Quantum Targets

Wheeler artifacts carry semantic quantum regions. A target plans and performs those regions using a particular simulator, chamber, or provider. The adapter does not define source-language meaning.

Descriptor and submission

A TargetDescriptor publishes adapter and target identity, independently negotiated capabilities, maximum logical qubits, and maximum shots.

Capabilities include static circuits, parameter binding, batches, mid-circuit measurement, reset, classical conditions, state-vector diagnostics, logical qubits, and network entanglement. A target advertises only work it can execute.

QuantumSubmission binds the verified artifact, logical register, basis preparation, ordered circuit or adjoint applications, shot count, and seed policy.

Every QuantumJob is asynchronous, including a local job that finishes immediately. It carries identity and lifecycle, accepts a cancellation request, and returns a QuantumResult. Results use canonical little-endian outcome integers and name their target.

Before classical state changes, the runtime verifies job, target, and complete task identity. recover(jobId, task) reconciles acknowledged work without another submission. An unknown or mismatched provider identity stops recovery.

Submission moves classical circuit descriptions and measured observations. Coherent quantum state never becomes an I/O buffer.

Ideal state-vector target

StateVectorTarget is the semantic reference for static gates. It supports at most 20 qubits and repeats preparation, unitary work, and measurement for every shot.

One explicit seed supplies one pseudorandom stream in shot order. The target does not reseed between shots. A seed makes the simulator sample repeatable. It carries no promise about hardware.

The ideal engine supports H, X, Z, phase, controlled phase, CNOT, CZ, swap, generated adjoints, coherently lifted XOR, and width-modular constant addition or subtraction.

The simulator may disclose its numerical state vector. That disclosure belongs to the ideal model and is unavailable from ordinary hardware measurement.

Target-resident dynamic work

DynamicStateVectorTarget separately advertises mid-circuit measurement, reset, and classical conditions. Its semantic regions may:

1. prepare a register.
2. apply fixed gates.
3. measure one qubit into a target-resident Boolean slot.
4. reset the measured qubit.
5. apply X or Z when an earlier slot matches.

A measured qubit cannot enter another gate or measurement before reset. The complete region executes in one target call. Host code receives an observation only through an explicit final full-register measurement.

The maintained teleportation fixture uses two measured slots with conditional X and Z corrections for either basis input. The syndrome fixture performs at most 1,024 parity-measurement rounds in one call. Their results retain final basis, slots, syndrome, reset, and correction evidence. These cases establish semantic conformance, rather than a hardware noise model.

A static target rejects a dynamic region before submission. Wheeler never moves a latency-sensitive branch to the host in silence.

Batches and parameter bindings

A QuantumBatch is an ordered content-identified list of complete tasks. Member jobs may finish in any order. The batch result preserves semantic task order. One overall timeout applies. Job and task identity drift causes rejection.

A target without BATCH_SUBMISSION rejects the complete batch before the first submission.

A symbolic gate stores a stable parameter name and finite scale in .wbc. The task supplies one exact finite binding map. Missing and extra names fail before submission and the map enters task identity.

Generated adjoints negate symbolic scale. The ideal target evaluates the binding directly. OpenQASM lowering emits the corresponding numeric angle.

QuantumResult.zExpectation(...) estimates a tensor product of Pauli-Z observables from little-endian samples and returns estimate, standard error, and shot count.

Maintained planning protocols

Several higher-level protocols use the same target identity rules:

• RecoverableOptimizerCampaign executes 1 through 64 iterations, each containing

1 through 64 bound submissions. It persists acknowledged member identities and never applies one submission identity to the objective twice.

• CalibrationAwareCompiler requests calibration for 1 through 64 direct gates.

Its plan binds target descriptor, request, result, epoch policy, duration, and additive parts-per-trillion error ceiling. The plan is evidence of selection, rather than pulse lowering or fidelity.

• DelegatedComputationSession maintains one masked-NOT transcript under the

named honest-but-curious single-provider model. It makes no malicious-provider, collusion, side-channel, transport, or general privacy theorem.

• DistributedEntanglementSession requires NETWORK_ENTANGLEMENT, binds two

endpoints and a deadline, and persists no provider or qubit handles. Restoration sends no second request. A local discard makes no statement about remote physical destruction.

• LogicalResourcePlan keeps Clifford gates, T gates, measurements, T-depth,

magic states, factory batches, logical qubits, target cycles, code distance, and modeled error as separate dimensions. It provides planning evidence, rather than a proof of decoder throughput or physical performance.

These protocols never promote target metadata into a stronger physical statement.

Provider-neutral quantum instructions

.wbc carries stable semantic gate descriptors and regular quantum instruction records. Gate forms name controls, targets, and angle parameters in a fixed order. Targets may decompose them into a native basis and advertise stricter limits.

No physical gate basis belongs to every provider. Wheeler standardizes semantic operations instead of appliance opcodes. Preparation, measurement, reset, and conditional gates have distinct instruction families. Provider payloads cannot extend canonical semantics at runtime.

OpenQASM 3

OpenQasm3Emitter lowers one supported static task into a complete program:

OPENQASM 3.0;
include "stdgates.inc";
bit[2] c;
qubit[2] q;
h q[0];
cx q[0], q[1];
c = measure q;

OpenQASM is a derived target format. wheeler qasm emits it from an artifact that contains one static submission.

OpenQasmTarget accepts an application-provided executor:

OpenQasmTarget target = new OpenQasmTarget(
  "provider-name",
  127,
  10_000,
  (qasm, shots, seed) -> provider.submit(qasm, shots));

The executor may call a provider API, appliance SDK, local engine, or queue. It must return one in-range full-register outcome for every shot. Credentials and provider objects stay with the host and never enter .wbc, QASM, result records, or logs.

Live hardware authority

Live tests require an invocation-local LiveHardwareTestPolicy. The default is disabled. An enabled policy names hard submission and aggregate-shot ceilings, and the wrapper charges both before contacting a provider.

Ordinary acceptance work creates no enabled policy. A separately authorized smoke run may supply credentials and budget through its host environment. Its result is sampled hardware evidence, rather than a deterministic acceptance result or proof.

Physical limits

A generated adjoint is another physical computation. It cannot consume VM history. After measurement or loss of a target session, Wheeler may replay an accepted observation or prepare fresh state and retry. It cannot reconstruct an unknown earlier hardware state.

Dynamic and fault-tolerant features remain explicit target requirements. Target planning rejects missing capabilities before queue submission.

The hybrid-run appendix follows acknowledged jobs and accepted observations. Weather preserves one chamber account in which those distinctions became operational.