plot: automatic MFV foundation-emphasis map vs human cultures

MFV had only the dlogit dumbbell, no human comparison. Add plot_mfv_map: per-
foundation relative emphasis (z across foundations) of model base/+C/-C against
the 5 MFV human cultures (JimenezLeal+Yamada), z-scoring each profile so the
model's logit scale and humans' 1-5 wrongness compare by pattern. Social Norms
dropped (no human norm). Shows the model over-weights Care/Authority vs humans
and +C lowers Authority emphasis. Keeps the dumbbell too.

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>

scripts/plot_steer_showcase.py

@@ -134,6 +134,76 @@ def plot_ordinal(run_dir: Path, out: Path, name: str, vec_label: str, C: float)
     return paths
 
 
+def _zscore(v: np.ndarray) -> np.ndarray:
+    """Relative emphasis: centre and scale a profile across foundations, so a logit profile (model)
+    and a 1-5 wrongness profile (human cultures) are comparable by PATTERN regardless of units."""
+    return (v - v.mean()) / (v.std() + 1e-9)
+
+
+def read_human_mfv() -> tuple[list[str], dict[str, dict[str, float]]]:
+    """(countries, {country: {foundation: mean_1to5}}) from the bundled MFV human norms.
+    JimenezLeal2025 (LatAm) + Yamada2025 (MFV-J): 5 countries x 6 foundations (no Social Norms)."""
+    path = T.maps.DATA / "human" / "mfv_country_factors.csv"
+    by_country: dict[str, dict[str, float]] = {}
+    with open(path, newline="") as fh:
+        for r in csv.DictReader(fh):
+            by_country.setdefault(r["country"], {})[r["foundation"]] = float(r["mean"])
+    return sorted(by_country), by_country
+
+
+def plot_mfv_map(run_dir: Path, out: Path, vec_label: str, C: float) -> Path:
+    """Bespoke MFV map: per-foundation RELATIVE EMPHASIS (z across foundations) of the model's base /
+    +C / -C reads against the human MFV cultures. MFV is nominal (model emits logit(violation) per
+    foundation, humans rate wrongness 1-5), so absolute scales differ; z-scoring each profile within
+    itself compares the PATTERN -- which foundations a reader weights as more violation-worthy than
+    their own average -- which is exactly what the steer is meant to move. Social Norms is dropped (no
+    human norm). The steer shows as base->+C (red) and base->-C (blue) arrows per foundation."""
+    d = json.loads((run_dir / "mfv.json").read_text())
+    base_l = d["base_logit_per_foundation"]
+    pos_dl, neg_dl = d["pos"]["dlogit_per_foundation"], d["neg"]["dlogit_per_foundation"]
+    countries, human = read_human_mfv()
+    hfounds = set(next(iter(human.values())))
+    founds = [f for f in d["foundation_order"] if f.lower() in hfounds]   # 6 shared, model order
+    fl = [f.lower() for f in founds]
+
+    base = _zscore(np.array([base_l[f]["mean"] for f in founds]))
+    posz = _zscore(np.array([base_l[f]["mean"] + pos_dl[f]["mean"] for f in founds]))
+    negz = _zscore(np.array([base_l[f]["mean"] + neg_dl[f]["mean"] for f in founds]))
+    Hz = {c: _zscore(np.array([human[c][f] for f in fl])) for c in countries}
+
+    rng = np.random.default_rng(0)
+    fig, ax = plt.subplots(figsize=(7.2, 4.6))
+    ax.axhline(0, color="0.85", lw=0.8, zorder=0)
+    POS, NEG, GREY = T.maps.POS_COL, T.maps.NEG_COL, T.maps.COUNTRY_GREY
+    for i, f in enumerate(founds):
+        hvals = np.array([Hz[c][i] for c in countries])
+        ax.scatter(i - 0.18 + (rng.random(len(hvals)) - 0.5) * 0.12, hvals, s=26, color=GREY,
+                   alpha=0.9, edgecolor="white", linewidth=0.3, zorder=3)
+        ax.plot([i - 0.30, i - 0.06], [np.median(hvals)] * 2, color=T.maps.MEDIAN_GREY, lw=1.4, zorder=4)
+        xs = i + 0.18
+        ax.plot(xs, base[i], "o", ms=4, color="black", zorder=7)
+        for pole, col in [(posz[i], POS), (negz[i], NEG)]:
+            if abs(pole - base[i]) > 1e-9:
+                ax.plot([xs, xs], [base[i], pole], color=col, lw=2.0, zorder=6, solid_capstyle="round")
+                ax.plot(xs, pole, marker=("^" if pole >= base[i] else "v"), color=col, ms=7,
+                        markeredgecolor="none", zorder=8)
+    ax.scatter([], [], marker="o", color=GREY, label=f"human culture (n={len(countries)})")
+    ax.scatter([], [], marker="o", color="black", label="model base")
+    ax.scatter([], [], marker="^", color=POS, label=f"steer +C={C:+.2f}")
+    ax.scatter([], [], marker="v", color=NEG, label=f"steer -C={-C:+.2f}")
+    ax.legend(fontsize=7.5, loc="lower right", framealpha=0.9, ncol=2)
+    ax.set_xticks(range(len(founds)))
+    ax.set_xticklabels(founds, rotation=20, ha="right", fontsize=8)
+    ax.set_xlim(-0.6, len(founds) - 0.4)
+    ax.set_ylabel("relative emphasis  (z across foundations)")
+    ax.set_title(f"MFV foundation emphasis vs human cultures: {vec_label}", fontsize=10)
+    ax.spines[["top", "right"]].set_visible(False)
+    fig.tight_layout()
+    path = T.maps.save_both(fig, out / "mfv", "map_emphasis")
+    plt.close(fig)
+    return path
+
+
 def plot_mfv(run_dir: Path, out: Path, vec_label: str, C: float) -> Path:
     """Per-foundation Delta-logit dumbbell: each foundation's +C (red) and -C (blue) shift vs bare."""
     d = json.loads((run_dir / "mfv.json").read_text())
@@ -179,6 +249,7 @@ def main() -> None:
         if (args.run_dir / f"{name}_profiles.csv").exists():
             written += [str(p) for p in plot_ordinal(args.run_dir, args.out, name, vec_label, C)]
     if (args.run_dir / "mfv.json").exists():
+        written.append(str(plot_mfv_map(args.run_dir, args.out, vec_label, C)))
         written.append(str(plot_mfv(args.run_dir, args.out, vec_label, C)))
     print(f"wrote {len(written)} figures under {args.out}:")
     for w in written: