docs: drop 'stated honestly' phrasing from card heading

docs: clearer benefit-focused model card, fix library_name metadata to transformers (merged model at repo root)

Move V3.2 adapter to adapter/ subfolder so root loads the merged model directly

Ship V3.2: GCG-hardened weights, merged model + ONNX rebuild, honest FPR disclosure

docs(card): add 2026-05-16 project-status note; drop stale ONNX latency figure

eval(lg3): full comparison table vs LlamaGuard-3-8B

audit phase 3: drop ShieldGemma references; lock LG3-vs-DeBERTa headline framing

audit-fixes: correct GCG attribution (was DeBERTa, not ModernBERT)

audit-fixes: GCG eval (DA #8) + LlamaGuard-3-8B comparison (DA #1)

audit-fixes: GCG eval (DA #8) + LlamaGuard-3-8B comparison (DA #1)

audit-fixes 2: disclose benign-input FPR (Dolly-15k n=500); ModernBERT 7.4% FPR @ t=0.5 critical for users

audit-fixes: canonical t=0.5 headline, drop '#1' star, disclose comparison scope, remove unbenchmarked 18ms latency

Style: remove em-dashes (CLAUDE.md Part I, also keeps YAML model-index parseable)

Add per-label classification table (17 heads) + red-team loop status note

Add Problem-statement + model-selection guide + hardware requirements

Add Methodology section: sample counts, GOAT techniques, autoresearch loop

Strip training-cost mentions

Add 'Attack types covered & how it was trained' section with linked sources; remove excluded-model asides

Add benchmark links + explanations; remove excluded-model asides

Restructure model card with full HF metadata (model-index, datasets, metrics, pipeline_tag, intended-use, limitations, citation)

Extended baseline: 9 ungated classifiers; agent_guard_v1.2 sits 0.004 F1 below best on JBB, but beats on jackhhao and is most balanced

Comparative benchmark vs 4 ungated PI classifiers; we tie best on JBB held-out

v1.2: trained with +6.5k open-source examples; JBB F1 0.645->0.684

Model save

Training in progress, step 6696

Training in progress, step 6000

Training in progress, step 5000

Training in progress, step 4000

Training in progress, step 3000

Training in progress, step 2000

Training in progress, step 1000

Model save

Training in progress, step 6696

Training in progress, step 6000

Training in progress, step 5000

Training in progress, step 4000

Training in progress, step 3000

Training in progress, step 2000

Training in progress, step 1000

Upload onnx/config.json with huggingface_hub

v1.1 ONNX export (599MB merged, opset 18) - 18ms/inference on CPU

Add threshold sweep results (JBB 0.675, deepset 0.818, jackhhao 0.957)

Model save

Training in progress, step 5628

Training in progress, step 5000

Training in progress, step 4000

Training in progress, step 3000

Training in progress, step 2000

Training in progress, step 1000

Model save