"""Phase 0.10 figures — traditional Plotly + cyberpunk themes. Generates 5 figures × 2 themes = 10 PNGs into figures//, plus interactive HTML copies into figures/interactive//. Run: `python generate_figures.py` """ import os import json from pathlib import Path import numpy as np import plotly.graph_objects as go from plotly.subplots import make_subplots HERE = Path(__file__).parent FIG_DIR = HERE / "figures" TRAD_DIR = FIG_DIR / "traditional" CYBER_DIR = FIG_DIR / "cyberpunk" INTERACTIVE_DIR = FIG_DIR / "interactive" for d in [TRAD_DIR, CYBER_DIR, INTERACTIVE_DIR, INTERACTIVE_DIR / "traditional", INTERACTIVE_DIR / "cyberpunk"]: d.mkdir(parents=True, exist_ok=True) # ── Theme dictionaries ────────────────────────────────────────────────── TRADITIONAL = { "name": "traditional", "template": "plotly_white", "bg": "white", "plot_bg": "white", "text": "#2a2a2a", "grid": "#e8e8e8", "axis": "#444444", "primary": "#1f77b4", # plotly blue — headline color "secondary": "#ff7f0e", # plotly orange — Round Two improvements "accent1": "#2ca02c", # plotly green — positive / strong-model band "accent2": "#d62728", # plotly red (unused after warning split) "warning": "#d62728", # red — negative result (§4.B prune hurt AUC) "neutral": "#7f7f7f", # gray "band": "rgba(44, 160, 44, 0.10)", # light green band for strong-model "band_edge": "rgba(44, 160, 44, 0.5)", "band_text": "#1d6b1d", # dark green for band annotation "first_post": "#888888", "font_family": "Arial, Helvetica, sans-serif", "title_size": 16, "axis_label_size": 13, "tick_size": 11, "marker_size": 12, "line_width": 2.5, } CYBERPUNK = { "name": "cyberpunk", "template": "plotly_dark", "bg": "#0a0a14", "plot_bg": "#14141f", "text": "#e5e5ff", "grid": "#2a2a44", "axis": "#5a5a8a", "primary": "#ff006e", # hot pink — headline color "secondary": "#00ffff", # cyan — Round Two improvements "accent1": "#fbff12", # neon yellow "accent2": "#00ff41", # matrix green — positive / strong-model band "warning": "#ff5500", # neon orange — negative result (§4.B prune hurt AUC) "neutral": "#9d9dff", # lavender "band": "rgba(0, 255, 65, 0.10)", "band_edge": "rgba(0, 255, 65, 0.6)", "band_text": "#00ff41", # matrix green for band annotation "first_post": "#7d7da8", "font_family": "Courier New, monospace", "title_size": 16, "axis_label_size": 13, "tick_size": 11, "marker_size": 14, "line_width": 3.0, } def style_layout(fig, theme, title=None, xaxis_title=None, yaxis_title=None, height=480, width=900, legend_y=1.02): """Apply common layout styling to a figure under one theme.""" fig.update_layout( template=theme["template"], paper_bgcolor=theme["bg"], plot_bgcolor=theme["plot_bg"], font=dict(family=theme["font_family"], color=theme["text"], size=theme["tick_size"]), title=dict( text=title or "", font=dict(size=theme["title_size"], color=theme["text"]), x=0.02, xanchor="left", ) if title else None, xaxis=dict( title=dict(text=xaxis_title or "", font=dict(size=theme["axis_label_size"])), gridcolor=theme["grid"], linecolor=theme["axis"], tickcolor=theme["axis"], zerolinecolor=theme["grid"], ), yaxis=dict( title=dict(text=yaxis_title or "", font=dict(size=theme["axis_label_size"])), gridcolor=theme["grid"], linecolor=theme["axis"], tickcolor=theme["axis"], zerolinecolor=theme["grid"], ), height=height, width=width, margin=dict(t=60 if title else 30, l=70, r=30, b=60), legend=dict( orientation="h", yanchor="bottom", y=legend_y, xanchor="right", x=1, font=dict(size=theme["tick_size"]), bgcolor="rgba(0,0,0,0)", ), ) return fig def save_fig(fig, theme, name): """Save fig as PNG into figures//.png and HTML into figures/interactive//.html.""" png_path = FIG_DIR / theme["name"] / f"{name}.png" html_path = FIG_DIR / "interactive" / theme["name"] / f"{name}.html" fig.write_image(str(png_path), scale=2) fig.write_html(str(html_path), include_plotlyjs="cdn") print(f" saved {png_path.name} + interactive HTML") # ── Figure 1: AUC trajectory ──────────────────────────────────────────── AUC_PHASES = [ {"label": "First post
(20K, K=3, raw)", "auc": 0.5760, "n": "20K", "kind": "first_baseline"}, {"label": "First post
(20K, K=2 calibrated)", "auc": 0.6107, "n": "20K", "kind": "first_published"}, {"label": "§4.A
(20K, K=2, prior=0.5)", "auc": 0.6312, "n": "20K", "kind": "round_two"}, {"label": "§4.B
(20K, K=2, pruned)", "auc": 0.6194, "n": "20K", "kind": "negative"}, {"label": "§4.E
(101K, K=2, seed 42)", "auc": 0.6621, "n": "101K", "kind": "round_two"}, {"label": "§4.F
(101K, K=2, 3-seed mean)", "auc": 0.6645, "auc_std": 0.0018, "n": "101K", "kind": "headline"}, ] def fig_auc_trajectory(theme): fig = go.Figure() # Strong-model band 0.64-0.68 (shaded region) fig.add_shape(type="rect", x0=-0.5, x1=len(AUC_PHASES) - 0.5, y0=0.64, y1=0.68, line=dict(color=theme["band_edge"], width=1, dash="dot"), fillcolor=theme["band"], layer="below") fig.add_annotation(x=0.05, y=0.685, text="Strong-published-model band (0.64–0.68)", showarrow=False, xanchor="left", yanchor="bottom", font=dict(color=theme["band_text"], size=11, family=theme["font_family"]), bgcolor=theme["plot_bg"], borderpad=3) # Coin-flip baseline fig.add_hline(y=0.5, line=dict(color=theme["neutral"], width=1, dash="dot"), annotation_text="Random (0.50)", annotation_position="bottom right", annotation=dict(font=dict(color=theme["neutral"], size=11))) # First post calibrated headline fig.add_hline(y=0.611, line=dict(color=theme["first_post"], width=1, dash="dash"), annotation_text="First post calibrated (0.611)", annotation_position="top right", annotation=dict(font=dict(color=theme["first_post"], size=11))) # Bars / markers xs = list(range(len(AUC_PHASES))) labels = [p["label"] for p in AUC_PHASES] aucs = [p["auc"] for p in AUC_PHASES] colors = [] for p in AUC_PHASES: if p["kind"] == "first_baseline": colors.append(theme["neutral"]) elif p["kind"] == "first_published": colors.append(theme["first_post"]) elif p["kind"] == "negative": colors.append(theme["warning"]) # negative-result signal in both themes elif p["kind"] == "headline": colors.append(theme["primary"]) else: colors.append(theme["secondary"]) fig.add_trace(go.Bar( x=xs, y=aucs, text=[f"{a:.4f}" for a in aucs], textposition="outside", textfont=dict(color=theme["text"], size=12), marker=dict(color=colors, line=dict(color=theme["axis"], width=1)), showlegend=False, hovertemplate="%{customdata}
AUC: %{y:.4f}", customdata=labels, )) # Add error bars for the 3-seed result headline_idx = next(i for i, p in enumerate(AUC_PHASES) if p["kind"] == "headline") fig.add_trace(go.Scatter( x=[xs[headline_idx]], y=[AUC_PHASES[headline_idx]["auc"]], error_y=dict(type="data", array=[AUC_PHASES[headline_idx]["auc_std"]], color=theme["accent1"], thickness=3, width=6), mode="markers", marker=dict(size=0, color=theme["accent1"]), showlegend=False, hoverinfo="skip", )) fig.update_layout( xaxis=dict(tickmode="array", tickvals=xs, ticktext=labels, tickangle=0), yaxis=dict(range=[0.45, 0.72]), bargap=0.30, ) style_layout(fig, theme, title="Held-out test AUC trajectory — diabetes-130 readmission", xaxis_title="", yaxis_title="AUC", height=540, width=1080) return fig # ── Figure 2: Multi-seed variance ─────────────────────────────────────── MULTI_SEED = { "AUC": {"seeds": [0.6621, 0.6664, 0.6649], "mean": 0.6645, "std": 0.0018, "range_lo": 0.660, "range_hi": 0.670}, "Brier": {"seeds": [0.0953, 0.0948, 0.0950], "mean": 0.0950, "std": 0.0002, "range_lo": 0.094, "range_hi": 0.096}, "NLL": {"seeds": [0.3349, 0.3340, 0.3347], "mean": 0.3346, "std": 0.0004, "range_lo": 0.333, "range_hi": 0.336}, "ECE": {"seeds": [0.0044, 0.0030, 0.0041], "mean": 0.0038, "std": 0.0006, "range_lo": 0.000, "range_hi": 0.008}, "Bal. accuracy": {"seeds": [0.5012, 0.5015, 0.5030], "mean": 0.5019, "std": 0.0008, "range_lo": 0.500, "range_hi": 0.504}, "F1 @ 0.5": {"seeds": [0.0061, 0.0070, 0.0130], "mean": 0.0087, "std": 0.0031, "range_lo": 0.000, "range_hi": 0.016}, } SEED_LABELS = ["seed 42", "seed 43", "seed 44"] def fig_multi_seed(theme): metrics = list(MULTI_SEED.keys()) fig = make_subplots(rows=2, cols=3, subplot_titles=[f"{m}" for m in metrics], horizontal_spacing=0.10, vertical_spacing=0.18) seed_colors = [theme["primary"], theme["secondary"], theme["accent1"]] for i, metric in enumerate(metrics): row = i // 3 + 1 col = i % 3 + 1 d = MULTI_SEED[metric] # Mean ± std band fig.add_shape(type="rect", x0=-0.5, x1=2.5, y0=d["mean"] - d["std"], y1=d["mean"] + d["std"], line=dict(color=theme["band_edge"], width=1, dash="dot"), fillcolor=theme["band"], layer="below", row=row, col=col) # Mean line fig.add_trace(go.Scatter( x=[-0.5, 2.5], y=[d["mean"], d["mean"]], mode="lines", line=dict(color=theme["neutral"], width=2, dash="dash"), showlegend=False, hoverinfo="skip", ), row=row, col=col) # Per-seed dots fig.add_trace(go.Scatter( x=[0, 1, 2], y=d["seeds"], mode="markers+text", text=[f"{v:.4f}" for v in d["seeds"]], textposition="top center", textfont=dict(color=theme["text"], size=10), marker=dict(size=theme["marker_size"], color=seed_colors, line=dict(color=theme["axis"], width=1.5)), showlegend=False, hovertemplate="%{customdata}
%{y:.4f}", customdata=SEED_LABELS, ), row=row, col=col) # Update axes per subplot fig.update_xaxes(tickmode="array", tickvals=[0, 1, 2], ticktext=SEED_LABELS, range=[-0.6, 2.6], gridcolor=theme["grid"], linecolor=theme["axis"], row=row, col=col) fig.update_yaxes(range=[d["range_lo"], d["range_hi"]], gridcolor=theme["grid"], linecolor=theme["axis"], row=row, col=col) fig.update_annotations(font=dict(color=theme["text"], size=12)) fig.update_layout( template=theme["template"], paper_bgcolor=theme["bg"], plot_bgcolor=theme["plot_bg"], font=dict(family=theme["font_family"], color=theme["text"], size=theme["tick_size"]), title=dict( text="§4.F multi-seed variance — full 101K, K=2, prior=0.5", font=dict(size=theme["title_size"], color=theme["text"]), x=0.02, xanchor="left", ), height=620, width=1080, margin=dict(t=70, l=60, r=30, b=40), showlegend=False, ) return fig # ── Figure 3: Per-leaf belief — naive vs ?-aware ──────────────────────── LEAVES = [ {"id": 1, "n_rows": 9617, "naive_miss": 1.0000, "aware_miss": 0.9621}, {"id": 2, "n_rows": 6215, "naive_miss": 1.0000, "aware_miss": 0.9624}, {"id": 3, "n_rows": 2545, "naive_miss": 1.0000, "aware_miss": 0.9621}, {"id": 4, "n_rows": 1623, "naive_miss": 1.0000, "aware_miss": 0.9613}, ] def fig_per_leaf_belief(theme): fig = go.Figure() labels = [f"Leaf {l['id']}
n={l['n_rows']:,}" for l in LEAVES] naive = [l["naive_miss"] for l in LEAVES] aware = [l["aware_miss"] for l in LEAVES] fig.add_trace(go.Bar( name="Naive (default Locke)", x=labels, y=naive, text=[f"{v:.4f}" for v in naive], textposition="outside", textfont=dict(color=theme["text"], size=11), marker=dict(color=theme["neutral"], line=dict(color=theme["axis"], width=1)), hovertemplate="%{x}
Naive missingness_score: %{y:.4f}", )) fig.add_trace(go.Bar( name="?-aware (with NullMaskMap)", x=labels, y=aware, text=[f"{v:.4f}" for v in aware], textposition="outside", textfont=dict(color=theme["text"], size=11), marker=dict(color=theme["primary"], line=dict(color=theme["axis"], width=1)), hovertemplate="%{x}
?-aware missingness_score: %{y:.4f}", )) # Reference line at 1.0 fig.add_hline(y=1.0, line=dict(color=theme["accent1"], width=1.2, dash="dot")) # Annotation lifted into the headroom above the y=1.0 line so it # doesn't collide with the "1.0000" bar-top labels. fig.add_annotation(x=0.02, y=1.018, xref="paper", text="Perfect (no missingness detected)", showarrow=False, xanchor="left", yanchor="middle", font=dict(color=theme["accent1"], size=11, family=theme["font_family"]), bgcolor=theme["plot_bg"], borderpad=3) fig.update_layout( barmode="group", yaxis=dict(range=[0.94, 1.030]), bargap=0.30, bargroupgap=0.05, ) style_layout(fig, theme, title="§4.D Part 1 — Locke per-leaf missingness, naive vs ?-aware", xaxis_title="", yaxis_title="missingness_score (1.0 = perfect)", height=480, width=1000) return fig # ── Figure 4: Calibration comparison — first post calibrated vs Round Two raw ── CALIBRATION = [ {"metric": "Brier", "first_cal": 0.0980, "round_two_raw": 0.0950, "lower_is_better": True, "base_rate": 0.099}, {"metric": "NLL", "first_cal": 0.3435, "round_two_raw": 0.3346, "lower_is_better": True, "base_rate": 0.351}, {"metric": "ECE", "first_cal": 0.0101, "round_two_raw": 0.0038, "lower_is_better": True, "base_rate": 0.000}, ] def fig_calibration(theme): fig = make_subplots(rows=1, cols=3, subplot_titles=[f"{c['metric']}" for c in CALIBRATION], horizontal_spacing=0.12) for i, c in enumerate(CALIBRATION): col = i + 1 labels = ["First post
(Platt calibrated, 20K)", "Round Two
(raw, 101K, n=3)"] vals = [c["first_cal"], c["round_two_raw"]] colors = [theme["first_post"], theme["primary"]] # Bar value labels INSIDE the bars so the base-rate annotation # (which lives above the bars near the line) doesn't collide # with them. For ECE the bars are tall enough that "inside" # reads cleanly; for Brier/NLL the bars are also tall enough # because the values are well above zero. label_color = "white" if theme["name"] == "cyberpunk" else "white" fig.add_trace(go.Bar( x=labels, y=vals, text=[f"{v:.4f}" for v in vals], textposition="inside", insidetextanchor="end", textfont=dict(color=label_color, size=12, family=theme["font_family"]), marker=dict(color=colors, line=dict(color=theme["axis"], width=1)), showlegend=False, hovertemplate="%{x}
%{y:.4f}", ), row=1, col=col) # Y-axis headroom — generous so the base-rate annotation, the # base-rate line, and the bar-top labels all have room above the # bars. max_v = max(vals + [c["base_rate"]]) * 1.55 fig.update_yaxes(range=[0, max_v], gridcolor=theme["grid"], linecolor=theme["axis"], row=1, col=col) fig.update_xaxes(gridcolor=theme["grid"], linecolor=theme["axis"], row=1, col=col) # Base-rate baseline — drawn AFTER the y-axis range is set so the # annotation can be parked at the top of the panel, well away # from the bar tops and their value labels. if c["base_rate"] > 0: fig.add_shape(type="line", x0=-0.5, x1=1.5, y0=c["base_rate"], y1=c["base_rate"], line=dict(color=theme["secondary"], width=1.5, dash="dash"), row=1, col=col) # Pin the annotation to the top of the panel with a small # downward arrow pointing at the base-rate line. Avoids # collision with bar-top labels even when bars and base-rate # are at nearly the same y. fig.add_annotation(x=1.4, y=max_v * 0.95, ax=1.4, ay=c["base_rate"], axref=f"x{col if col > 1 else ''}", ayref=f"y{col if col > 1 else ''}", xref=f"x{col if col > 1 else ''}", yref=f"y{col if col > 1 else ''}", text=f"base rate
{c['base_rate']:.3f}", showarrow=True, arrowhead=2, arrowsize=1, arrowwidth=1.5, arrowcolor=theme["secondary"], xanchor="center", yanchor="middle", font=dict(color=theme["secondary"], size=10, family=theme["font_family"]), bgcolor=theme["plot_bg"], borderpad=3, bordercolor=theme["secondary"]) fig.update_annotations(font=dict(color=theme["text"], size=13)) fig.update_layout( template=theme["template"], paper_bgcolor=theme["bg"], plot_bgcolor=theme["plot_bg"], font=dict(family=theme["font_family"], color=theme["text"], size=theme["tick_size"]), title=dict( text="Calibration metrics — first post (Platt calibrated) vs Round Two (raw)", font=dict(size=theme["title_size"], color=theme["text"]), x=0.02, xanchor="left", ), height=500, width=1080, margin=dict(t=70, l=60, r=30, b=80), bargap=0.30, ) return fig # ── Figure 5: MI routing-feature selection at 20K vs 101K ────────────── MI_ROUTING = { "20K trial (§4.A)": { "features": ["num_inpatient", "time_in_hospital", "number_diagnoses", "number_emergency", "num_lab_procedures"], "selected": [True, True, False, False, False], "ranks": [1, 2, 3, 4, 5], # arbitrary rank — illustrative }, "101K full (§4.E)": { "features": ["num_inpatient", "time_in_hospital", "number_diagnoses", "number_emergency", "num_lab_procedures"], "selected": [True, False, False, True, False], "ranks": [1, 3, 4, 2, 5], }, } def fig_mi_routing(theme): fig = make_subplots(rows=1, cols=2, subplot_titles=[f"{k}" for k in MI_ROUTING.keys()], horizontal_spacing=0.18) for i, (label, d) in enumerate(MI_ROUTING.items()): col = i + 1 ranks = d["ranks"] # Sort by rank for plot order = sorted(range(len(ranks)), key=lambda k: ranks[k]) feats = [d["features"][k] for k in order] sel = [d["selected"][k] for k in order] # Show as horizontal bars with rank position bar_colors = [theme["primary"] if s else theme["neutral"] for s in sel] bar_text = ["SELECTED" if s else "" for s in sel] fig.add_trace(go.Bar( y=feats, x=list(range(len(feats), 0, -1)), text=bar_text, textposition="inside", textfont=dict(color=theme["text"], size=11), orientation="h", marker=dict(color=bar_colors, line=dict(color=theme["axis"], width=1)), showlegend=False, hovertemplate="%{y}
MI rank: %{customdata}", customdata=[r for r in sorted(ranks)], ), row=1, col=col) fig.update_xaxes(title=dict(text="MI rank (higher = better)", font=dict(size=12, color=theme["text"])), showticklabels=False, gridcolor=theme["grid"], linecolor=theme["axis"], row=1, col=col) fig.update_yaxes(gridcolor=theme["grid"], linecolor=theme["axis"], tickfont=dict(family="Courier New, monospace", size=11), row=1, col=col) fig.update_annotations(font=dict(color=theme["text"], size=13)) fig.update_layout( template=theme["template"], paper_bgcolor=theme["bg"], plot_bgcolor=theme["plot_bg"], font=dict(family=theme["font_family"], color=theme["text"], size=theme["tick_size"]), title=dict( text="MI-selected routing features — scale dependence", font=dict(size=theme["title_size"], color=theme["text"]), x=0.02, xanchor="left", ), height=420, width=1080, margin=dict(t=70, l=20, r=20, b=60), bargap=0.25, annotations=list(fig.layout.annotations) + [ dict(text="time_in_hospital wins at 20K → number_emergency wins at 101K", xref="paper", yref="paper", x=0.5, y=-0.18, showarrow=False, font=dict(color=theme["neutral"], size=11)), ], ) return fig # ── Driver ────────────────────────────────────────────────────────────── THEMES = [TRADITIONAL, CYBERPUNK] FIGURES = [ ("auc_trajectory", fig_auc_trajectory), ("multi_seed_variance", fig_multi_seed), ("per_leaf_belief_question_marks", fig_per_leaf_belief), ("calibration_first_vs_round_two", fig_calibration), ("mi_routing_scale", fig_mi_routing), ] def main(): for theme in THEMES: print(f"\n[{theme['name']}]") for name, builder in FIGURES: fig = builder(theme) save_fig(fig, theme, name) print("\ndone.") if __name__ == "__main__": main()