DomainsSelect domain Diagnostics Safety Global · misdiagnosis Medicare · Public Data Intelligence Level 4 status Long-horizon robotics policies Robotics benchmark Robotics value Robotics experiment Robotics policies Loop · closed loop Recursive improvement Level 4 run ROE · Reciprocal Optimization Engine Proof Metamorphic proof pack Result History Proof-to-Payment Engine End-to-end proof journey Customer opportunity Acquisition agent Five evidence conditions Proof contract Workbench execution Authoritative evidence Human disposition Conversion Engine status Founder Reference Flywheel Live telemetry Drift review Automated proof Live middleware Stream harness Auto proof Start a durable run Offer Pilot Marketplace Workbench New run Diagnostics Marketplace Peer proof · extremely small
Same controller, two workbenches A / Self()
│ same controller
▼
MCP
┌┴────┐
▼ ▼
B1 B2
│ │
z1,e1 z2,e2
└──┬──┘
▼
Φ / experience Measure a true cold Self() run on B1 first (opt-in, slower). The controller picks each step by asking B1 to score the options, with no memory of earlier runs, so Φ is measured against a cold final e. B2 is a fixed fixture with no Self() run, so it always uses the fast baseline. Use the MCP commercial case for B1 (separate starting plant, constructed for this path: completeness 0.80, contradiction 0.08, missing 0.40). Same sealed sequence and same greedy cold controller; only the starting plant differs. Prove the peer Not run. An error win needs a real cold run (see the sealed sequence rule: final e is the win, a tie is not a win). Safety Φ stays on /workbench/proof. No settled payment.
B1 MCP B2 identity Safety proof