# Validation record
This is the original 2026-10-29 qualification record. The subsequent workspace
example suite and its shared core operations are covered by
[the 2026-11-07 example validation record](examples-validation.md).
Actual GPU runtime qualification subsequently became possible outside the
sandbox; see [CPU/GPU backend validation](backend-validation.md) for current results.
Validated locally on 2026-10-29 against the working tree, Burn **0.22.1** and
the repository's `Cargo.lock`. This document records implementation
qualification, not a published release, remote CI result, or quantum advantage.
## Environment
- Linux x86_64; Intel Core i9-12900K.
- Rust 1.95.0 (`59807616e`, 2026-04-14), Cargo 1.95.0.
- Flex CPU, explicit F64 reference tests and F32 examples.
- Local dependencies were available offline. In the execution environment the
snap Rust shim was unavailable, so the standalone toolchain's `bin` directory
was prepended to PATH. No nightly or local Burn patches were required.
Reproduce the main checks with `bash scripts/validate.sh`. Raw logs are written
to `artifacts/validation/` and intentionally excluded from the source package.
Use `CARGO_NET_OFFLINE=true` when dependencies are already cached.
## Passed checks
| `cargo fmt --all --check` | Pass |
| `cargo clippy --locked --all-targets -- -D warnings` | Pass |
| `cargo test --locked --all-targets` | 26 integration tests pass; example targets compile |
| `cargo test --locked --doc` | 1 executable documentation test passes |
| `cargo check --locked --no-default-features` | Pass, no execution backend required to compile |
| `cargo check --locked --no-default-features --features flex` | Pass, inference without the autodiff feature |
| `cargo check --locked --features wgpu,fusion --all-targets` | Pass |
| `cargo check --locked --all-features --all-targets` | Pass, including CUDA and WGPU |
| `RUSTDOCFLAGS='-D warnings' cargo doc --locked --no-deps` | Pass |
| `bell`, `qudit`, `train`, `hybrid` examples | Executed successfully, including convergence assertions |
| Release `profile` example | Executed at 4, 8, 12 qubits, one warmup and seven timings each |
| `cargo package --locked --allow-dirty` | Source archive built and independently compiled; generated logs excluded |
The test suite covers multiple cases per test, including:
- Bell/GHZ states at 2/3/5 qubits; maximally entangled qudits at dimensions 3/4/5.
- Independent scalar complex arithmetic against mixed-radix tensor application,
reversed and nonadjacent targets, two distinct batch entries, and gate inverses.
- Generalized Weyl relations, level-controlled Y gates, Pauli-Y coherences,
Hamiltonian sums, variances, marginals and seeded measurement statistics.
- X/Y/Z rotation gradients at dimensions 2/3/4 against finite differences and
appropriate parameter shifts, including spectator-level coherences.
- Shared affine parameters, per-example input gradients, and balanced Burn
checkpointing; incompatible complex-component execution contexts are rejected.
- Pure-to-density gradients through amplitude damping, Kraus completeness,
mixed-state positivity checks, complex reduced states, noise endpoints and
arbitrary qutrit dephasing.
- Burn `Config` serialization, burnpack weight-record round trips, module
device movement and actual Adam updates.
- Invalid matrix/IR invariants, versions, dimensions, duplicate wires, tensor
contexts, nonphysical values, overflow and resource-limit rejection.
## Executed prototype outcomes
| Bell state | Probabilities approximately `[0.5, 0, 0, 0.5]`; only 00/11 sampled |
| Qutrit Bell state | Equal 00/11/22 probabilities; reduced purity approximately 1/3 |
| One-angle variational optimization | Initial loss 3.8233407; final Z expectation -0.997102 after 100 Adam steps |
| Classical encoder + entangled circuit + classical head | Regression MSE 0.1360664 -> 0.0001202 after 120 Adam steps |
Both training problems are deterministic integration demonstrations. They do
not establish useful generalization, quantum advantage, or trainability at
large depth.
## Small release-mode timing probe
Flex CPU, F32, batch 1, H on each wire followed by a nearest-neighbor CNOT chain.
Each timed execution includes zero-state creation, gather planning, gate
execution, probability readback and normalization validation. The state-byte
column is a storage lower bound, **not measured peak memory**.
| 4 | 16 | 7 | 128 | 0.034 | 0.032 | 0.048 |
| 8 | 256 | 15 | 2,048 | 0.180 | 0.176 | 0.283 |
| 12 | 4,096 | 23 | 32,768 | 4.065 | 4.019 | 4.124 |
These small bounded measurements establish that the prototype runs and provide
a baseline for future prepared-plan/kernel work. They do not characterize
large-state throughput, training memory, tensor-network scaling, or GPU speed.
## Remaining qualification
- GPU runtime and numerical parity are **unverified**. `nvidia-smi` could not
communicate with the NVIDIA driver and `/dev/dri` was unavailable in this
execution environment. Feature compilation is separate from hardware proof.
The optional `wgpu_smoke` example checks a forward value and its analytic
derivative when run on an available adapter.
- CUDA runtime, browser/Wasm execution, distributed simulation and remote QPUs
have not been tested. No hardware fallback is implied by this record.
- CI is configured in `.github/workflows/ci.yml`; no remote run was triggered.
- This working tree has not been committed, pushed or published.