Core Concepts¶
Package, Benchmark, And Evidence Bundle¶
pqid-bench is the installable interface. PQID-Bench v1.0.0 is the frozen
benchmark release. The evidence bundle contains the prompts, archived
responses, evaluations, robustness analyses, repeatability records,
regeneration scripts, and reproducibility documentation. Unpublished
manuscript source and manuscript-facing publication derivatives are not part
of the public evidence bundle.
These are related but not interchangeable. The small Python wheel does not embed the evidence bundle.
Six Operations¶
Verification¶
Verification checks whether the release files have their declared byte sizes and SHA-256 digests. Full verification additionally checks frozen scientific counts and repeatability dimensions.
Reproduction¶
Reproduction recomputes the published summary from archived evaluation records. It does not run generated programs and does not contact providers.
Evaluation Summarization¶
Evaluation summarization computes the same metric family from a supplied canonical evaluation JSONL file. The file must already contain scored model-prompt cells.
Candidate Comparison¶
Comparison aligns a supplied evaluation bundle with the frozen benchmark and reports candidate-minus-frozen deltas. The default comparison requires the complete 154-prompt test set for every candidate model. Partial comparison is explicit and uses the same prompt subset on both sides.
Live Replication¶
Live replication exports frozen model-facing prompt messages to one declared OpenAI-compatible model route and records a new response panel, attempt history, raw-payload digests, and run manifest. It contacts a provider but does not execute generated code.
Executable Replay¶
Replay executes archived or newly collected generated Python in a credential-free Docker worker. It is the only package command that executes generated code.
Reproduction Versus Replication¶
Reproduction asks whether the frozen results can be reconstructed from the published evidence.
Replication asks whether new calls to models, providers, or routes produce similar results. Fresh generation can differ because models, aliases, decoding systems, provider infrastructure, and policies change. It therefore requires a new run identity. Package version 1.2.1 implements that replication path without treating its outputs as frozen-result reproduction.
The Frozen Primary Matrix¶
The primary matrix contains:
- 21 completed model routes;
- 154 held-out prompts;
- 144 evaluator-facing reference signatures; and
- 3,234 model-prompt cells.
The matrix is rectangular: each completed model row contains one canonical cell for every frozen prompt.
Execution And Reference Reconstruction¶
Execution means that the generated program runs in the frozen evaluator environment.
Reference-signature recovery means that an executable output matches the frozen target's:
- qubit count;
- classical-bit count; and
- complete evaluator-visible operation-type count map.
Scalar operation-count agreement follows from count-map equality under the frozen convention and is reported as a separate diagnostic.
Ordered operation-and-operand equality is stricter than signature recovery. Parameter-aware ordered equality is stricter again. None of these predicates proves unitary, measurement-distribution, physical, or semantic equivalence.
Identifiability¶
Four frozen prompts do not uniquely determine every exact component of their
stored reference signature. They remain in the primary 154-prompt
stress-inclusive matrix. A prespecified 150-prompt sensitivity subset excludes
them when assigning the cell-level term structural hallucination.
This distinction prevents prompt underspecification from being attributed automatically to the model.
Independent Version Dimensions¶
Every core machine-readable scientific report carries:
- package version;
- benchmark release;
- evaluator version;
- structural-predicate version;
- schema version;
- artifact-manifest version; and
- run type.
Executable replay adds a seventh distribution dimension:
- container-artifact version, together with its image identity or published OCI registry digest.
Container artifact 1.0.0 encapsulates evaluator
pqid-bench-evaluator-1.1.0-safe-builtins. Those values need not have the same
version number because one identifies a distribution object and the other
identifies the scientific execution contract. Before OCI publication, the
local image ID and container-archive SHA-256 are recorded separately and no
registry digest is claimed.
The evaluator may change without changing the prompt split. A schema may change without changing the structural predicate. Keeping the identifiers separate makes those changes auditable.
Trust Boundary¶
The ordinary offline commands treat archived evaluation records as data.
run-model crosses a separate network, retention, and billing boundary and
requires prompt-export acknowledgement. Generated Python is outside both
boundaries. Replay therefore requires a second acknowledgement and delegates
execution to Docker with networking disabled, read-only scientific inputs,
dropped capabilities, resource limits, and no credentials.