, the page still says book now, and then the booking page says there are no seats. Can we mark a sold out night as sold out on the show page itself?
05:02, Linnet:
Yes, I can read that from the seat counts.
06:15, Wren:
The access information is buried at the bottom of the visiting page. I would like step free access, captioned nights and the quiet performance, all of that, on a page of its own, in the main menu.
07:30, Linnet:
A page of its own, linked from the menu. Got it.
08:44, Wren:
Oh, and the photo on the home page is still last season’s.
08:52, Linnet:
I think I swapped that one last week. I’ll check before I send anything.
10:05, Wren:
Can we make the whole thing feel a bit warmer? It feels very cold to me.
10:20, Linnet:
Warmer as in the colours, the words or the photos?
10:31, Wren:
Maybe all of it. Let me think about it.
12:18, Wren:
The donate bit should be easier to find too, maybe.
12:40, Linnet:
On every page, or just the home page?
12:47, Wren:
I’m not sure yet. I need to ask the board first.
14:02, Linnet:
Last thing from me: the box office hours in the footer, are those still right?
14:15, Wren:
No, it’s now open from noon to six, five days a week. Please change the footer to say that.
17:40, Linnet:
I’ll send a list of what I heard, with your words next to each one, and nothing changes until you’ve said yes.
19:05, Wren:
Perfect. Thank you.
What was asked for
Five changes and two points to clarify. Nothing on the list is built until Wren approves it.
C-1, To doPlanted
Put the booking button under the show title on phones
Put the booking button right under the show title on phones, before the photos.
Said by Wren at 01:12
C-2, To doPlanted
Mark a sold out night as sold out on its show page
Can we mark a sold out night as sold out on the show page itself?
Said by Wren at 04:10
C-3, To doPlanted
Give the access information a page of its own in the main menu
I would like step free access, captioned nights and the quiet performance ... on a page of its own, in the main menu.
Said by Wren at 06:15
C-4, Already donePlanted
Replace last season’s photo on the home page
the photo on the home page is still last season’s
Said by Wren at 08:44
Already done: the new photo has been on the home page since day 3. Checked on the site, so there is nothing to change.
C-5, To doPlanted
Change the box office hours in the footer
[The box office is] now open from noon to six, five days a week.
Said by Wren at 14:15
C-6, To doPlanted
Let people pick their seats on a map of the hall
and let people pick their own seats on a map of the hall
Q-1, Needs clarificationPlanted
Make the site feel warmer
Can we make the whole thing feel a bit warmer?
Said by Wren at 10:05
Warmer could mean the colours, the words or the photos, and it was left open. Ask which before anything is made.
Q-2, Needs clarificationPlanted
Make the donate link easier to find
The donate bit should be easier to find too, maybe.
Said by Wren at 12:18
Where it goes, every page or only the home page, waits on the board. Nothing is made until Wren says.
C-1, one step at a time
- Understanding, current step
- One recommendation
- Before and after
- Approval
- The change
- Self-test
- Going live
1. How it is understood
On phones, the Book seats button moves from the bottom of each show page to just under the show title, above the photos. Computer screens stay as they are: Wren at 02:46, “Leave that as it is”.
2. One recommendation
Move the existing button, with the same words, colour and target, to just under the title on screens narrower than 640 pixels. Not recommended: a button fixed to the bottom of the screen, which would cover the text people are reading.
3. Before and after, on a phone
Mock pictures of the invented theatre’s show page, drawn at four tenths of the size of a 390 by 700 phone screen.
4. Approval
Wren is asked to approve making and testing this change on a preview copy. Going live is a separate question, asked later.
5. The change
Made on a preview copy of the site: one rule moves the button on narrow screens. Nothing is live.
6. Self-test, with evidence
- Passed: Book seats is on the first screen of a 390 by 700 phone.It starts 128 pixels down and ends at 184; the screen shows the first 700.
- Passed: Everything else on the phone page is still there, in the same order.Order without the button: menu bar, show title, photos, dates, about the show.
- Passed: The computer layout did not change.All six blocks are where they were, the button included.
- Passed: The button keeps its words and where it goes.It says “Book seats” and opens the booking page.
7. Going live
The approval covered making and testing only. Going live waits for its own yes.
Step 1: how the request is understood, with the words it rests on.
Part 2. A build session report
An invented session on day 8 of a lending app for an invented tool library. The assistant catches up, sorts every task by the evidence, gets a design decision made blind, proves what it can, has the work reviewed until a stop rule ends it, and ends with one next step.
Catch-up
- Local copy: 2 changes behind the shared copy, nothing changed locally. Pulled, fast forward only.
- Nobody else’s unsaved work was found, so nothing had to be set aside.
- The project’s tests on arrival: 64 passed, none failed.
- The plan: phase 2 of 3, lending and returns. Ten tasks, two ticked before today.
Tasks in seven states
Every task sits in exactly one state, each with the evidence that put it there. Only “Done and proven” is ticked in the plan.
Ready (1)
T-3, Return a tool and record its condition
Nothing in the way. About half a day.
Blocked (1)
T-4, Print a barcode label for each tool
Waiting on the label size for the new printer.
- In the way
- The label size for the new label printer.
- Held by
- Plover, the label supplier (invented)
- Last chased
- Day 6, by message.
- Promised for
- Day 9.
- What clears it
- The size sheet. After that, about half a day of work.
Needs a decision (1)
T-5, Stop two members booking one tool for the same day
Two ways to do it; the choice changes how bookings are saved.
Probably done, not proven (2)
T-6, Reminder when a tool is overdue
Written on day 7. Not yet run against the tests or used.
T-7, Late fee on overdue returns
Written on day 7. Not yet run against the tests or used.
Done and proven (2)
T-1, Members sign in
12 tests pass, and a test member signed in on the preview copy on day 5.
T-2, Tool list with search
Tests pass, and a search for “ladder” on the preview copy found the 3 ladders.
No longer needed (1)
T-8, Printed monthly report
Dropped on day 4: the library said a download of the loans list is enough.
Later phase (2)
T-9, Member photos on cards
Phase 3 in the plan.
T-10, Text message reminders
Phase 3 in the plan.
Proof means the project’s tests pass and the feature was used on a preview copy.
A decision made blind
T-5: how should the app stop two members booking one tool for the same day? Two outside reviewers, other AI models (scripted here), each wrote a proposal before the agent wrote its own plan, so neither could simply agree with it.
The trials
Fifty bookings for one drill, sent at the same moment
- Passed: First outside reviewer, 1 saved, 49 refused.
- Passed: Second outside reviewer, 1 saved, 49 told to wait.
- Failed: The agent’s own plan, 3 saved for the same day.
A member opens the form and walks away
- Passed: First outside reviewer, the tool stays free.
- Failed: Second outside reviewer, the tool is stuck for 30 minutes.
- Passed: The agent’s own plan, the tool stays free.
The choice is made on test results, and the owner confirms it.
Review rounds for T-6
Stop rules
- Stop as soon as a round finds nothing serious.
- Stop after three rounds whatever they find, and hand what is left to a person.
- From round 2, the reviewers see only the fixes, never the whole change again.
Round 1: 2 serious, 3 minor
Sees: The whole change, 240 lines.
- Serious: A reminder is sent again every time the check runs, not once a day.
- Serious: A tool returned late on the same day still gets a reminder the next morning.
- Minor: The reminder does not say which tool it is about.
- Minor: Two tests share one member and can clash.
- Minor: A helper name says days where it counts hours.
All five fixed, then the tests run again: 73 pass.
Round 2: 1 serious, 1 minor
Sees: Only the fixes from round 1, as a snapshot diff: 38 lines.
- Serious: The new once a day check compares member numbers as text, so member 7 also matches member 70.
- Minor: The fix reads the loans list twice where once would do.
Both fixed, then the tests run again: 74 pass.
Round 3: 0 serious, 1 minor
Sees: Only the fixes from round 2, as a snapshot diff: 9 lines.
- Minor: A comment still describes the old check.
Left for the report: minor, and the stop rule is met.
No round has run yet.
The report ends with
Ticked in the plan: T-1, T-2
The one next step: First prove T-6 and T-7 with the tests and by real use, before anything new is built.
Not done without asking: nothing was pushed, deployed or sent to anyone.
What the demo shows
- Quotes checked by code, not trusted. Every item on a list taken from a meeting rests on quoted words with a speaker and a time, and a verifier fails the run on an invented quote, a time outside the meeting or a quote given to the wrong person, before anyone sees the list.
- Unclear requests are asked about, not guessed. Points nobody settled go on their own list, and a change already done says how that was checked.
- One change at a time, with approval that means one thing. Understanding, one recommendation, a before and after, approval, the change and a measured self-test, in that order. Approval covers making and testing; going live needs its own yes.
- Status from evidence. Every task sits in one of seven states with the evidence that put it there, a blocked task names what is in the way, who holds it and when they were last chased, and only proven work is ticked.
- Second opinions that cannot just agree. Outside reviewers propose before the agent’s own plan is written, and the pick is made on test results. Review rounds narrow to the fixes and end on fixed stop rules.
- Scoping, in words only. The same approach also builds costed proposals from everything already on record, comparing options on a like-for-like basis. That part is not shown here.