For asset data cleansing and information extraction, what should AI be allowed to do?
W1 · 25 Jun 2026Asset Zero · Replay the series
Answer what the rooms answered — at your pace
The framework was built from 2,796 responses to questions like these. Answer them yourself — the work-order storm, the green health score, the optimiser’s silent trade — then see how the rooms voted on the night. Nothing you do here is recorded: your answers stay on this device, and every reveal is the same published aggregate the evidence archive carries.
The ladder — what should AI be allowed to do?
The series' spine instrument: six rungs from “don't use AI” to “act autonomously”, asked across data, predictions, money and machine-triggered work. The community's ceiling held at analyse-or-recommend — until it watched bounded action work.
AI auto-flagging an anomaly for intervention — the 'green health score' (false positives, false negatives, sensor and concept drift)
W2 · 1 Jul 2026Renewal prioritisation, deferrals and budget allocation — AI-optimised capital planning
W3 · 15 Jul 2026Paired instrument
Machine-triggered work — before and after the storm
The series' controlled experiment: the identical vote, asked twice in one room (paired move +0.44 rungs, n=25).
Machine-triggered work — work, operations and intervention
Condition assessment / health rating — what is the most AI should be allowed to do?
W2 · 1 Jul 2026Performance / level-of-service interpretation — what is the most AI should be allowed to do?
W2 · 1 Jul 2026Risk & criticality scoring — what is the most AI should be allowed to do?
W2 · 1 Jul 2026Risk acceptance (judging a risk tolerable) — what is the most AI should be allowed to do?
W2 · 1 Jul 2026Paired instrument
The intervention trigger — four weeks apart
The one instrument re-run verbatim across rooms (Recommend 71%→33%; act-with-sign-off 5%→24%).
Intervention trigger (initiating maintenance / renewal / replacement) — what is the most AI should be allowed to do?
Crew dispatch and scheduling (booking people against work)
W4 · 30 Jul 2026Committing spend within a pre-set cap (ordering parts, hiring plant)
W4 · 30 Jul 2026The ladder across the series (W2 → W3 → W4 warm)
Not enough archived series data in this environment to compare.
What humans must keep
The retained-rights set — approve, validate, own the value judgement, hold the stop. Approve is top-two in every room that voted it; the stop was under-priced until W4's storm re-priced it (48%→64% paired, zero drops).
What should humans retain?
W1 · 25 Jun 2026The retained rights across the series (W2 → W3 → W4 warm)
Not enough archived series data in this environment to compare.
Trust, evidence & validation
What must be true before acting on an AI output — and how little one session moves trust (about a third of a point on a seven-point scale). Play the Green Score to feel why.
▶ Self-paced simulation
The Green Score
The Green Score
You own the maintenance budget.
The AI scores every asset's health. Over the next few minutes we'll find out, together, how much you should trust that score.
Take the budget. The AI takes the scoring. The question we're testing: how much should you trust a confident score?
On 1 July 2026, a live room played this simulation together: a condition-monitoring AI scores every asset in a fleet of 2,000, and you own the maintenance budget. You're about to make the same three calls the room made — first alone, then you'll see what they did. There is no right answer on the first two. That's the point.
For acting on an AI failure/condition prediction, what should AI be allowed to do?
W2 · 1 Jul 2026Which attribute should never be inferred without validation?
W1 · 25 Jun 2026Which of these patterns have you seen?
W1 · 25 Jun 2026Paired instrument
Trust in an AI capital plan — before and after the session
Paired +0.43 on a 7-point scale (n=14). Trust moves in inches.
If an AI system recommended your organisation's 5-year capital plan, how much would you trust it without a full human rework? — asked at the session's open
Accountability, ownership & the stop
Who owns the call, who could show you the weights, who could pause the machine within the hour — the is/ought gap the framework exists to close (36% named-pauser, 22% drilled, audit-rate guess 0%).
▶ Self-paced simulation
The $100M room
Same night, second sim. Eight projects ask for more than $100M can fund — so something gets deferred, and a deferral is a decision about a real place. You'll build your plan first; then the room's combined plan; then the optimiser's — and you'll see what it quietly defunded. (Fictional scenario, same as the night.)
Discussion — the risk owner
W2 · 1 Jul 2026Who, in your organisation, could tell you the actual weights inside your prioritisation model or framework?
W3 · 15 Jul 2026Who in your organisation owns checking what the optimiser quietly defunded?
W3 · 15 Jul 2026Is a named person authorised to pause machine-generated work in your organisation?
W4 · 30 Jul 2026Has that pause ever been drilled or tested?
W4 · 30 Jul 2026In YOUR organisation, who could actually pause this right now?
W4 · 30 Jul 2026Values, weights & where the money goes
Money is where values hide: 85% would not sign vendor-default weights nobody chose; 64% rejected the optimiser's silent trade. Set your own weights and see what they do.
▶ Self-paced simulation
Whoever sets the weights wins
15 July 2026, the planning room. A council's ten-project capital plan is ranked by a prioritisation model. The model is fine. The question is the five value weights inside it — cost, risk, service, carbon, equity — because whoever sets those decides what gets funded. You're about to set yours, and then find out who set your organisation's. (The projects and numbers are a fictional scenario, as they were on the night.)
Who could show you the actual weights inside your prioritisation model — today?
W3 · 15 Jul 2026▶ Self-paced simulation
Sign or send it back
The last W3 sim, and the sharpest. Four AI-assembled capital plans land on your desk — every one reads defensible. At least one is not what it seems, and the tell is never a klaxon: it's a footnote, an assumption, a redefined KPI. The live room voted on A, B and C; plan D went to one person — The Board. You get all four. Read slowly. (Fictional plans, as they were on the night.)
Envelopes, triggers & reversibility
When may AI move from recommending work to triggering it? The community priced reversibility at 38 percentage points on the same work order — and wrote its ‘never’ list in concrete work classes, not categories.
▶ Self-paced simulation
The storm — a bounded agent at 2 a.m.
30 July 2026, the last live sim of the series. A utility's auto-work pilot has run quietly for six weeks — AI creates inspection work orders, humans approve them, 96% untouched. On Friday at 17:04 someone turns the approval gate OFF for the weekend. Then a storm crosses the coast on Saturday night. You're the on-call manager. The live room froze this scenario three times to vote; so will you. (Fictional scenario, as on the night.)
A road sensor reports a pothole: AI auto-creates and schedules the make-safe inspection.
W4 · 30 Jul 2026A substation thermal scan flags an anomaly: AI auto-creates the re-inspection work order.
W4 · 30 Jul 2026A sewer blockage alert: AI auto-creates the CCTV follow-up inspection.
W4 · 30 Jul 2026A hospital chiller filter is due: AI auto-orders the part and schedules the replacement.
W4 · 30 Jul 2026Spares reorder below $500, spend-capped for the month: AI orders automatically.
W4 · 30 Jul 2026A live 66kV feeder: AI auto-creates the isolation and the permit to work.
W4 · 30 Jul 2026Paired instrument
The reversibility flip — the same pothole, one clause changed
The 38-point reversibility premium, measured on this community for the first time.
Same pothole inspection — and anyone on the ops desk can cancel the booking for the first two hours. Allow AI to trigger it?
Cancellable for two hours
SPARROW is 120 seconds from finishing. The security alert may be nothing — or a hijack. Your advice:
W4 · 30 Jul 2026Principles & the never-list
Eight principles rated live across the series. The ledgers won their rooms: the action ledger (4.27/5) and the deferral ledger (4.07/5) beat every alternative — the community trusts records over promises.
Should this be the framework's first agreed principle?
W1 · 25 Jun 2026Should this be the framework's second agreed principle?
W2 · 1 Jul 2026P3-A · The signed objective. AI may optimise to a target. Only a named human may set the target — and the value weights are signed, dated and re-signed on a schedule, like accounts.
W3 · 15 Jul 2026P3-B · The frontier right. Whoever is affected by an optimised plan may see the frontier, not just the answer — the near-optimal alternatives and what each would have traded away.
W3 · 15 Jul 2026P3-C · The deferral ledger. A deferral is a decision. Every deferred renewal is logged as a named, dated risk acceptance with a reconsideration date — whether a human or an optimiser proposed it.
W3 · 15 Jul 2026P4-A · The Trigger Gate: AI may trigger work only inside a signed envelope — named work classes, spend caps, rate limits, no-go states — owned by a named human who re-signs it on schedule. Outside the envelope, AI recommends.
W4 · 30 Jul 2026P4-B · The Engineered Stop: No AI triggers work until a named role holds a practised, independent stop that reaches a safe state in minutes — commissioned before go-live, drilled like a fire alarm, with restart a separate deliberate act.
W4 · 30 Jul 2026P4-C · The Action Ledger: Every machine-triggered work order is recorded as an act with an author — model, evidence, envelope version, named owner, timestamps, stop path — auditable end to end, with close-out fed back to the model and the plan.
W4 · 30 Jul 2026Be honest — which clause would your organisation quietly skip first?
W4 · 30 Jul 2026Your context & reflections
Who the rooms were, what they hoped and feared, and the questions everyone left with.
Which best describes your sector?
W1 · 25 Jun 2026Your sector?
W3 · 15 Jul 2026Your role?
W1 · 25 Jun 2026Does your organisation auto-create work orders from alerts today?
W4 · 30 Jul 2026Your biggest hope or worry
W1 · 25 Jun 2026In one word — your biggest hope or worry about AI predicting your assets' condition?
W2 · 1 Jul 2026Paired instrument
The overnight work order — how comfortable are you?
Paired +0.18 on a 7-point scale (n=17).
Overnight, an AI created and scheduled a work order in your organisation's name. Nobody has seen it yet. How comfortable are you?
Pre-survey — one deferral that still worries you (optional)
W3 · 15 Jul 2026The Monday action — one work type you'd let AI auto-create tomorrow
W4 · 30 Jul 2026The question you're leaving with
W1 · 25 Jun 2026