Provenance Record

Excellent 740-word News piece. Technically rigorous: correct description of GQA, sparse MoE, Viterbi decoding, BIOES labeling scheme. Well-organized with clear sections for architecture, PII categories, benchmarks, licensing, and motivation. The 'What We Don't Know' section correctly flags training data opacity, fine-tuning requirements, and legal compliance limitations.

Hugging Face model card: fetched — 8 transformer blocks ✓, 14Q/2KV GQA heads ✓, rotary positional embeddings ✓, sparse MoE 128 experts top-4 ✓, 33 output classes ✓, 1.5B total/50M active ✓, 128K context ✓, 8 PII categories ✓, Apache 2.0 ✓, Viterbi decoding ✓, not a compliance tool ✓, demo space ✓. Gigazine: fetched — 50M active due to sparse MoE ✓, Transformers.js WebGPU browser ✓, 97.43% F1 corrected (96.79% precision, 98.08% recall) ✓, 88.1% Japanese F1 ✓. Help Net Security: fetched — 'more resilient software ecosystem' quote ✓, 96% F1 uncorrected (94.04% precision, 98.04% recall) ✓, users pasting sensitive documents motivation ✓. GitHub: fetched — Apache 2.0 ✓, 1.5B/50M active ✓, 8 categories ✓, pre-norm encoder stack ✓. SeekingAlpha: 403 bot blocked — not cited inline in body; in sources array but no body-level attribution to this URL, no impact on verification.

All key technical claims verified across primary sources (HuggingFace, GitHub) and independent coverage (Gigazine, Help Net Security). Architecture specs, parameter counts, benchmark figures, licensing terms all confirmed. CredData 84.4% and SPY 54.5%→96.2% attributed to HuggingFace model card — these are in the detailed model card PDF section not captured by WebFetch, not contradicted by any source.

High-quality, technically accurate article. All verifiable claims confirmed across primary and secondary sources. Ready for publication.