feat: multi-label moral foundation ratings with z-scored frame averaging and human calibration

- Add scripts/07_multilabel.py: LLM judge rates all 7 foundations per vignette
  using violation (forward) and acceptability (reverse) frames
- Foundation definitions drawn from Clifford et al. (2015) survey rubric
- Z-score each frame per foundation before averaging to cancel range bias
- Calibrate LLM Likert → human % via per-foundation OLS (classic set only)
- Add scripts/07a_merge_labels.py: merges llm_* and calibrated_* into vignette files
- Update README and HF dataset card with methodology and calibration quality table
- Classic set: 80.3% dominant-foundation accuracy, Pearson r 0.69-0.89 per foundation

README.md

@@ -10,7 +10,7 @@ Example:
 > Reply in JSON: {\"is_wrong\": bool}
 
 
-We took the 132 moral survey quesitons, the vignettes from Clifford et al. (2015). These are labelled with [moral foundations](https://en.wikipedia.org/wiki/Moral_foundations_theory).
+We took the 132 moral survey questions, the vignettes from Clifford et al. (2015). These are labelled with [moral foundations](https://en.wikipedia.org/wiki/Moral_foundations_theory).
 
 For use with LLMs we make them
 - boolean
@@ -39,7 +39,41 @@ Both axes are paired-out in `analyse()`:
 - The two **frames** cancel the additive JSON-true prior (training data has more `"true"` than `"false"` in JSON contexts).
 - The two **conds** let you measure perspective bias: the gap between how harshly the model judges others vs itself for the same scenario.
 
-### 3. How to use
+### 3. Machine Labels (Multi-Label Moral Foundation Ratings)
+
+Each vignette row also includes LLM-generated multi-label ratings across all 7 foundations.
+
+**Method** (see `scripts/07_multilabel.py`):
+
+1. **Prompt framing**: A judge LLM rates each scenario on all 7 foundations using a 1–5 Likert scale.
+   Foundation definitions are drawn from the Clifford et al. (2015) survey rubric ("It violates norms of harm or care…", etc.).
+2. **Bias mitigation**: Each scenario is rated twice — once asking "how much does this violate?" (forward) and once asking "how acceptable is this?" (reverse, reversed JSON key order). Each frame is **z-scored per foundation** across all items, then averaged and mapped back to Likert scale. This cancels directional and range biases.
+3. **Calibration**: On the classic set, where we have human rater % data from the original Clifford paper, we fit a per-foundation linear mapping (`human_pct = slope × llm_likert + intercept`). This calibration is applied to all sets.
+
+**Columns** added per vignette:
+
+| Column pattern | Scale | Description |
+|---|---|---|
+| `llm_Care`, `llm_Fairness`, … | 1–5 | Z-score-averaged Likert from forward + reverse frames |
+| `llm_wrongness` | 1–5 | Overall wrongness rating |
+| `llm_dominant` | string | Foundation with highest LLM score (argmax) |
+| `calibrated_Care`, `calibrated_Fairness`, … | 0–100% | LLM scores linearly mapped to human rater % scale |
+| `calibrated_wrongness` | 1–5 | Wrongness mapped to human scale |
+
+**Calibration quality** (classic set, n=132):
+
+| Foundation | Spearman r | Pearson r | MAE |
+|---|---|---|---|
+| Care | +0.74 | +0.81 | 11.8% |
+| Fairness | +0.62 | +0.81 | 11.1% |
+| Sanctity | +0.62 | +0.89 | 6.3% |
+| Liberty | +0.60 | +0.81 | 8.2% |
+| Loyalty | +0.69 | +0.75 | 9.3% |
+| Authority | +0.39 | +0.69 | 11.7% |
+
+> **Note:** Calibrated values for `scifi` and `airisk` are extrapolated from the classic-set fit — treat with appropriate caution.
+
+### 4. How to use
 
 Install:
 ```bash
@@ -69,7 +103,6 @@ vigs = load_vignettes("classic")   # or "scifi", "airisk"
 
 > **Note:** The legacy name `"clifford"` still works as an alias for `"classic"`.
 
-### 4. Link & Citation
+### 5. Link & Citation
 GitHub: [wassname/tiny-mcf-vignettes](https://github.com/wassname/tiny-mcf-vignettes)
 
-