A built thing you can walk across, and look under.
In short — One working build, audited against piece 5’s five requirements — including where it falls short. It refuses to fake agreement (that’s a runtime error, not a setting), leaves barred seats empty, decides nothing itself, and seals a tamper-evident record. Don’t trust us; check the working.
Reading on — Piece 7: one room is only a pilot. What happens when the same question runs in many rooms at once?
The last piece wrote a list: five things a room must do to hold a disagreement instead of faking its way out of one. This piece hands you one real attempt at building that room, and asks you to do the obvious thing — check it against the list, including where it falls short. Don’t take my word for any of it. That’s not modesty; it’s the whole design.
Start with the hardest requirement: keep the disagreement, word for
word. Most “consensus” tools quietly do the opposite — they average,
they tidy, they find a middle nobody quite holds. This build cannot.
When it draws many rooms’ results together, a rule runs over the output
before anything is saved: if a real split has been blended into a smooth
agreement, the process throws an error and refuses to
save.1 Faking agreement isn’t discouraged
here; it’s a runtime error. The words that would let it happen —
average, percentage, overall score, tally — are banned
outright; ask for them and the build stops.
✓ requirement 1
Requirement four: leave the missing seat visibly empty. In a
rehearsal, the panel is filled with composites — drafted stand-ins built
from published sources — and it would be easy — and false — to draft one
for everyone. So the build refuses two ways. It refuses to fabricate a
tangata whenua voice: those seats are barred to any
synthetic stand-in, ambiguity defaults to refuse, and the seat is left
registered and empty, routed to real people.2 And
it never lets the composites pass as real: they are labelled, in their
own records, as what they are. The caveat and the mechanism are the same
object — the empty seats are what keep this room truthful.
✓ requirement 4
Requirement five: decide nothing itself. There is no path in the code
from anything the machine produces to an actual decision. A binding
result can only be created inside a human confirmation — a person’s
hand, recorded — and the rehearsal’s own records open with their status
set, permanently, to draft for contestation.3 The
machine argues, summarises, maps, provokes. Then it stops, and hands
back. The software rehearses; people decide.
✓ requirement 5
So what does it produce? Not a recommendation — there is no field in
the record that could hold one. It produces a map:
where the argument converged, where it split and along which lines, what
was left thin on evidence, where a fast agreement looks suspicious
rather than reassuring.4 A map of the disagreement is the
opposite of a verdict, and it is what a rehearsal is for.
✓ requirement 3 — every AI hand shown
Requirement two: show its working. Every turn the machine takes is stamped with a fingerprint of exactly what went in and what came out; the whole record is then sealed, write-once, so it cannot be quietly rewritten or backdated.5 The point is not to ask you to trust the people who ran it. It’s the move piece 2 left hanging: when you can’t trust the deliberator, you stop trusting the deliberator and check the deliberation instead. Don’t trust me — check the process.
And here is the shortfall, named rather than hidden. Today the seal
is tamper-evident — you can tell if it’s been altered — but it
is not yet something a total stranger can verify without trusting us
about one key. So on the requirement that matters most, this build does
not yet fully pass; it strains. Saying so is the requirement working on
itself. ⚠ requirement 2 — passes, with a gap
One more thing, plainly. This is built and it runs — but it is switched off in production, and that is a refusal, not an apology. Turning it on for real communities is itself a decision that has to pass the test: it needs real seats, real people, real consent, and a seal a stranger can check — and those conditions aren’t all met yet. Switching it on before they are would be exactly the kind of move this series argues against.
“Off in production” answers can you use it today — not does it exist. It exists: there is a published rehearsal, a real run over a hard question, with a sealed record you can point at.6 Look at the run, not the promise.
A test you only run on other people is a weapon. So, graded against the same fairness scorecard we’ll hand you in the next piece, this build earns clear marks on keeping dissent and showing its working, a fair mark on whose voices are present, and a low one on stranger-verifiability and on ever being used for real. A low letter we didn’t hide is the sign the grading is real.
The room can be built. This is one attempt, with its seams showing and one wall still to finish. Which is the only kind worth trusting:
Held in kōrero, not collapsed to a number.
Next, and last: “Many Rooms, One Map” — why one room is only a pilot, how many rooms become a country hearing itself, and why we would rather you built your own than adopted ours.