Inspectable software
Executable code, documented inputs, testable behavior and clear failure cases.
SELECTED ENGINEERING WORK · 01 / 03
A tenant needs help.
Can AI prepare the work safely?
A small, inspectable case about what AI can prepare, what a teammate still decides, and why missing evidence should stop a workflow.
Try the 60-second caseFictional scenarioNo external contactHuman approval · Tested gate
Tenant operations · new case
Heating failure · Weserstr. 18
A tenant says the heating is completely cold while hot water still works.
01 · What came in
Tenant message
The same inputs an operator would receive.
Supporting records · 3 (optional to inspect)
02 · Bounded worker
Prepare, then stop for review
The worker may classify, resolve and draft. Release remains a separate authority decision.
GoalPrepare review-ready work without contacting the tenant or contractor.
✋
Human approval is a real gatePreparation and release are separate steps.
classify→resolve property→match trade→prepare
Prepared work
Human review requiredReady for review
Not released. The worker prepared everything but took no external action.
Review complete?Only this gate can release the prepared action.
Approved for release — simulated. In this public synthetic demo no external system is connected, so no real message is sent.
What we'd measure in a pilothandling time · corrections · approval rate · escalations · false completions
Trust testRemove the property record. Should the worker continue?
WHY THIS MATTERS · FOR THE PERSON DOING THE WORK
A prepared hand-off is useful only if the teammate can check the evidence, correct mistakes and stay in control.
BEFORE · MANUAL PREPARATIONFind the right records.
Read a tenant request, identify the property, check the relevant trade and prepare a response.
WITH AI SUPPORT · DEMOReview prepared work.
See the proposed classification, linked source records and a draft hand-off in one place.
HUMAN RESPONSIBILITYDecide what happens next.
If evidence is missing, the workflow stops. A person checks and approves consequential actions.
Fictional case. No measured time savings, live contractor integration or automatic outreach. See the tested gateway boundary ↗
Earned autonomy
Why not approve forever after enough clean cases?
That can be the right direction — but autonomy should be promoted by evidence and scope, not by a magic click count. Repeated approvals are a signal; outcome quality and failure behaviour are the gate.
Shadow prepare only; every case reviewed
Copilot human approves every consequential action
Limited auto auto-release only low-risk, well-bounded cases
Trusted broader authority only after sustained evidence
Illustrative promotion evidence: 100+ reviewed cases · near-zero material corrections · 0 unsafe/false-completion events · stable source coverage · explicit rollback and monitoring. Promotion still requires an intentional policy change.
Want to inspect the engineering? +
The demo uses synthetic fixtures. The underlying Factory implements explicit capabilities, evidence requirements, human approval and replayable tests. The approval above is simulated because this public page is intentionally disconnected from external systems.
Verified safety / retry release gate ↗ · Read executable test ↗ · Full repository ↗ENGINEERING, WITHOUT THE MAGIC TRICK
What is real. What stays human.
People stay responsible
No external action or consequential decision is silently taken by this public demonstration.
No production claims
Fictional examples are not evidence of customer adoption, saved hours or deployment readiness.