Skip to content

Glossary

Artifact manifest

The TSV inventory recording each public release path, byte size, and SHA-256 digest.

Archived replay

Re-execution of already collected generated programs under the named evaluator. It does not call the model provider again.

Assembly admissibility

The operational state (A=1): an executable selected circuit serializes successfully to OpenQASM 3 under the evaluator. It does not denote execution of the assembly on a simulator or hardware backend.

AS-Gap

Assembly-Structure Gap. The aggregate contrast (\bar A-\bar M^{\mathrm{sig}}). On the frozen release, (M^{\mathrm{sig}}\preceq A), so it is the assembly-admissible subset of the ES-Gap and counts (A=1,M^{\mathrm{sig}}=0) cells.

Benchmark record

A source/governance row containing a stable row ID, instruction, and optional reference program and release bucket.

Benchmark release

The versioned prompt, split, target, and frozen scientific contract. It is independent of package and evaluator versions.

Candidate

One or more supplied model rows being summarized or compared with the frozen benchmark.

Canonical evaluation

One normalized scored model-prompt cell satisfying the package record and endpoint invariants.

Canonical response

The normalized provider/model response retained for one prompt, including generated text/code or a normalized error.

Cell

One (model, prompt_id) observation.

Classical-bit count

The number of classical bits in the evaluated circuit, one component of the reference signature.

Complete matrix

The rectangular 21-model by 154-prompt frozen primary evaluation, containing 3,234 cells.

Container artifact

The separately versioned image or archive used for isolated executable replay. It encapsulates an evaluator contract but is not itself the evaluator version.

Container image ID

The local content-addressed Docker image identity. It is distinct from the archive SHA-256 and from an OCI registry manifest digest.

Count map

A mapping from evaluator-visible operation type to multiplicity. It records more information than the set of observed gate types.

Denominator alignment

The requirement that candidate and comparator rates use the same prompt set.

Deployment variation

Outcome variation caused by changing model services, routes, provider infrastructure, policies, or serving systems.

ERSD

Executable reference-signature disagreement rate. ES-Gap count divided by execution count on the same cell set.

ES-Gap

Execution-Structure Gap. The aggregate rate of cells that execute but fail the frozen reference-signature predicate.

Evaluation record

A canonical scored cell containing model, prompt, execution, signature, optional assembly admissibility, and optional stricter reconstruction endpoints.

Evaluator

The versioned implementation that executes generated code, extracts circuit descriptors, and computes evaluation endpoints.

Executable output

Generated code that runs successfully under the named evaluator.

Evidence bundle

The complete frozen public artifact directory containing records, analyses, scripts, documentation, and the internal manifest. Manuscript-facing publication derivatives are excluded.

Frozen

Declared immutable for the named benchmark release.

Gate vocabulary

The set of operation types present, without multiplicities. It must not be used interchangeably with the complete operation-type count map.

Identifiable prompt

A prompt whose wording uniquely specifies all components compared by the frozen signature predicate.

Matched-subset comparison

An explicit partial comparison in which frozen rows are restricted to the candidate's exact prompt IDs.

Model route

The evaluated combination of requested model, provider, and concrete serving route represented by a canonical model label.

Live replication

A newly identified, explicitly authorized model-provider run produced by run-model. It is new stochastic evidence, not historical reproduction.

Operation count

The scalar number of counted operations. Under the frozen predicate it follows from complete count-map equality and is reported separately as a diagnostic.

Ordered operation-and-operand tape

The sequence of operation names together with their quantum and classical operands, compared in order.

Parameter-aware match

An ordered match that additionally recovers the evaluator-normalized parameter-value sequence.

Predicate

The versioned rule determining whether an output passes a reconstruction endpoint.

Prompt ID

The stable identifier joining prompt, response, evaluation, and provider attempt records.

Provider attempt

One normalized request attempt, including route, model identities, timestamps, status, hashes, usage, and error/transport metadata.

QASM3 export

Successful export of an evaluated circuit to OpenQASM 3. It is a diagnostic endpoint and is not equivalent to reference reconstruction.

Qubit count

The number of qubits in the evaluated circuit, one component of the reference signature.

Reference signature

The frozen tuple operationalized by qubit count, classical-bit count, and the complete operation-type count map.

Replay

Execution of archived or newly collected generated code. In package version 1.2.1 it occurs only inside the isolated Docker evaluator.

Replication

A new experiment, generally involving new model/provider calls and new run identity.

Reproduction

Deterministic reconstruction of published quantities from frozen evidence.

Resolved model

The model identity reported or resolved by a provider, which may differ from the requested alias.

Run manifest

The seven-field version and run-type record attached to an experiment or comparison.

Schema

The machine-readable JSON record contract. Schema version is independent of benchmark and evaluator versions.

Semantic equivalence

Task-level or physical equivalence of circuit behavior. It is not established by the frozen signature, ordered, or parameter-aware predicates.

SHR

Structural-hallucination rate. The conditional frequency of (E=1,M=0) among executable cells on the identifiable prompt subset.

Structural hallucination

On an identifiable prompt, an executable output that fails the frozen reference-signature predicate.

Structural predicate

The explicitly limited reference-reconstruction rule used for signature matching.

Target signature

The frozen reference signature associated with a prompt's audited target circuit.

Transport affected

A provider-attempt flag indicating that network, service, quota, refusal, or another transport/deployment condition affected observation.

Trust boundary

The separation among local data processing, third-party network collection, and generated-code execution.

Version dimension

One independently recorded identity: package, benchmark, evaluator, predicate, schema, artifact manifest, or run type.