Research artifacts for Model Consolidation and the Family-Balanced Shared-Core Optimizer for Multi-Source Identity-Document Fraud Detection
Description
This record contains the redistributable research materials supporting “Model Consolidation and the Family-Balanced Shared-Core Optimizer for Multi-Source Identity-Document Fraud Detection.” Version 1 comprises a conventional ZIP archive with one root directory, a top-level README, a file inventory, a SHA-256 checksum list, and a byte-identical copy of the article’s Supplementary Material. The archive provides a single command-line entry point (artifact.py) for integrity verification, executable demonstrations, workflow discovery, public RQ1 and RQ2 analyses, and independent FB-SCO retraining on a user’s own lawful document collection using precomputed 512-dimensional features. Training preserves the reported shared-core and corpus-source-local parameter ownership and supports calibrated checkpoint inference. The record contains frozen configurations, aggregate and non-identifying measurements, source tables, provenance records, and a claim–evidence map. It excludes protected identity-document images, restricted row-level records, credentials, and pretrained weights. Exact manuscript replay requires independent lawful access to the cited datasets and models, as specified in FULL_DATASET_VALIDATION.md; independent retraining is distinguished explicitly from exact replay. Original code and documentation are licensed as identified within the archive. AIForge-Doc-derived aggregate and protocol files retain CC BY-NC-SA 4.0.
Files
Steps to reproduce
1. Download all five top-level files and verify their SHA-256 digests against SHA256SUMS.txt. 2. Extract FB_SCO_Research_Artifacts.zip and enter the FB_SCO_Research_Artifacts directory. 3. Run the dependency-free checks: python artifact.py verify python artifact.py demo python artifact.py list 4. Create an isolated Python environment, activate it, and install the declared dependencies: python -m venv .venv-artifact Windows: .venv-artifact\Scripts\python -m pip install -r requirements-artifact.txt macOS/Linux: .venv-artifact/bin/python -m pip install -r requirements-artifact.txt 5. Reproduce the public analyses: python artifact.py run rq1 python artifact.py run rq2 python artifact.py test 6. Test a complete independent training run: python artifact.py example-training --output scratch/own_data_example 7. To use a lawful document collection, first read OWN_DATA_TRAINING.md and print the fixed feature contract: python artifact.py feature-contract python artifact.py init-dataset --output my_document_library python artifact.py validate-dataset --manifest my_document_library/manifest.csv python artifact.py train-features --manifest my_document_library/manifest.csv --output my_document_library/fb_sco_fit --device auto python artifact.py predict-features --checkpoint my_document_library/fb_sco_fit/model.pt --manifest my_document_library/prediction_manifest.csv --output my_document_library/predictions --device auto The independent-training interface expects one normalized 512-dimensional .npy feature vector per row and does not replace the reported Stable Diffusion v1.5 extractor with another backbone. Users remain responsible for lawful images, feature provenance, corpus-source/provenance strata, and leakage-safe component construction. FULL_DATASET_VALIDATION.md describes the additional restricted inputs required for exact image-level manuscript replay.