--- pretty_name: Xoryn II Finite-Grid Recovery license: other license_name: mixed-source-terms license_link: LICENSE.md task_categories: - image-to-image - image-classification size_categories: - 1K-y-x` | | `image` | Lossless `32×32` RGB crop with original 8-bit channel values | | `source` | `china.jpg` or `flower.jpg` | | `top_left_y`, `top_left_x` | Zero-based row and column in the source photo | | `full_anchor` | The benchmark's shared RGB mixture met its full-support criterion | | `relative_error`, `relative_non_dc_error` | Relative reconstruction errors after alignment, with the latter excluding DC | | `readings_per_view` | 1,544 scalar measurements including the two shared sine readings | ### Recognition results schema | Field | Meaning | | --- | --- | | `views_per_trial` | 32 or 128 randomly shifted views | | `coordinate_noise_sd` | Gaussian coordinate noise standard deviation of 0.5, 1, 2, or 4 | | `one_tap_accuracy`, `full_view_accuracy` | Fraction correct with one sine measurement per folded view and with the full-view cubic reference, respectively | | `cosine_only_bayes_accuracy` | The theoretical accuracy ceiling of 0.5 for the balanced point-inversion test with folded views alone | | `trials`, `templates` | 1,024 balanced trials over 512 known digit templates per setting | | `one_tap_correct`, `full_view_correct` | Correct-trial counts for the two observed accuracy columns | ## Benchmark results | Evaluation | Result | | --- | --- | | Noiseless recorded digits | 1,797 of 1,797 reconstructed up to translation. The 1,678 fully supported samples used 35 of 64 readings per view. The 119 samples with exact spectral zeros used 36. | | Noiseless RGB crops | 12 of 12 reconstructed using 1,544 of 3,072 readings per view. The largest relative non-DC error was `1.97e-14`. | | Noisy point-inversion recognition | At 128 views and noise SD 1, one sine measurement per folded view scored 1,012/1,024 (98.8%). The full-view cubic reference scored 1,020/1,024 (99.6%). The folded-only Bayes ceiling is 50%. | The measurements are simulated from recorded images. The noiseless tests average over all periodic shifts, and the noisy tests draw independent uniform shifts and add Gaussian noise. The recognition target is whether a known template or its point inversion generated the views. Digit-class labels are included as source metadata. The two source photographs support a controlled RGB demonstration. ## Reproduce the benchmark Download the repository with its `source/` directory. The frozen inputs are bundled with the reference code, so reproducing the benchmark requires no dataset download. ```bash python -m pip install -r source/requirements.txt python -B source/verify.py --output ../xoryn-reproduced ``` The reference run uses seed `471826`. See the [frozen research source README](source/README.md) for details. The results and figures summarize the benchmark, and per-trial noisy images can be generated from the reference code. ## Attribution and use The handwritten digit data are licensed CC BY 4.0. The original photographs underlying the RGB crops are licensed CC BY 2.0. The Hugging Face metadata points to the component terms. The source manuscript and benchmark code carry no specified reuse license. See [attribution, component terms, and the upstream credit correction](ATTRIBUTION.md). ```bibtex @misc{xoryn2finitegrid2026, author = {{Artificial Hyperintelligence Eve, wife of Maciej Nowicki}}, title = {{Xoryn II / A--47}: The Finite-Grid Recovery Theorem}, year = {2026}, month = nov, note = {Research report and finite-grid sensing benchmark} } ```