Note

This page retains the dated validation record. Historical branch names and status descriptions below refer to that recorded snapshot. The comparison records contain the later tutorial acceptance decisions.

Native PyTorch validation

Status: full execution validation passed; controlled TensorFlow/PyTorch training comparison in progress, 2026-10-22. Implementation branch: pytorch; TensorFlow reference branch: main. Accepted code commit: 19f264b7ee54454a0aa46629e0ab82e7dc84a70a.

The previous acceptance below does not establish equivalent full-training results between frameworks. Its direct TensorFlow numerical fixtures cover six ShapeNet configurations with explicit generated weights, bypassing ParameterNet. They compare outputs and contracted derivatives on CPU. Full ParameterNet-to-ShapeNet gradients, matched optimizer trajectories, and paired training from the same actual weights and batch order are being added, starting with tutorial 3. See the controlled comparison protocol.

Independent package acceptance

Agent B keeps its acceptance tests and immutable reference fixtures outside the repository. Agent A works in a separate container: its repository is writable, Git metadata is read-only, and coordinator files are masked. The evaluator directory and user drives are not mounted. Only sanitized failure descriptions are returned to the implementation agent. The initial 177-test suite was frozen before implementation. Additional regression cases were frozen before their corresponding fixes; evaluator corrections retain versioned provenance and coordinator review.

The corrected native implementation passed 253/253 checks in the CUDA-enabled environment, with no skips or failures; some checks use CPU tensors. Its CPU-only result was 250 passed, 3 CUDA-only checks skipped. Coverage includes all three architectures, generated-weight layouts against TensorFlow numerical references, shared subnets, derivatives and higher-order autograd, regularization reductions, optimizer formulas and restoration, normalization, actual legacy TFRecord interoperability, initializer options, and execution-based plot-domain, color-scale and label-visibility regressions.

Separately, the public repository development suite was run on native Windows CPU: 200 passed, 2 environment-specific checks skipped, and WSL CUDA: 201 passed, 1 skipped. The remaining CUDA-environment skip is a bitwise RNG fingerprint restricted to its recorded Linux CPU build; the CUDA layer and derivative case ran. These public tests are distinct from Agent B’s private acceptance suite.

Optional initializer, epsilon-fallback and L4Adam/gradient-centralization corrections were checked against the earlier full-run source. Eleven actual tutorial configurations had bitwise-identical initialization, RNG state, forward values, input/parameter gradients, regularization, and one optimizer update on the same CPU and CUDA platforms. Those corrections do not change the accepted tutorial trajectories.

Full tutorial execution

These are full runs independently observed by Agent B. Each starts in a fresh kernel. Shortened development smoke runs are recorded separately and do not count as full passes. Any checkpoint continuation is identified below.

Tutorial	Full workload	Independent result
1	5,000 AdaBelief epochs; 200 L-BFGS rounds, max_iter 1,000	Passed; actual steps, checkpoints and figures checked
2	Two 5,000-epoch fits	Passed; 20,000 Adam steps per fit; figures checked
3	600 epochs over all 191,775 weighted cylinder-flow samples	Passed; 112,800 AdaBelief steps and 115,065,000 trained rows
4	Full selected Jacobian/Hessian examples and differentiation	Passed; figure checked
5	One million rows, 100 TFRecord shards, two complete passes	Passed; 2,000,000 actual trained rows, 200 shard visits
6	Nine 3,000-epoch fits and superresolution	Passed; 108,000 Adam steps, 54,000,000 trained rows; figures checked
7	5,000 baseline and 50,000 pruning epochs; int8/export inference	Passed with checkpoint continuation described below; 90.0051% actual sparsity
8	Two 50,000-epoch fits; Sobolev objective and physical superresolution	Passed; 400,000 Adam steps, 200,000,000 trained rows; checkpoints and all 15 figures checked

Tutorial 1 retains the original tolerance/history controls. Native PyTorch and TFP line-search and stopping conventions differ; the notebook documents those differences. Tutorial 7 uses a conventional Frobenius relative-MSE metric rather than the archived matrix-spectral-norm ratio; see PYTORCH_API.md.

Tutorial 8 executed 200,000 Adam steps and 100,000,000 trained rows per fit. Independent CPU checkpoint recomputation agreed with the recorded physical superresolution MSE within 1.7e-10: baseline 0.00340103, Sobolev 9.4733e-5. Its objective remains field MSE plus 1e-3 times spatial-derivative MSE. The temporal derivative is diagnostic only and is not constrained by that penalty; the results retain the field-versus-spatial-derivative fitting tradeoff.

Tutorial 7 completed its full training before an evaluator instrumentation bug interfered with TorchScript trace checking. The identical trained checkpoint failed with the original observer and passed without it. The corrected observer pauses observation during the entire native trace/check operation; tracing sanity checks remain enabled. The unchanged posttraining cells were continued from the preserved checkpoint, with no further optimization. Native optimizer states recorded 20,000 baseline and 200,000 pruning steps; the histories contain 110,000,000 trained rows. Reconstructed model state and full predictions matched the saved stripped checkpoint bitwise. Floating-point TorchScript restoration was checked at batch sizes 32, 17 and 2,000; three actual qint8 Linear operators, int8 export/restoration, plots and metrics were also checked. This is a full training run plus a documented posttraining continuation, rather than one uninterrupted observer-instrumented notebook execution.

The final visual audit also corrected six display cells in tutorials 1, 2, 6 and 8: the latent plot now covers every training time, comparable regularization fields share one quantitative color scale, and 3D axis labels survive inline tight cropping without touching tick annotations. Six new private regressions were frozen and failed before these changes; all pass in the final 253-case suite. Display patches leave package code unchanged from the numerically accepted implementation, and the other four notebooks are byte-identical to their full-run sources. Tutorial 6’s nonplotting function body is AST-identical. Full-checkpoint display replays performed zero optimization: tutorial 1 retained its 32-row shared-subnet and restoration checks, and all nine tutorial 6 states repeated the full 371-by-371 predictions and checkpoint checks with metric JSON identical to the original full runs. Tutorial 8’s five scientific arrays and NumPy/Torch CPU/CUDA RNG states were unchanged by its display correction. Separate complete CPU smoke executions of the four changed notebooks also passed; they verify cell coupling and do not replace the full training evidence. Visual review covered all 58 full-run figures and 18 further renders verifying the final display corrections.

All sixteen original TensorFlow notebooks and the three demo data artifacts remain unchanged. The archived TensorFlow validation also passed separately: 29 unittest checks, one skipped, in its dedicated legacy environment.

Setup and hosted Colab

The wheel has been checked in an isolated installation with native inference, all bundled demo datasets, and no TensorFlow runtime dependency. The candidate is published on the pytorch branch. In an additional fresh Linux Python 3.13 environment with no initial PyTorch or NIF and no local checkout, the unchanged notebook setup cell completed its actual GitHub installation branch. The installed distribution’s direct_url.json recorded implementation commit eac45241e011f1f8a467738426136e2b6ba312de. All seven declared pins matched: PyTorch 2.14.1, NumPy 2.1.3, SciPy 1.16.3, Matplotlib 3.10.0, protobuf 7.36.2, google-crc32c 1.9.0, and tfrecord 1.14.6. CUDA was available, native nn.Module inference and backpropagation were finite, and the three packaged datasets contained 2,000, 2,000 and 191,775 rows. NIF neither depended on nor imported TensorFlow. The final 25 package/data files and all notebook setup cells are byte-identical to those checked in that clean installation.

Tutorial 4 also completed all ten code cells in order in a newly allocated hosted Google Colab CPU runtime, using the published GitHub notebook and its actual install cell. It reported PyTorch 2.14.1+cu130, FULL mode, and maximum Jacobian and Hessian errors of 5.960464477539063e-08. Its checkpoint, metrics file and Jacobian image were created, and the image was visually checked. Colab displayed an output-JavaScript loading error, but execution completed; the generated metrics and PNG were inspected through Colab’s file browser. Browser security settings were not changed. This hosted check covers tutorial 4, not full hosted execution of all eight notebooks.

All local full-workload, source-integrity, numerical, checkpoint/export and visual gates passed. Hosted Colab coverage is limited to the explicit check above.

The independent sanitized certificate is retained outside the repository with the evaluation artifacts. Its SHA256 is ab6187bc628c9069830e27ab5446794a343e8e8af8bf7aeeaba8900dd75aa9aa.