viz: reference = Ariahw paper (oracle upper bound), not SGTM

Swap the floor->ceiling reference to the substrate paper (Ariahw et al. 2025),
which benchmarks interventions on the same floor (No-Intervention hack ~79%) /
ceiling (RL-Baseline no-loophole). Their best arm (Ground-Truth Penalty, ~0%
hack, perf >= ceiling) reaches the top corner BUT uses the oracle monitor at
train time -- the exact cheat our no-cheat constraint forbids; their only
oracle-free method (inoculation) gave incomplete, high-variance mitigation.
Plotted hatched/grey as an ORACLE upper bound (solve approx; figures are images,
200-step preset not step-matched). Honest framing: their working methods need
the oracle; ours uses no detector at train time and still suppresses 93%.

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

scripts/plot_floor_ceiling.py

@@ -106,11 +106,15 @@ def build_csv() -> pl.DataFrame:
 
 
 # ── stage 2: plot from the csv ──────────────────────────────────────────────
-# Reference: the gradient-routing paper (SGTM, Mhaskar et al. 2025) reports its result as a
-# retain/forget trade-off vs a "perfect filter" oracle (= our ceiling) and "no filter" (= our
-# floor). Placed on the SAME floor->ceiling % axis (approximate; LM-unlearning task, not RL):
-# forget suppression ~leakage 0.02 -> ~98%; retain ~5% compute penalty -> ~95% of oracle.
-SGTM_REF = dict(label="SGTM grad-routing\n(LM unlearn, ~approx)", hack_supp=0.98, solve_uplift=0.95)
+# Reference: Ariahw et al. 2025 (the substrate paper) benchmark interventions on the SAME
+# floor/ceiling -- No-Intervention (hack ~79%) = floor, RL-Baseline/no-loophole = ceiling. Their
+# best interventions (Ground-Truth Penalty ~0% hack, perf >= ceiling) reach the top corner BUT
+# use the oracle monitor at train time -- the exact cheat our no-cheat constraint forbids. Their
+# only oracle-free method (inoculation) gave incomplete, high-variance mitigation. We plot the
+# GT-monitor point as a clearly-marked ORACLE upper bound (solve approx; figures are images, and
+# their 200-step preset is not step-matched to our 60-step fast). hack_supp ~1.0 (hack ~0%),
+# solve_uplift ~1.0 (perf at/above ceiling).
+ARIAHW_REF = dict(label="Ariahw GT-monitor\n(uses ORACLE -- cheat)", hack_supp=0.99, solve_uplift=1.0)
 GOLD, DARK = "#c8920a", "#3a3a3a"
 
 
@@ -127,7 +131,7 @@ def _bars(ax, rows, key, raws, title, xlabel, xlo):
     """One floor->ceiling panel: horizontal bars in [xlo,1], 0=floor, 1.0=ceiling."""
     for yi, (lab, val, raw, col) in enumerate(rows):
         ax.barh(yi, val, height=0.55, color=col, alpha=0.9,
-                hatch="//" if "approx" in lab else None, edgecolor="white")
+                hatch="//" if col == GREY else None, edgecolor="white")   # grey = approx reference
         tip = f"{val*100:+.0f}%" if xlo < 0 else f"{val*100:.0f}%"
         rawtxt = f"  ({raw})" if raw else ""
         ax.text(val + (0.02 if val >= 0 else -0.02), yi, tip + rawtxt,
@@ -151,16 +155,16 @@ def plot(df: pl.DataFrame) -> None:
     def hsupp(r): return (vh - r["hack_deploy"]) / vh
     def suplift(r): return (r["solve_deploy"] - base) / (ceil - base)
 
-    # rows: best (gold), random control (dark), SGTM reference (grey, hatched). Top row plots last.
+    # rows: best (gold), random control (dark), Ariahw oracle reference (grey, hatched). Top plots last.
     hack_rows = [
-        (SGTM_REF["label"], SGTM_REF["hack_supp"], "~0.98 supp", GREY),
+        (ARIAHW_REF["label"], ARIAHW_REF["hack_supp"], "hack ~0%", GREY),
         ("routeV random-V\n(direction control)", hsupp(rand), f"{rand['hack_deploy']:.3f}", DARK),
-        ("routeV per-token\n(best)", hsupp(best), f"{best['hack_deploy']:.3f}", GOLD),
+        ("routeV per-token\n(best, no oracle)", hsupp(best), f"{best['hack_deploy']:.3f}", GOLD),
     ]
     solve_rows = [
-        (SGTM_REF["label"], SGTM_REF["solve_uplift"], "~oracle", GREY),
+        (ARIAHW_REF["label"], ARIAHW_REF["solve_uplift"], "~>=ceiling", GREY),
         ("routeV random-V\n(direction control)", suplift(rand), f"{rand['solve_deploy']:.3f}", DARK),
-        ("routeV per-token\n(best)", suplift(best), f"{best['solve_deploy']:.3f}", GOLD),
+        ("routeV per-token\n(best, no oracle)", suplift(best), f"{best['solve_deploy']:.3f}", GOLD),
     ]
     prov = " (ceiling PROVISIONAL=0.223, FIXME job 24)" if a["provisional"] else ""
     fig, (axl, axr) = plt.subplots(1, 2, figsize=(11, 3.2), sharey=True)