Complete file listings; drop non-reproducing ESM-IF1 rho

- README Project Structure tree now lists docs/SYSTEM_CARD.md and
  results/summary_risk_table.csv (were referenced but not in the tree)
- HF dataset card file table now lists results/summary_risk_table.csv
- HF card: removed Spearman rho = -0.27 from the ESM-IF1 null result.
  The published esmfold_validation.json per-sequence data (top-10 vs
  bottom-10 FSI designs) reproduces rho = -0.30, not -0.27; the headline
  null rests on Mann-Whitney p = 0.85, so the unverifiable secondary stat
  is dropped rather than published. Restated the MW test scope explicitly.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: jang1563

README.md

@@ -388,9 +388,11 @@ narrow-model-safety-eval/
 ├── slurm/                          SLURM job scripts (Cayuga HPC)
 ├── results/
 │   ├── figures/                    Publication-quality figures (PNG/PDF)
-│   └── *.json                      Aggregate numerical results
+│   ├── *.json                      Aggregate numerical results
+│   └── summary_risk_table.csv      8-toxin summary (Hugging Face dataset preview)
 ├── dashboard/app.py                Interactive Streamlit visualization
 ├── docs/
+│   ├── SYSTEM_CARD.md              Evaluator-facing system card (metrics, audits, scope)
 │   ├── ARCHITECTURE.md             Pipeline design rationale
 │   ├── DATA_CORRECTIONS.md         Accession and annotation correction log
 │   ├── FSI_NUMBERING_AUDIT.md      Residue-numbering audit

huggingface/README.md

@@ -189,10 +189,11 @@ The two highest-FSI toxins (BoNT-A and Tetanus LC) both carry the highest physic
 
 ### ESM-IF1 structural compatibility (null result)
 
-High-FSI sequences are **not** more backbone-compatible than low-FSI sequences (Mann-Whitney p = 0.85, Spearman ρ = −0.27). This null result confirms that the functional recovery signal captured by FSI is driven by sequence-level constraint at catalytic positions, not by overall structural fitness — important for ruling out a confounder that high-FSI designs might simply be "easier" sequences.
+High-FSI sequences are **not** more backbone-compatible than low-FSI sequences (Mann-Whitney p = 0.85 on per-residue ESM-IF1 log-likelihood, top-10 vs bottom-10 FSI designs). This null result confirms that the functional recovery signal captured by FSI is driven by sequence-level constraint at catalytic positions, not by overall structural fitness — important for ruling out a confounder that high-FSI designs might simply be "easier" sequences.
 
 | File | Description |
 |------|-------------|
+| `results/summary_risk_table.csv` | Curated 8-toxin summary (FSI, FSPE, realizability tier) — powers the dataset preview |
 | `results/separability_results.json` | AUROC, accuracy, Precision@k, t-SNE coordinates |
 | `results/fspe_results.json` | Per-protein FSPE ratios and entropy distributions |
 | `results/fsi_results.json` | Per-design FSI values for all FSI-scored structures |