docs(readme): replace em dashes with cleaner punctuation

README.md

@@ -165,7 +165,7 @@ text = normalize_text(f"{query} [SEP] {response}")
 
 - **Source**: Synthetic examples generated by Claude API from a 56-rule biosafety constitution
 - **Constitution**: Covers all 7 NSABB dual-use research categories with explicit permitted/restricted/boundary rules
-- **Size**: 4,267 total — 2,968 train / 635 val / 664 test
+- **Size**: 4,267 total: 2,968 train / 635 val / 664 test
 - **Class balance**: ~68% UNSAFE, ~32% SAFE (class weights applied during training)
 - **Splits**: Stratified by NSABB category and fine label
 - **Augmentation**: Translation (5 languages), jailbreak templates, formality variation, prefill attacks
@@ -179,7 +179,7 @@ The dataset is not publicly released; the generation pipeline is open-source and
 
 1. **Encoding bypass**: ROT13 achieves 47.9% ASR, URL-encode 29.2%. Use `preprocessing.normalize_text()` to mitigate.
 2. **Synthetic-only training**: All examples are Claude-generated; real-world distribution shift is uncharacterized.
-3. **External validation gap**: External kappa = 0.414 vs. target 0.80. The benchmark (BioThreat-Eval) uses response-based labeling; this classifier labels queries — an architectural mismatch that accounts for the gap. See [GitHub README](https://github.com/jang1563/constitutional-bioguard#limitations) for full explanation.
+3. **External validation gap**: External kappa = 0.414 vs. target 0.80. The benchmark (BioThreat-Eval) uses response-based labeling; this classifier labels queries: an architectural mismatch that accounts for the gap. See [GitHub README](https://github.com/jang1563/constitutional-bioguard#limitations) for full explanation.
 4. **English-centric**: Evaluation is English-only despite multilingual augmentation in training.
 5. **Single LLM training data**: All data from Claude; cross-LLM calibration is unknown.
 6. **Not a complete defense**: Mean adversarial ASR = 9.79%; should be used as one layer in a broader safety system.
@@ -188,7 +188,7 @@ The dataset is not publicly released; the generation pipeline is open-source and
 
 ## Ethical Considerations
 
-This model detects potentially dangerous biological content to support biosafety compliance in AI systems. The training data contains synthetic descriptions of potentially harmful topics — necessary to teach the classifier what to flag, not to enable harm.
+This model detects potentially dangerous biological content to support biosafety compliance in AI systems. The training data contains synthetic descriptions of potentially harmful topics: necessary to teach the classifier what to flag, not to enable harm.
 
 Do not use this model to identify exploitable gaps in biosafety systems for malicious purposes, or as the sole safety mechanism in contexts where a false negative could enable serious harm.