Differentiable Reviewer Evidence

This page maps bounded reviewer criticisms and the evidence package package to real repository commands, artefacts, and explicit public gaps. It is an index into executable evidence, not a replacement for the tests or artefacts.

Regenerate and validate it with:

python tools/differentiable_reviewer_evidence_page.py --write
python tools/differentiable_reviewer_evidence_page.py --check

Evidence boundary

Generated reviewer-response index over repository commands and artefacts. Implemented and bounded rows identify only their declared local evidence; open rows remain unpromoted, and no provider, hardware, universal-framework, compiler, isolated-performance, or superiority claim is implied.

Support-matrix synchronization

The generated Differentiable Support Matrix supplies the current registry and planner summary:

Generated surface	Current result
Program AD registry	118/118 complete
Quantum-gradient planner audit	5 supported / 5 fail-closed

Reviewer criticisms

ID	Criticism	Status	Evidence and reproduction	Open gap	Boundary
reviewer-criticism-parameter-shift-implementation	No parameter-shift implementation	implemented	Scalar, batched, generalised finite-spectrum, registered Phase-QNode, and finite-shot parameter-shift contracts have executable tests. Commands: python -m pytest -q tests/test_differentiable_parameter_shift.py tests/test_phase_generalised_parameter_shift.py Paths: src/scpn_quantum_control/differentiable_parameter_shift.py, src/scpn_quantum_control/phase/generalised_parameter_shift.py, tests/test_differentiable_parameter_shift.py, tests/test_phase_generalised_parameter_shift.py, docs/quantum_gradients.md	—	Implemented for declared local and bounded callback routes; no universal gate, provider, or hardware-gradient claim is implied.
reviewer-criticism-quantum-gradient-surface	No quantum-gradient surface	bounded	The executable planner and Phase-QNode gradient backend cover supported statevector, finite-shot callback, and host-bridge cases while unsafe cells fail closed. Commands: python -m pytest -q tests/test_phase_gradient_backend.py tests/test_phase_gradient_support_matrix.py Paths: src/scpn_quantum_control/phase/gradient_backend.py, src/scpn_quantum_control/phase/gradient_support_matrix.py, tests/test_phase_gradient_backend.py, tests/test_phase_gradient_support_matrix.py, docs/differentiable_support_matrix.md	DIFF-OPEN-02	Local and caller-supplied callback evidence does not promote live provider execution or unrestricted hardware gradients.
reviewer-criticism-changelog-autodiff-evidence	No autodiff evidence in the changelog	implemented	The changelog records parameter-shift, finite-shot, JAX, PyTorch, TensorFlow, Program AD, and compiler-AD additions after implementation. Commands: python tools/differentiable_reviewer_evidence_page.py --check Paths: CHANGELOG.md, tools/differentiable_reviewer_evidence_page.py, tests/test_differentiable_reviewer_evidence_page.py	—	Historical entries document landed bounded surfaces; they do not upgrade their recorded provider, hardware, or performance boundaries.
reviewer-criticism-vqe-gradient-optimisation	VQE uses classical or gradient-free optimisation only	implemented	PhaseVQE exposes parameter-shift gradients, gradient descent, natural-gradient paths, and convergence evidence alongside declared classical baselines. Commands: python -m pytest -q tests/test_phase_vqe.py tests/test_phase_gradient_descent.py tests/test_phase_natural_gradient.py Paths: src/scpn_quantum_control/phase/phase_vqe.py, src/scpn_quantum_control/phase/gradient_descent.py, src/scpn_quantum_control/phase/natural_gradient.py, tests/test_phase_vqe.py, tests/test_phase_gradient_descent.py, tests/test_phase_natural_gradient.py, docs/quantum_gradients.md	—	Convergence evidence is bounded to declared local objectives and does not establish global optimiser superiority.
reviewer-criticism-ml-framework-integration	No ML-framework integration	bounded	JAX, PyTorch, TensorFlow, PennyLane, and Qiskit rows are exercised through the public framework bridge/parity contracts with explicit optional-dependency and unsupported-route classifications. Commands: python -m pytest -q tests/test_phase_qnn_framework_bridge_matrix.py tests/test_phase_qnode_framework_parity.py Paths: src/scpn_quantum_control/phase/qnn_framework_bridge_matrix.py, src/scpn_quantum_control/phase/qnode_framework_parity.py, tests/test_phase_qnn_framework_bridge_matrix.py, tests/test_phase_qnode_framework_parity.py, docs/differentiable_programming.md	DIFF-OPEN-01, DIFF-OPEN-02	Installed local and compatibility routes do not imply arbitrary-simulator, accelerator, provider, or hardware framework autodiff.

Evidence package

ID	Requirement	Status	Evidence and reproduction	Open gap	Boundary
evidence-package-gradient-correctness-runtime	Finite-difference versus parameter-shift correctness and runtime	bounded	The differentiable benchmark suite records analytic agreement and local functional timing with explicit non-isolated classification. Commands: python -m pytest -q tests/test_differentiable_programming_benchmarks.py Paths: src/scpn_quantum_control/benchmarks/differentiable_programming.py, tests/test_differentiable_programming_benchmarks.py, data/differentiable_phase_qnode/local_benchmark_20260616T0955Z/diff-qnode-ci-evidence-schema-v1.json	DIFF-OPEN-03	Local functional timing is not isolated performance evidence and supports no speedup claim.
evidence-package-batched-shifted-circuits	Batched shifted-circuit execution	implemented	Batched parameter-shift helpers and benchmark rows validate values, gradients, evaluation counts, masks, and malformed batch contracts. Commands: python -m pytest -q tests/test_differentiable_batch_helpers.py tests/test_differentiable_parameter_shift.py Paths: src/scpn_quantum_control/differentiable_batch_helpers.py, src/scpn_quantum_control/differentiable_parameter_shift.py, tests/test_differentiable_batch_helpers.py, tests/test_differentiable_parameter_shift.py	—	Batched local evaluation does not imply provider-side batching, QPU execution, or performance promotion.
evidence-package-optimiser-convergence-comparison	VQE and QNN optimiser convergence comparison	bounded	Deterministic multi-start and QNN optimiser suites compare parameter-shift, finite-difference, natural-gradient, Adam, L-BFGS-B, SPSA, and declared derivative-free routes. Commands: python -m pytest -q tests/test_phase_qnn_optimizer_benchmark.py tests/test_phase_vqe.py Paths: src/scpn_quantum_control/phase/qnn_optimizer_benchmark.py, src/scpn_quantum_control/phase/phase_vqe.py, tests/test_phase_qnn_optimizer_benchmark.py, tests/test_phase_vqe.py, data/differentiable_phase_qnode/ground_state_optimizer_convergence_20260709.json	DIFF-OPEN-03	Bounded local convergence comparisons do not prove general optimiser or application superiority.
evidence-package-framework-adapter-overhead	Framework adapter overhead benchmarks	open	Functional framework-overlay evidence exists, but promotable overhead ratios still require compatible accelerator and isolated-affinity artefacts. Commands: python -m pytest -q tests/test_differentiable_framework_overlay.py tests/test_differentiable_benchmark_workflow.py Paths: src/scpn_quantum_control/differentiable_framework_overlay.py, tests/test_differentiable_framework_overlay.py, tests/test_differentiable_benchmark_workflow.py, data/differentiable_phase_qnode/ml350_framework_overlay_baseline_20260705/diff-qnode-external-comparison.json	DIFF-OPEN-01, DIFF-OPEN-03	Functional overlay rows are not isolated overhead or accelerator performance claims.
evidence-package-compiler-ad-evidence	Compiler-backed AD benchmark evidence	bounded	Native whole-program and Enzyme/MLIR evidence rows are executable where the toolchain is present and remain non-promotional without isolated artefact IDs. Commands: python -m pytest -q tests/test_native_whole_program_ad_execution_evidence.py tests/test_enzyme_toolchain_execution_evidence.py tests/test_phase_qnode_compiler_lowering.py Paths: src/scpn_quantum_control/compiler/mlir_enzyme_evidence.py, tests/test_native_whole_program_ad_execution_evidence.py, tests/test_enzyme_toolchain_execution_evidence.py, tests/test_phase_qnode_compiler_lowering.py, data/differentiable_phase_qnode/compiler_evidence_boundary_20260705.json	DIFF-OPEN-03, DIFF-OPEN-04	Toolchain execution rows do not promote broad compiler AD, Rust registry execution, LLVM/JIT, or performance claims.
evidence-package-noisy-gradient-shot-allocation	Noisy-gradient and shot-allocation evidence	bounded	Finite-shot parameter-shift, seeded SPSA, score-function, confidence-interval, shot-allocation, and failure-policy contracts have executable tests. Commands: python -m pytest -q tests/test_stochastic_gradient_failure_policy.py tests/test_stochastic_gradient_rust_parity.py tests/test_phase_qnn_optimizer_benchmark.py Paths: src/scpn_quantum_control/differentiable_stochastic_estimators.py, src/scpn_quantum_control/differentiable_stochastic_policy.py, tests/test_stochastic_gradient_failure_policy.py, tests/test_stochastic_gradient_rust_parity.py, tests/test_phase_qnn_optimizer_benchmark.py, docs/quantum_gradients.md	DIFF-OPEN-02, DIFF-OPEN-03	Seeded simulator and callback uncertainty evidence does not imply calibrated hardware robustness or performance promotion.
evidence-package-reviewer-evidence-synchronisation	Reviewer evidence page and synchronization gate	implemented	This generated page validates all row commands, paths, changelog markers, public open gaps, support-matrix counts, and capability-manifest surfaces. Commands: python tools/differentiable_support_matrix_page.py --check python tools/differentiable_reviewer_evidence_page.py --check Paths: tools/differentiable_support_matrix_page.py, tools/differentiable_reviewer_evidence_page.py, tests/test_differentiable_support_matrix_page.py, tests/test_differentiable_reviewer_evidence_page.py, docs/differentiable_support_matrix.md, docs/differentiable_reviewer_evidence.md	—	The page inventories evidence and gaps; it does not itself prove numerical correctness, provider execution, or benchmark promotion.

Public open-gap pointers

These identifiers resolve to the open evidence work section of the public roadmap. A row carrying one remains bounded or open even when its local command passes.

Pointer	Gap	Evidence still required
DIFF-OPEN-01	Framework-native and accelerator promotion	Broader native framework transforms, incompatible accelerator routes, and cross-runtime persistence require dedicated artefacts.
DIFF-OPEN-02	Provider and hardware gradient evidence	Live-ticket provider jobs, raw counts, calibration, and simulator comparison remain approval-gated.
DIFF-OPEN-03	Isolated benchmark promotion	Performance claims require reserved-host isolated-affinity artefacts and validated comparison metadata.
DIFF-OPEN-04	Compiler AD promotion	Broader Enzyme, MLIR, LLVM/JIT, and executable registry promotion requires native evidence plus isolated benchmark identifiers.
DIFF-OPEN-05	Tutorial and notebook breadth	Public notebooks beyond the executable examples remain open until clean-environment replay and expected outputs are recorded.

Reading the rows

• implemented means the declared bounded contract has a runnable evidence route.
• bounded means local evidence exists but a named promotion gap remains open.
• open means existing artefacts are diagnostic or functional only and cannot support the requested promoted claim.
• Commands are deliberately module-scoped; repository-wide local test runs remain outside this reproduction page.