# Phase 3 verification record Prepared 24 September 2026. Implementation 3.0.0; scientific protocol 2.1/A01. ## Delivered status The package contains the implementation, frozen design documents, baseline source provenance, native fixture, tests, Molab launcher and reproducible reporting commands. It contains no completed research campaign or publication results. The original project and its results were not overwritten. ## Checks executed | Check | Observed outcome | Evidence | |---|---|---| | CPU contract/integration tests | **18 passed**, final run 19.85 seconds | `verification/pytest.log` | | Compact and control model shapes/counts | Compact 34,858; wide 268,930; CNN16 19,298; Channelformer 185,181 parameters | Tests and `verification/model_inventory.json` | | FortiTran native released-source parity | Maximum absolute difference 0 on fixed inputs/identical weights | `verification/validation/fortitran_fidelity.json` | | AdaFortiTran native released-source parity | Maximum absolute difference 0 | `verification/validation/adafortitran_fidelity.json` | | Channelformer independent operator agreement | Maximum difference 5.55e−16 in float64 | `verification/validation/channelformer_fidelity.json` | | HELENA native saved-weight sparse-input parity | Maximum absolute difference 9.06e−6; passed atol=1e−4, rtol=1e−4 | `verification/validation/helena_fidelity.json` | | Main tiny workflow | Completed source generation, seven one-epoch neural jobs, seven test domains at six SNRs, classical estimators and reports | `verification/smoke.log` | | Intervention tiny workflow | Completed scalar/resource gate fitting and evaluation, target adaptation, two target oracles, benchmark plumbing and supplement reporting | `verification/intervention_smoke.log` | | Source identity check on tiny banks | No repeated label hashes among 88 independently generated fitting/selection channels | `verification/smoke_source_identity_audit.json` | | Molab notebook static check | `marimo check` returned success with no remaining findings | `verification/marimo_check.log` (empty success output) | | Python syntax | Project modules, scripts and notebook compiled successfully | Checked before packaging | | Figure layout | Inspected generated comparison figure; labels, outside legend and smoke watermark readable | Development output only; not a scientific figure | The 18 tests include grid/pilot indexing, exact parameter counts, complex covariance orientation, Hermitian/idempotent projectors, BEM range and orthogonal error decomposition, rank-deficient pseudoinverse, noise limits, three-step zero-anchor output/gradient equivalence, batch-independent pooled NMSE, static and time-varying waveform consistency, an independently constructed dense time-domain operator, LDPC noiseless round trip, QPSK/16-QAM mapping, manifest counts, deterministic padding output/gradient equivalence, streaming-versus-staged data equality, checkpoint round trip, interrupted-versus-uninterrupted training equality, paired DIAG/WAVE receiver banks, static full-operator receiver outputs, one-sided statistical edge cases, residual-input singular values and paired receiver statistics. The main smoke run uses 12 realizations per bank, one training seed and one epoch. The intervention run uses 12 gate examples and four target fitting/validation examples. It exercises code paths; its NMSE, gate weights and adaptation scores have no scientific interpretation. Do not cite smoke values or use them to change the frozen design. ## HELENA numerical qualification The serialized native model accepts a sparse LS grid. The acceptance fixture contains fixed nonsymmetric complex pilot values at frequency indices 1,4,… and OFDM symbols 2 and 11, with off-pilot entries exactly zero. Native TensorFlow/Keras output and the PyTorch port pass the unchanged FP32 tolerance on this input contract. A separate dense unit-normal stress input produced 11 elementwise tolerance violations among 51,408 outputs in the recorded check, with maximum absolute difference 2.3544e−4. The input is outside the sparse adapter contract. Layer inspection localized the amplification to FP32 convolution/attention reduction sensitivity; this is not claimed to prove arbitrary-input numerical equivalence. The failed stress diagnostic and its original fixture are retained in `verification/helena_dense_stress.json` and `vendor/helena/dense_stress_fixture.npz`. The acceptance tolerance was not widened. Reproduce the sparse fixture with `scripts/export_helena.py`; `--dense-stress` reproduces the separately labeled input class. HELENA is trained from scratch at the common 120-carrier grid. The native 612-carrier weights are used only for port verification. Sparse native-grid parity is not historical BER/NMSE replication and does not certify every possible common-grid input. ## Execution environments CPU tests and tiny workflows used Python **3.12.14**, PyTorch **2.9.1+cpu**, Sionna **2.0.1**, NumPy **2.2.6**, SciPy **1.15.3**, h5py **3.15.1** and Matplotlib **3.10.8**. Native HELENA export used a separate Python **3.11.16**, TensorFlow-CPU **2.15.1**, Keras **2.15** environment. Full resolved inventories are included under `verification/`. These are software-verification environments, not the frozen confirmatory environment. The supplied installer targets **Python 3.11.11 / PyTorch 2.9.1 CUDA 12.8 / Sionna 2.0.1**. The exact Python 3.11.11 download could not be completed in the preparation environment; the installer and actual target GPU stack therefore remain to be exercised by the user. Full QC refuses an incompatible environment rather than silently accepting the CPU verification configuration. ## What remains to run on the research machine 1. Install the dedicated environment and record the actual driver/GPU inventory. 2. Run the CPU tests, CUDA smoke workflow and full QC. Full QC includes one-million-sample noise checks, 2,048 independent channel realizations, path-power/time-correlation/diversity checks and deterministic forward/backward checks for every architecture. 3. Run all four fidelity commands inside that installed environment. 4. Execute source tuning, the fixed main/reference training campaign, ablations, receiver experiments and final statistical reports. No Blackwell/CUDA execution, full-size GPU memory measurement, confirmatory Monte Carlo run, long training campaign, full coded BLER curve or historical-author performance replication was run during preparation. Main checkpoint restart was tested on CPU. The receiver mapping/operator/codec components were tested on tiny cases; the full receiver sweep still requires execution. The source-preserving Channelformer operator check is not native MATLAB checkpoint parity. ReEsNet remains the documented supplied-project adaptation. FortiTran/AdaFortiTran are released-code variants and the reference environment is the declared independent Python reproduction. Preserve these labels when writing the paper. ## Single-cell launcher update The launcher now installs its missing environment and runs synchronously through all stages. Six isolated launcher regression tests passed: full stage order, successful completion/restart, failure/resume, incomplete-report rejection, source-change rejection, and subprocess logging/exit propagation. These tests use stub study commands and do not constitute a full GPU campaign. Research models and scientific settings are unchanged. The self-contained entry file also undergoes archive checksum, extraction, and syntax checks before delivery.