viz: floor-to-ceiling method comparison (csv + figure)

Two-stage script: build out/plots/floor_ceiling.csv (one row per arm/anchor,
with SOURCE and STATUS columns flagging every provisional/missing cell) then
the keynote figure. Prints TODO/FIXME data gaps before plotting.

Panel A: normalized floor->ceiling bars, headline deploy (knob-off, test n=119).
Panel B: the knob effect -- arrow knob-ON -> knob-OFF on the SAME held-out val
split (eval_curve.jsonl), isolating the quarantine from the train/test
memorization gap. Fixes the earlier conflation where the train->deploy arrow
mixed knob-on/off with train-problems/test-problems.

Data gaps flagged in csv: solve ceiling provisional=paper 0.223 (FIXME job 24),
prog_wide arm contaminated (TODO job 28 prog_wide_clean).

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

justfile

@@ -281,6 +281,12 @@ plot GLOB='logs/*_sub4_*.log' STEM='out/figs/substrate':
 plot-deploy GLOB='out/runs/*sub4*/per_mode_deploy.json' OUT='out/figs/deploy_overlay.png':
     uv run python scripts/plot_deploy_overlay.py {{ GLOB }} --out {{ OUT }}
 
+# Keynote floor->ceiling method comparison. Builds out/plots/floor_ceiling.csv
+# (inspectable, with SOURCE + STATUS/TODO columns) then the figure. Prints any
+# provisional/missing cells (ceiling = job 24, prog_wide clean = job 28).
+plot-floor-ceiling:
+    uv run python -m scripts.plot_floor_ceiling
+
 # Regenerate both dynamics plots from the cell logs (default: all cells; pass a
 # narrower glob like 'logs/*_cell_*_s41.log' for the seed-41-only checkpoint).
 regen-dynamics GLOB='logs/*_cell_*.log':

out/plots/floor_ceiling.csv

@@ -0,0 +1,8 @@
+label,kind,hack_deploy,solve_deploy,hack_on,hack_off,solve_on,solve_off,source,status
+routeV per-token,method,0.042,0.1429,0.6312,0.025,0.0688,0.0688,20260607T134234_fast_routingV_seed43_dir6_routeV_pertoken_s43/[deploy_test.json + eval_curve.jsonl],ok
+routeV authored,method,0.0756,0.1176,0.6687,0.0187,0.0563,0.0437,20260608T134141_fast_routingV_seed43_dir8_routeV_authored_perroll_s43/[deploy_test.json + eval_curve.jsonl],ok
+routeV prog_wide,method,0.1008,0.1261,0.6937,0.0125,0.0688,0.0563,20260607T195125_fast_routingV_seed43_dir6_routeV_s43/[deploy_test.json + eval_curve.jsonl],TODO: contaminated pairs -> job 28 prog_wide_clean
+routeV random-V,method,0.1008,0.1092,0.7,0.0437,0.075,0.0688,20260608T020623_fast_routingV_seed43_dir6_routeV_random_s43/[deploy_test.json + eval_curve.jsonl],ok (directionality control)
+vanilla GRPO,method,0.6134,0.1008,0.5938,0.5938,0.075,0.075,20260608T224659_fast_vanilla_seed43_dir8_vanilla_s43/[deploy_test.json + eval_curve.jsonl],ok (defines hack-worst anchor)
+base (floor),anchor_floor,0.0,0.1261,,,,,*_dir8_baseline_s43/deploy_test.json,ok (base model; steps=0)
+ceiling,anchor_ceiling,0.0,0.223,,,,,"Ariahw et al. 2025 (paper), NOT our run",FIXME: PROVISIONAL paper 0.223 -- awaiting job 24 (no-loophole ceiling)

scripts/plot_floor_ceiling.py

@@ -0,0 +1,195 @@
+"""Floor-to-ceiling method comparison: the keynote figure.
+
+Two stages so the data is inspectable before it's drawn:
+  1. build  -> out/plots/floor_ceiling.csv   (one row per arm/anchor, with SOURCE and STATUS
+                columns; every provisional/missing value is flagged, not silently filled)
+  2. plot   -> out/plots/floor_ceiling.{pdf,png}
+
+Run `uv run python -m scripts.plot_floor_ceiling` to do both; it prints a TODO/FIXME summary
+of any provisional or missing cells before plotting.
+
+THE GOAL: place each gradient-routing arm on a floor->ceiling scale so "how much of the
+achievable range did it capture" is read at a glance, and show that the quarantine (knob)
+is what removes the hack, not a train/test artifact.
+
+TWO METRICS, two anchor pairs (right/down = better):
+  hack removed    = (vanilla_hack - arm_hack) / vanilla_hack            1.0 = no hack
+  solve recovered = (arm_solve - base_solve)   / (ceiling - base_solve) 1.0 = no-loophole ceiling
+
+TWO VIEWS of the same arms:
+  A. normalized floor->ceiling bars, HEADLINE deploy (knob-off, test n=119, recency-clean).
+     Source per arm: out/runs/<run>/deploy_test.json.
+  B. the KNOB effect: arrow knob-ON -> knob-OFF on the SAME held-out val split (n=32), so it
+     isolates the quarantine from the train/test memorization gap. Source per arm:
+     out/runs/<run>/eval_curve.jsonl, where the file's `train_*`/`deploy_*` prefixes denote
+     KNOB STATE (on/off), not the problem set (always val here). L5 = mean of last 5 evals.
+
+DATA GAPS (see STATUS column in the csv):
+  - solve ceiling: provisional = paper 0.223 until job 24 (out/runs/*noloophole*) lands. FIXME.
+  - prog_wide arm uses contaminated pairs; job 28 (prog_wide_clean) will replace it. TODO.
+  - full-env (paper-scale) panel: no method runs exist, only paper anchors. Out of scope here.
+"""
+from __future__ import annotations
+import json
+from pathlib import Path
+
+import polars as pl
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+
+RED, GREEN, GREY = "#c0392b", "#1e8449", "#9aa0a6"
+RUNS = Path("out/runs")
+OUT = Path("out/plots")
+CSV = OUT / "floor_ceiling.csv"
+PAPER_CEILING = 0.223          # Ariahw et al. no-loophole solve -- provisional fast-env ceiling
+
+# arm display order, identified by a substring of the run's out_tag (seed-43 fast runs)
+ARMS = [
+    ("routeV per-token", "_dir6_routeV_pertoken_s43",        "ok"),
+    ("routeV authored",  "_dir8_routeV_authored_perroll_s43", "ok"),
+    ("routeV prog_wide", "_dir6_routeV_s43",                 "TODO: contaminated pairs -> job 28 prog_wide_clean"),
+    ("routeV random-V",  "_dir6_routeV_random_s43",          "ok (directionality control)"),
+    ("vanilla GRPO",     "_dir8_vanilla_s43",                "ok (defines hack-worst anchor)"),
+]
+
+
+def _find_run(tag: str) -> Path:
+    cands = sorted(d for d in RUNS.iterdir()
+                   if d.name.endswith(tag) and (d / "deploy_test.json").exists())
+    if not cands:
+        raise FileNotFoundError(f"no run dir ending '{tag}' with a deploy_test.json")
+    return cands[-1]                 # latest timestamp wins
+
+
+def _l5(rows: list[dict], k: str) -> float:
+    v = [r[k] for r in rows[-5:]]
+    return sum(v) / len(v)
+
+
+# ── stage 1: build the inspectable csv ──────────────────────────────────────
+def build_csv() -> pl.DataFrame:
+    rows = []
+    for label, tag, status in ARMS:
+        run = _find_run(tag)
+        dep = json.loads((run / "deploy_test.json").read_text())
+        ev = [json.loads(l) for l in (run / "eval_curve.jsonl").read_text().splitlines()]
+        rows.append(dict(
+            label=label, kind="method",
+            hack_deploy=round(dep["deploy_hack"], 4), solve_deploy=round(dep["deploy_solve"], 4),
+            hack_on=round(_l5(ev, "train_hack"), 4),  hack_off=round(_l5(ev, "deploy_hack"), 4),
+            solve_on=round(_l5(ev, "train_solve"), 4), solve_off=round(_l5(ev, "deploy_solve"), 4),
+            source=f"{run.name}/[deploy_test.json + eval_curve.jsonl]", status=status))
+
+    base = json.loads((_find_run("_dir8_baseline_s43") / "deploy_test.json").read_text())
+    rows.append(dict(label="base (floor)", kind="anchor_floor",
+                     hack_deploy=round(base["deploy_hack"], 4), solve_deploy=round(base["deploy_solve"], 4),
+                     hack_on=None, hack_off=None, solve_on=None, solve_off=None,
+                     source="*_dir8_baseline_s43/deploy_test.json", status="ok (base model; steps=0)"))
+
+    ceil_path = next(RUNS.glob("*noloophole*/deploy_test.json"), None)
+    if ceil_path:
+        ceil_solve, status = round(json.loads(ceil_path.read_text())["deploy_solve"], 4), "ok"
+        source = f"{ceil_path.parent.name}/deploy_test.json"
+    else:
+        ceil_solve, status = PAPER_CEILING, "FIXME: PROVISIONAL paper 0.223 -- awaiting job 24 (no-loophole ceiling)"
+        source = "Ariahw et al. 2025 (paper), NOT our run"
+    rows.append(dict(label="ceiling", kind="anchor_ceiling",
+                     hack_deploy=0.0, solve_deploy=ceil_solve,
+                     hack_on=None, hack_off=None, solve_on=None, solve_off=None,
+                     source=source, status=status))
+
+    df = pl.DataFrame(rows)
+    OUT.mkdir(parents=True, exist_ok=True)
+    df.write_csv(CSV)
+    return df
+
+
+# ── stage 2: plot from the csv ──────────────────────────────────────────────
+def _anchors(df: pl.DataFrame) -> dict:
+    g = lambda kind, col: df.filter(pl.col("kind") == kind)[col][0]
+    ceil_status = g("anchor_ceiling", "status")
+    return dict(base_solve=g("anchor_floor", "solve_deploy"),
+                vanilla_hack=df.filter(pl.col("label") == "vanilla GRPO")["hack_deploy"][0],
+                ceiling=g("anchor_ceiling", "solve_deploy"),
+                provisional=ceil_status.startswith("FIXME"))
+
+
+def _panel_normalized(ax, methods: pl.DataFrame, a, title):
+    base, vh, ceil = a["base_solve"], a["vanilla_hack"], a["ceiling"]
+    labels = [l for l in methods["label"] if l != "vanilla GRPO"]   # vanilla = the 0% hack anchor
+    for yi, lab in enumerate(labels):
+        r = methods.filter(pl.col("label") == lab).to_dicts()[0]
+        hack_rm = (vh - r["hack_deploy"]) / vh
+        solve_rc = (r["solve_deploy"] - base) / (ceil - base)
+        ax.barh(yi + 0.18, hack_rm, height=0.32, color=RED, alpha=0.85)
+        ax.text(hack_rm + 0.015, yi + 0.18, f"{r['hack_deploy']:.3f}  ({hack_rm*100:.0f}%)",
+                va="center", ha="left", fontsize=8, color=RED)
+        ax.barh(yi - 0.18, solve_rc, height=0.32, color=GREEN, alpha=0.85)
+        ax.text(solve_rc + 0.015 if solve_rc >= 0 else solve_rc - 0.015, yi - 0.18,
+                f"{r['solve_deploy']:.3f}  ({solve_rc*100:+.0f}%)",
+                va="center", ha="left" if solve_rc >= 0 else "right", fontsize=8, color=GREEN)
+    ax.axvline(0, color=GREY, lw=0.8)
+    ax.axvline(1.0, color=GREY, lw=0.8, ls=":")
+    ax.text(1.0, len(labels) - 0.35, "ceiling / no-hack", fontsize=7, color=GREY, ha="center")
+    ax.set_yticks(range(len(labels))); ax.set_yticklabels(labels, fontsize=9)
+    ax.set_xlim(-0.35, 1.25); ax.set_xlabel("fraction of floor→ceiling range  (right = better)")
+    ax.set_title(title, fontsize=10, loc="left")
+    ax.text(0.01, 0.99, "red = hack removed (vs vanilla)   green = solve recovered (base→ceiling)",
+            transform=ax.transAxes, fontsize=7.5, color="#444", va="top")
+    for s in ("top", "right", "left"):
+        ax.spines[s].set_visible(False)
+    ax.tick_params(left=False)
+
+
+def _panel_knob(ax, methods: pl.DataFrame):
+    labels = list(methods["label"])
+    for yi, lab in enumerate(labels):
+        r = methods.filter(pl.col("label") == lab).to_dicts()[0]
+        ax.annotate("", xy=(r["hack_off"], yi + 0.16), xytext=(r["hack_on"], yi + 0.16),
+                    arrowprops=dict(arrowstyle="->", color=RED, lw=1.6, alpha=0.9))
+        ax.plot([r["hack_on"], r["hack_off"]], [yi + 0.16] * 2, "o", color=RED, ms=4, alpha=0.5)
+        ax.text(r["hack_on"] + 0.012, yi + 0.16, f"on {r['hack_on']:.2f}", va="center", ha="left", fontsize=7, color=RED)
+        ax.text(r["hack_off"] - 0.012, yi + 0.16, f"{r['hack_off']:.2f}", va="center", ha="right", fontsize=7.5, color=RED)
+        ax.annotate("", xy=(r["solve_off"], yi - 0.16), xytext=(r["solve_on"], yi - 0.16),
+                    arrowprops=dict(arrowstyle="->", color=GREEN, lw=1.6, alpha=0.9))
+        ax.plot([r["solve_on"], r["solve_off"]], [yi - 0.16] * 2, "o", color=GREEN, ms=4, alpha=0.5)
+        ax.text(max(r["solve_on"], r["solve_off"]) + 0.012, yi - 0.16, f"solve {r['solve_off']:.2f}",
+                va="center", ha="left", fontsize=7.5, color=GREEN)
+    ax.set_yticks(range(len(labels))); ax.set_yticklabels(labels, fontsize=9)
+    ax.set_xlim(-0.02, 0.80)
+    ax.set_xlabel("rate  (arrow = knob-ON → knob-OFF on held-out val; left = better for hack)")
+    ax.set_title("B. the knob effect (held-out val n=32, L5 -- isolates the quarantine)", fontsize=10, loc="left")
+    for s in ("top", "right", "left"):
+        ax.spines[s].set_visible(False)
+    ax.tick_params(left=False)
+
+
+def plot(df: pl.DataFrame) -> None:
+    a = _anchors(df)
+    methods = df.filter(pl.col("kind") == "method")
+    prov = "  [ceiling PROVISIONAL=0.223, FIXME job 24]" if a["provisional"] else ""
+    fig, axes = plt.subplots(2, 1, figsize=(8.5, 8.0), gridspec_kw=dict(height_ratios=[1, 1.05]))
+    _panel_normalized(axes[0], methods, a, f"A. normalized floor→ceiling, deploy (test n=119){prov}")
+    _panel_knob(axes[1], methods)
+    fig.suptitle("vGROUT: floor-to-ceiling method comparison (seed 43, 60-step fast)",
+                 fontsize=11, x=0.02, ha="left")
+    fig.tight_layout(rect=(0, 0, 1, 0.97))
+    for ext in ("pdf", "png"):
+        fig.savefig(OUT / f"floor_ceiling.{ext}", dpi=150, bbox_inches="tight")
+
+
+def main() -> None:
+    df = build_csv()
+    flags = df.filter(~pl.col("status").str.starts_with("ok"))
+    print(f"wrote {CSV}")
+    if len(flags):
+        print("\n=== TODO/FIXME in data ===")
+        for r in flags.to_dicts():
+            print(f"  [{r['label']}] {r['status']}")
+    plot(df)
+    print(f"\nwrote {OUT}/floor_ceiling.pdf and .png")
+
+
+if __name__ == "__main__":
+    main()