Datasets:

PureOne
/

Eidrion

This guide separates executed evidence, fresh reruns, and unrun extensions. The public name is EIDRION; the source-locked avenyra_omega implementation and historical experiment filenames are preserved.

1. Assemble and verify

Extract Eidrion_v1.0.0_Research_Package.zip. The main package is enough for the core/API checks and reading all reports. Large companion evidence is distributed in numbered binary parts. Keep the parts in one directory, then use release_tools/restore_evidence.py and the included companion_parts.json to verify and reconstruct the four ZIPs. Extract all four into the same parent as the main archive: their shared root is Eidrion_v1.0.0/.

python release_tools/restore_evidence.py --parts-dir PATH_TO_DOWNLOADS --restore

Omit --restore to verify without writing ZIPs. Use --output-dir PATH to restore elsewhere. The parts and reconstructed archive both consume disk space until the user chooses which copies to keep.

Use Python 3.12 in a fresh virtual environment. The captured CPU environment used Python 3.12.14, torch 2.14.1+cpu, numpy 2.3.5, scipy 1.17.0, and scikit-learn 1.8.0. From the extracted project directory:

python -m pip install torch==2.14.1 --index-url https://download.pytorch.org/whl/cpu
python -m pip install -r requirements-reproduction.txt
python -m pip install -e . --no-deps
python release_tools/verify_release.py --mode quick --output ../checks_quick_new

Every output path above must be new. Quick verification checks all main-group hashes, 39 grouped mathematical/numerical fixtures, the ordinary-training example, and generated-data extension smokes. Companion groups are explicitly marked NOT_AUDITED, whether present or missing. These fixture checks are not population generalization experiments.

After all companions are assembled:

python release_tools/verify_release.py --mode hashes --output ../checks_all_hashes_new

This checks every listed file against its recorded size and SHA-256. A manifest supplies integrity consistency, not a digital signature or independent proof of authorship. Generated local outputs are outside its scope. Approximately 12 GB of uncompressed release/evidence plus 7.7 GB of compressed evidence are involved; allow additional working space if retaining both parts and reconstructed ZIPs.

2. Run the API example

python examples/quickstart.py --output ../quickstart_new

The example trains all layers of the bounded raw residual, checks frozen external features, checkpoint/optimizer restoration and a matched resumed update, and exercises fresh acceptance/refusal. Read examples/README.md for the actual retention contract. The public import is eidrion; captured benchmark imports remain avenyra_omega.

3. Stage official image data

Raw downloaded archives are intentionally excluded. experiments/data/archive_manifest.json records the official URLs and SHA-256 digests. Both class-stratified training_partition_ids.npz files are included in main and must be preserved. Download the original archives into their existing dataset directories:

python experiments/image_data.py --data-dir experiments/data --datasets mnist fashion_mnist

Use experiments/DATA_SOURCES.md and the archive manifest for provenance and dataset terms. That note was written during staging: its statement that no official test images had yet been evaluated describes that historical stage, not the completed release. Later, all final locked runs evaluated the official held-out test sets.

4. Reproduce all 260 locked image runs

The final source lock is af4ac3635681ac02faa2b36181830f45654c3f9549601258e2fe4aa879308825. The portable wrapper verifies the captured source, exact installed package versions, and raw dataset hashes before starting. It copies only locked source into a fresh output directory; it does not overwrite captured predictions or checkpoints.

python experiments/reproduce_final.py --captured-root experiments/final_standard --output-dir ../final_rerun_new --data-root experiments/data --workers 4

For setup validation only, add --prepare-only and use a different new output directory. Prepare-only was executed and checked; a second complete 260-run training reproduction was not performed for this release. The supplied 260-run capture itself completed without execution failures or relocking.

The protocol has two datasets, ten paired seeds, thirteen methods, five two-class arrivals, and shared ten-class inference without task identity. The 6 MiB persistent-learner budget counts retained data and calibration evidence. The joint reference is explicitly exempt; four-epoch ER-ACE is a declared compute-relaxed control. See experiments/FINAL_PROTOCOL_NOTES.md and FINAL_RESULTS.md for complete qualifications. The locked accounting harness uses Unix resource; use Linux or WSL. Native Windows training reproduction is not claimed.

5. Independently audit the captured image evidence

These are read-only reconstructions, not training reruns. Complete prediction/state companions and staged official datasets are needed. Create a new output directory, for example ../audit_new, before calling the scripts:

python prior_art/audit_final_predictions.py --root experiments/final_standard --data-root experiments/data --output ../audit_new/predictions.json
python prior_art/audit_final_checkpoint_contracts.py --root experiments/final_standard --output ../audit_new/contracts.json
python prior_art/audit_final_aggregation.py --root experiments/final_standard --prediction-audit ../audit_new/predictions.json --output ../audit_new/aggregation.json

Do not use --allow-partial for a complete-release audit. Captured independent audit reports are already included in prior_art/. They cover 260 runs, 2,860 prediction/evidence NPZ files, 800 controller checkpoints, and 7,136 parameter/controller binding checks. Calibration probabilities apply under the stated sampling assumptions; saved-data agreement alone does not verify those population assumptions.

6. Reproduce the 96 structural runs

python experiments/structured_stream/run_structured_stream.py --mode evaluation --output ../structured_rerun_new
python experiments/structured_stream/audit_saved_results.py --results ../structured_rerun_new --output ../structured_audit_new.json

The locked seed list and noise settings are in that experiment's protocol. Keep all settings, including the failures at nonzero noise. The captured repair record documents three incomplete NPZ archives recovered by exact locked reruns; surviving original arrays matched. Original partial files and recovery logs remain in the release. Recovery is not a routine reproduction step.

7. Inherited H-OSP v3 reproduction

The unchanged original release is inherited/Avenyra_Harmonic_OSP_Research_Package_v3.0.0.zip. Extract it separately and follow its own README and research/v3/retraction_study/run_retraction.py command-line help. Its full v1–v3 source/evidence hierarchy is inside that archive. The fresh v3 outputs used here are under verification/inherited_v3_fresh/; the independent inherited inventory is experiments/inherited_audit.json.

experiments/audit_inherited.py --research-root PATH --output NEW_JSON expects the extracted inherited research directory containing experiments/, v2/, and v3/, not the EIDRION project root. It audits existing inherited evidence and does not train a model. The reproduced 104-parameter neural study is explicitly support-informed.

8. Rebuild scientific figures and manuscript

Work in a disposable copy if rebuilding files covered by the release manifest. Figure generation reads the completed evidence and accepts a fresh output directory:

python manuscript/make_figures.py --root . --final-root experiments/final_standard --output ../figures_new --require-final

manuscript/make_tables.py rebuilds its tables from captured JSON. manuscript/build_report.py writes the PDF from numbered manuscript sources, tables, and figures. These two tools intentionally write their project outputs; do not run them against the sole immutable archive copy. PDF generation additionally uses ReportLab and Matplotlib; PDF inspection used PyMuPDF. The distributed 58-page PDF has already passed layout inspection and visual review of all rendered pages.

9. Unrun extensions and publication

extensions/README.md gives generated-NPZ engineering tests, the reserved-data CIFAR-100/ResNet18 converter, Permuted-MNIST converter, and bounded LoRA example. Generated-data smoke tests ran. Actual CIFAR-100, Permuted-MNIST, pretrained transformer adaptation, and GPU/distributed experiments did not run and have no reported benchmark scores.

The Windows-friendly publish_windows.ps1 and Python publisher are dry-run-only by default. No Hugging Face authentication, repository creation, or upload was performed. A complete assembled package and the user's explicit publication authorization are required before executing an upload. The release assigns no invented public license or repository identifier.