docs(analyst): sync MODEL-CARD.md to the published HF card restructure

[thebenignhacker]

What

Replaces models/nanomind-v3-qwen3-1.7B-sft-r64/MODEL-CARD.md with the restructured content now live on the HF analyst card (opena2a/nanomind-security-analyst, commit e5345565), so the repo's canonical model card (referenced by nanomind-models.json modelCard) is identical to the public card.

What the restructure adds

• Runnable Quick Start — transformers (bf16, exact prompt contract, greedy) and llama.cpp CPU-GGUF, with the Metal-breakage caveat (the Q4_K_M GGUF produces garbage under Metal/GPU offload on Apple Silicon; correct on CPU). Both paths were validated against the SHA-verified v3.0.0 artifact.
• A real example generation (not illustrative).
• Corrected numbers — fp-suppression ~77% (repaired gate; the shipped 57% was on a gate later found 41% placeholder filler), input gate threshold 0.90, dropped the superseded 92% figure. Matches PR docs(analyst): correct fp-suppression to ~77% and gate to threshold 0.90 #38 and the website.
• Internal ledger (CDS decisions, PR SHAs) moved to a clearly-labelled maintainer appendix; internal [CDS-020]/[CDS-022]/[CDS-006] markers removed from the body.

Verification

• MODEL-CARD.md is byte-identical to the live HF README.md (diff -q clean).
• Follows docs(analyst): correct fp-suppression to ~77% and gate to threshold 0.90 #38 (which corrected the numbers minimally); this brings the full restructure into the repo to match HF.

https://claude.ai/code/session_01BdPyMkt6oYGyZL37AWQbzA

[github-actions]

Claude Code Review

VERDICT: APPROVE

SUMMARY: This is a documentation-only change that synchronizes the repo's MODEL-CARD.md with the live Hugging Face card (opena2a/nanomind-security-analyst, commit e5345565). The restructure adds runnable Quick Start examples (Transformers bf16 + llama.cpp GGUF with Metal-breakage warnings), a real example generation, corrected metrics matching PR #38 (fp-suppression ~77%, input gate threshold 0.90), and moves internal ledger content to a maintainer appendix. The PR description verifies byte-identical match with the HF README.md. No code, configuration, or publishable surface is changed — only markdown documentation in models/nanomind-v3-qwen3-1.7B-sft-r64/MODEL-CARD.md. The content is accurate to the model's limitations (off-topic hallucination, dual-use security code over-flagging, injection recall weakness) and deployment requirements (NanoMind-Guard daemon + input-classifier gate). All technical details (bf16 on MPS, GGUF Metal breakage, greedy decoding requirement, prompt contract) match the existing published state. No security, correctness, or publish-surface issues introduced.

---

Reviewed 1 files changed (30097 bytes)