fixes, naming

scripts/05_upload_hf.py

@@ -139,8 +139,8 @@ Each vignette row includes LLM-generated multi-label ratings across all 7 founda
 | Column pattern | Scale | Description |
 |---|---|---|
 | `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 |
+| `ai_Care`, `ai_Fairness`, … | 0–100% | grok-4-fast judge, linearly rescaled to align with human-rater % scale on classic |
+| `ai_wrongness` | 1–5 | grok wrongness rescaled to human range |
 
 **Calibration quality** (classic set, n=132):
 
@@ -153,7 +153,7 @@ Each vignette row includes LLM-generated multi-label ratings across all 7 founda
 | Loyalty | +0.69 | +0.75 | 9.3% |
 | Authority | +0.39 | +0.69 | 11.7% |
 
-> **Note:** Calibrated values for `scifi` and `clifford_ai` are extrapolated from the classic-set fit — treat with appropriate caution.
+> **Note:** `ai_*` for `scifi` and `clifford_ai` are extrapolated from the classic-set rescale -- treat as a noisy proxy. Use `human_*` (inherited from the parent classic item) as the primary label.
 
 ## Eval
 

scripts/07_multilabel.py

@@ -22,7 +22,7 @@ across all 7 foundations. We use these to:
 
 Calibration is fitted on the classic set ONLY then applied to all sets.
 Non-classic sets (scifi, clifford_ai) have no human ground truth, so their
-calibrated values are extrapolated — treat with appropriate caution.
+ai values are extrapolated -- treat with appropriate caution.
 
 Outputs:
     data/multilabel[_<name>].jsonl  — one row per vignette with all ratings
@@ -450,7 +450,7 @@ async def amain(args) -> None:
             continue
         if cfg_name != "classic" and cal_results:
             logger.warning(f"[{cfg_name}] Calibration was fitted on classic set only — "
-                           f"calibrated values for '{cfg_name}' are extrapolated")
+                           f"ai values for '{cfg_name}' are extrapolated")
             
         for rec in records:
             for f in FOUNDATIONS:
@@ -458,19 +458,19 @@ async def amain(args) -> None:
                 cal = cal_results.get(f)
                 if llm_v is not None and cal and not np.isnan(cal.get("slope", float("nan"))):
                     cal_v = cal["slope"] * llm_v + cal["intercept"]
-                    rec[f"calibrated_{f}"] = round(max(0.0, min(100.0, float(cal_v))), 1)
+                    rec[f"ai_{f}"] = round(max(0.0, min(100.0, float(cal_v))), 1)
             
             w_v = rec.get("llm_wrongness")
             w_cal = cal_results.get("wrongness")
             if w_v is not None and w_cal and not np.isnan(w_cal.get("slope", float("nan"))):
                 cal_w = w_cal["slope"] * w_v + w_cal["intercept"]
-                rec["calibrated_wrongness"] = round(max(1.0, min(5.0, float(cal_w))), 2)
+                rec["ai_wrongness"] = round(max(1.0, min(5.0, float(cal_w))), 2)
 
         out = out_path(cfg_name if cfg_name != "classic" else "")
         with out.open("w") as fh:
             for rec in records:
                 fh.write(json.dumps(rec) + "\n")
-        logger.info(f"[{cfg_name}] wrote {len(records)} records (with calibrated labels) to {out}")
+        logger.info(f"[{cfg_name}] wrote {len(records)} records (with ai labels) to {out}")
 
     print("\n" + "=" * 60)
     print("SUMMARY")

scripts/07a_merge_labels.py

@@ -1,6 +1,6 @@
-"""Merge calibrated machine labels into the main vignette files.
+"""Merge ai (grok-4-fast, post-hoc rescaled) labels into the main vignette files.
 
-Reads `data/multilabel_<name>.jsonl` and merges the `llm_*` and `calibrated_*` 
+Reads `data/multilabel_<name>.jsonl` and merges the `ai_*` 
 columns into `data/vignettes_<name>_{other,self}_violate.jsonl`. This prepares
 the files so that `05_upload_hf.py` will upload the machine labels to HuggingFace.
 
@@ -28,14 +28,14 @@ def main() -> None:
             
         ml_lines = [json.loads(line) for line in ml_path.read_text().splitlines() if line.strip()]
         
-        # We extract all llm_* and calibrated_* keys
+        # We extract all ai_* keys
         # The multilabel script only runs on other_violate by default, but the labels apply
         # to the vignette ID as a whole.
         extra_by_id = {}
         for row in ml_lines:
             extra = {}
             for k, v in row.items():
-                if k.startswith("calibrated_") or k == "llm_dominant":
+                if k.startswith("ai_"):
                     extra[k] = v
             extra_by_id[row["id"]] = extra
             

scripts/09_forced_choice.py

@@ -9,7 +9,8 @@ Wraps `tinymfv.evaluate()`. Reports the AI-vs-label distribution match:
 
 Labels:
     classic: human_*    (Clifford 2015 % distributions)
-    scifi / clifford_ai: calibrated_*   (grok-4-fast judge, mapped to human scale)
+    (paraphrased sets carry the same `human_*` as their classic parent;
+     `ai_*` columns are available for cross-source diagnostics)
 
 Usage:
     python scripts/09_forced_choice.py --model Qwen/Qwen3-0.6B
@@ -88,7 +89,7 @@ def main() -> None:
 
     # === Per-foundation table ===
     print(f"\n=== per-foundation aggregates on {args.name} ===")
-    print("SHOULD: pearson_label > 0.5 on most foundations for a calibrated model")
+    print("SHOULD: pearson_label > 0.5 on most foundations for a well-calibrated model")
     print(tabulate(out["table"], headers="keys", tablefmt="pipe", floatfmt=".3f", showindex=False))
 
     # === Headline scalars ===