Bring one real consequence.
Not a transformation programme. Not an AI strategy deck. One bounded workload where success, failure and evidence can be agreed before we start.
Propose a test →Good collaboration has a kill condition
Commercial pilot
Bring a costly or valuable machine workload. We define the current result, the target and what evidence would count. If the result does not survive the agreed test, it does not become a claim.
Research
Bring a falsifiable question, competing explanation, benchmark, replication or method. Negative findings are output, not resistance.
Physical / edge
Bring a machine, sensor path or retrofit problem where intelligence must cross a governed boundary before physical consequence.
Authority / risk
Bring a workflow where a machine can decide but must not create its own right to act, or where authorization and underwriting conditions must remain independent.
What we will not sell you
No claim of universal autonomy. No “AI said so” evidence. No mechanism disclosure masquerading as value. No pilot whose success criteria are invented after the run.
The useful object is a verified change in quality, cost, latency, throughput, risk or control — with the boundary visible.