Just Because a Robot Says It Can Doesn't Mean It Can Right Now: Checking an AI-Chosen Robot Task Against the Robot's Real Condition, Safety State, Timing and What Actually Happened

AI planners and agents increasingly select robot skills from machine-readable descriptions — “pick and place, 8 kg payload, ±0.1 mm repeatability” — and dispatch them to physical machines. A skill that a controller advertises, however, says nothing about whether this robot, with its current tool, payload, calibration, temperature, wear and safety state, can perform it now, within the required deadline, and with the intended physical result. Attested firmware, an accepted command or a reported protective stopare equally weak as evidence of physical capability, actuation or zero motion. This disclosure describes a vendor-neutral, non-actuating method for qualifying a robot skill against evidence. Each skill is bound to an issuer and revision and to a generation-qualified robot instance (hardware, controller, tool and sensor generations). Declared capability is kept separate from measured capability, and task requirements in payload, reach, torque, speed and precision are tested against calibrated limits after subtracting measurement uncertainty, with calibration generation, freshness, scope, metrological traceability and coordinate-frame transforms checked explicitly. The state of an independent safety controller is read, never overridden, and its freshness is verified. Command admission requires an end-to-end response bound within the deadline, and dispatch, controller acceptance, actuator application and physical completion are recorded as separate evidence. During execution the method supports monotone capability downgrade with residual-work requalification, and on cross-robot transfer it refuses to reuse source calibration or safety state without target-specific evidence. Predicates are combined with three-valued logic under an explicit objective: a proven contradiction yields INCONSISTENT, complete evidence yields CONSISTENT, and missing evidence yields UNRESOLVED. UNRESOLVED never implies zero motion; the physical response remains with the separately validatedsafety architecture. The document provides a 48-stage, six-phase engineering workflow with flowcharts, formal definitions, six detailed embodiments, 70 formal relations, 40 engineering checkpoints, a comparison with existing approaches (ROS 2 and MoveIt, OPC UA Robotics and Industrie 4.0 capability-skill models, ISO 9283 performance testing, ISO/IEC 17025 and GUM metrology, ISO 10218 and ISO 13849 functional safety, IETF RATS attestation, real-time schedulability analysis and TSN, and fleet task interfaces such as VDA 5050) and a non-actuating reference evaluator in Python, JavaScript and Go with 79 shared synthetic test vectors, all passing. It is not a safety certification and makes no SIL or PL claims; the laboratory is synthetic, self-authored and not peer reviewed. The author is open to acquihire or strategic collaboration. Keywords: AI safety; robotics; robot skills; capability qualification; calibration; measurement uncertainty; metrological traceability; coordinate transforms; functional safety; independent safety controller; protective stop; real-time deadlines; worst-case response time; command admission; physical-effect verification; capability derating; skill transfer; three-valued logic; ROS 2; OPC UA