scripts: add NEET 2026 chart generator

Reads live from results/*NEET_2026* and emits leaderboard, cost-vs-accuracy,
and per-subject heatmap PNGs into charts/neet_2026/. Charts directory is
gitignored — regenerate before publishing.

Co-Authored-By: Claude Opus 4.7 <[email protected]>

.gitignore

@@ -87,3 +87,6 @@ lerna-debug.log*
 # Other
 *.swp
 *~
+
+# Generated charts (regenerate via scripts/make_neet_2026_charts.py)
+charts/

scripts/make_neet_2026_charts.py

@@ -0,0 +1,271 @@
+"""Generate publication-quality charts for the NEET 2026 AI benchmark.
+
+Reads live from `results/*NEET_2026*/summary.{md,jsonl}` and `predictions.jsonl`
+so the charts always reflect the current run set. Outputs PNGs into
+`charts/neet_2026/`.
+"""
+
+import argparse
+import json
+import re
+from collections import defaultdict
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+import numpy as np
+
+# Display label and brand color per OpenRouter model id.
+# New models get a default fallback color; add an entry for nicer styling.
+MODEL_META = {
+    "google/gemini-3-flash-preview": ("Gemini 3 Flash", "#4285F4"),
+    "google/gemini-3.1-pro-preview": ("Gemini 3.1 Pro", "#1A73E8"),
+    "openai/gpt-5.5": ("GPT-5.5", "#10A37F"),
+    "openai/gpt-5.4": ("GPT-5.4", "#7CC4A8"),
+    "qwen/qwen3-vl-235b-a22b-thinking": ("Qwen3-VL 235B Thinking", "#615CED"),
+    "anthropic/claude-opus-4.7": ("Claude Opus 4.7", "#C9622D"),
+    "anthropic/claude-sonnet-4.6": ("Claude Sonnet 4.6", "#D97757"),
+    "anthropic/claude-haiku-4.5": ("Claude Haiku 4.5", "#E8B496"),
+    "z-ai/glm-4.6v": ("GLM-4.6V", "#FFB000"),
+    "x-ai/grok-4.20": ("Grok 4.20", "#1DA1F2"),
+}
+DEFAULT_COLOR = "#888888"
+
+NEET_SUBJECTS = ["Physics", "Chemistry", "Biology"]
+NEET_SUBJECT_MAX = {"Physics": 180, "Chemistry": 180, "Biology": 360}
+
+plt.rcParams.update({
+    "font.family": "DejaVu Sans",
+    "font.size": 11,
+    "axes.spines.top": False,
+    "axes.spines.right": False,
+    "axes.titleweight": "bold",
+})
+
+
+def parse_run(run_dir: Path) -> dict | None:
+    summary_md = run_dir / "summary.md"
+    summary_jsonl = run_dir / "summary.jsonl"
+    predictions_jsonl = run_dir / "predictions.jsonl"
+    if not (summary_md.exists() and summary_jsonl.exists() and predictions_jsonl.exists()):
+        return None
+
+    md = summary_md.read_text()
+    score_m = re.search(r"\*\*Overall Score:\*\* \*\*(\d+)\*\* / \*\*(\d+)\*\*", md)
+    cost_m = re.search(r"\*\*Total Cost:\*\* \$(\S+)", md)
+    if not score_m:
+        return None
+
+    subject_by_qid = {}
+    with predictions_jsonl.open() as f:
+        for line in f:
+            r = json.loads(line)
+            subject_by_qid[r["question_id"]] = r.get("subject", "Unknown")
+
+    correct = incorrect = failed = 0
+    subj_score = defaultdict(int)
+    subj_correct = defaultdict(int)
+    subj_total = defaultdict(int)
+    with summary_jsonl.open() as f:
+        for line in f:
+            r = json.loads(line)
+            status = r["evaluation_status"]
+            if status == "correct":
+                correct += 1
+            elif status == "incorrect":
+                incorrect += 1
+            else:
+                failed += 1
+            sub = subject_by_qid.get(r["question_id"], "Unknown")
+            subj_total[sub] += 1
+            subj_score[sub] += r.get("marks_awarded", 0)
+            if status == "correct":
+                subj_correct[sub] += 1
+
+    # Recover model id from dir name: provider_model_..._NEET_2026_TIMESTAMP
+    name = run_dir.name.split("_NEET_2026")[0]
+    # First underscore separates provider from rest of model slug.
+    model_id = name.replace("_", "/", 1)
+
+    return {
+        "model": model_id,
+        "score": int(score_m.group(1)),
+        "total": int(score_m.group(2)),
+        "correct": correct,
+        "incorrect": incorrect,
+        "failed": failed,
+        "cost": float(cost_m.group(1)) if cost_m else 0.0,
+        "subjects": {s: subj_score.get(s, 0) for s in NEET_SUBJECTS},
+        "run_dir": str(run_dir),
+    }
+
+
+def load_results(results_dir: Path) -> list[dict]:
+    runs = []
+    for d in sorted(results_dir.glob("*NEET_2026*")):
+        if not d.is_dir():
+            continue
+        parsed = parse_run(d)
+        if parsed:
+            runs.append(parsed)
+    # Keep most recent run per model (in case of reruns)
+    by_model = {}
+    for r in runs:
+        prev = by_model.get(r["model"])
+        if prev is None or r["run_dir"] > prev["run_dir"]:
+            by_model[r["model"]] = r
+    return sorted(by_model.values(), key=lambda x: -x["score"])
+
+
+def label_for(model_id: str) -> str:
+    return MODEL_META.get(model_id, (model_id, DEFAULT_COLOR))[0]
+
+
+def color_for(model_id: str) -> str:
+    return MODEL_META.get(model_id, (model_id, DEFAULT_COLOR))[1]
+
+
+def chart_main_leaderboard(data: list[dict], out_path: Path):
+    data = sorted(data, key=lambda x: x["score"])
+    labels = [label_for(d["model"]) for d in data]
+    scores = [d["score"] for d in data]
+    colors = [color_for(d["model"]) for d in data]
+    pcts = [s / 720 * 100 for s in scores]
+
+    fig, ax = plt.subplots(figsize=(11, max(5, 0.85 * len(data) + 1.5)), dpi=200)
+    fig.patch.set_facecolor("white")
+    bars = ax.barh(labels, scores, color=colors, edgecolor="white", linewidth=1.5, height=0.72)
+
+    ax.axvline(715, color="#E63946", linestyle="--", linewidth=1.4, alpha=0.8, zorder=0)
+    ax.text(715, len(labels) - 0.35, "  AIR-1 cutoff zone", color="#E63946",
+            fontsize=10, fontweight="bold", va="bottom")
+
+    for bar, score, pct in zip(bars, scores, pcts):
+        ax.text(bar.get_width() + 8, bar.get_y() + bar.get_height() / 2,
+                f"{score}  ({pct:.1f}%)", va="center", fontsize=11, fontweight="bold",
+                color="#222")
+
+    ax.set_xlim(0, 800)
+    ax.set_xlabel("Score (out of 720)", fontsize=12, fontweight="bold")
+    ax.set_title("NEET 2026 — Frontier AI Models Take India's Hardest Medical Exam",
+                 fontsize=15, pad=14)
+    ax.text(0, len(labels) + 0.35,
+            "180 vision questions • zero shot • image input • exam date: 3 May 2026",
+            fontsize=10.5, color="#666", style="italic")
+    ax.tick_params(axis="y", labelsize=11.5)
+    ax.set_axisbelow(True)
+    ax.grid(axis="x", alpha=0.22, linestyle="-", linewidth=0.6)
+
+    fig.text(0.99, 0.01, "github.com/Reja1/jee-neet-benchmark", ha="right",
+             fontsize=8.5, color="#999", style="italic")
+    plt.tight_layout()
+    plt.savefig(out_path, bbox_inches="tight", facecolor="white")
+    print(f"wrote {out_path}")
+    plt.close()
+
+
+def chart_cost_vs_accuracy(data: list[dict], out_path: Path):
+    fig, ax = plt.subplots(figsize=(11, 7), dpi=200)
+    fig.patch.set_facecolor("white")
+
+    for d in data:
+        if d["cost"] <= 0:
+            continue
+        pct = d["score"] / 720 * 100
+        ax.scatter(d["cost"], pct, s=320, color=color_for(d["model"]),
+                   edgecolors="white", linewidths=2, zorder=3, alpha=0.95)
+        ax.annotate(label_for(d["model"]), (d["cost"], pct),
+                    xytext=(d["cost"] * 1.05, pct + 0.6),
+                    fontsize=10.5, fontweight="bold", color="#222")
+
+    pareto_pts = sorted([(d["cost"], d["score"] / 720 * 100) for d in data if d["cost"] > 0])
+    frontier, best = [], -1
+    for c, p in pareto_pts:
+        if p > best:
+            frontier.append((c, p))
+            best = p
+    if frontier:
+        fx, fy = zip(*frontier)
+        ax.plot(fx, fy, "--", color="#888", linewidth=1.4, alpha=0.5, zorder=1, label="Pareto frontier")
+
+    ax.set_xscale("log")
+    ax.set_xlabel("Total cost for full exam (USD, log scale)", fontsize=12, fontweight="bold")
+    ax.set_ylabel("Accuracy (%)", fontsize=12, fontweight="bold")
+    ax.set_title("NEET 2026 — Cost vs Accuracy (upper-left = best value)",
+                 fontsize=14, pad=14)
+    pcts = [d["score"] / 720 * 100 for d in data]
+    ax.set_ylim(min(pcts) - 5, 102)
+    ax.grid(alpha=0.25, linestyle="-", linewidth=0.5)
+    ax.set_axisbelow(True)
+    ax.legend(loc="lower right", frameon=False, fontsize=10)
+
+    fig.text(0.99, 0.01, "github.com/Reja1/jee-neet-benchmark",
+             ha="right", fontsize=8.5, color="#999", style="italic")
+    plt.tight_layout()
+    plt.savefig(out_path, bbox_inches="tight", facecolor="white")
+    print(f"wrote {out_path}")
+    plt.close()
+
+
+def chart_subject_heatmap(data: list[dict], out_path: Path):
+    data = sorted(data, key=lambda x: -x["score"])
+    matrix = np.array([
+        [d["subjects"].get(s, 0) / NEET_SUBJECT_MAX[s] * 100 for s in NEET_SUBJECTS]
+        for d in data
+    ])
+    labels = [label_for(d["model"]) for d in data]
+
+    fig, ax = plt.subplots(figsize=(8.5, max(4, 0.8 * len(data) + 1.5)), dpi=200)
+    fig.patch.set_facecolor("white")
+
+    im = ax.imshow(matrix, cmap="RdYlGn", vmin=40, vmax=100, aspect="auto")
+    ax.set_xticks(range(len(NEET_SUBJECTS)))
+    ax.set_xticklabels(NEET_SUBJECTS, fontsize=12, fontweight="bold")
+    ax.set_yticks(range(len(labels)))
+    ax.set_yticklabels(labels, fontsize=11)
+    ax.set_xticks(np.arange(-0.5, len(NEET_SUBJECTS), 1), minor=True)
+    ax.set_yticks(np.arange(-0.5, len(labels), 1), minor=True)
+    ax.grid(which="minor", color="white", linewidth=2)
+    ax.tick_params(which="minor", length=0)
+
+    for i in range(matrix.shape[0]):
+        for j in range(matrix.shape[1]):
+            val = matrix[i, j]
+            text_color = "white" if val < 65 else "#222"
+            ax.text(j, i, f"{val:.0f}%", ha="center", va="center",
+                    fontsize=11, fontweight="bold", color=text_color)
+
+    ax.set_title("Per-Subject Accuracy — NEET 2026", fontsize=14, pad=12)
+    cbar = fig.colorbar(im, ax=ax, shrink=0.85)
+    cbar.set_label("% correct", fontsize=10)
+
+    fig.text(0.99, 0.01, "github.com/Reja1/jee-neet-benchmark",
+             ha="right", fontsize=8.5, color="#999", style="italic")
+    plt.tight_layout()
+    plt.savefig(out_path, bbox_inches="tight", facecolor="white")
+    print(f"wrote {out_path}")
+    plt.close()
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--results-dir", default="results")
+    ap.add_argument("--output-dir", default="charts/neet_2026")
+    args = ap.parse_args()
+
+    out_dir = Path(args.output_dir)
+    out_dir.mkdir(parents=True, exist_ok=True)
+    data = load_results(Path(args.results_dir))
+    if not data:
+        raise SystemExit(f"No NEET 2026 runs found under {args.results_dir}")
+    print(f"loaded {len(data)} runs from {args.results_dir}")
+    for d in data:
+        print(f"  {label_for(d['model']):30s} {d['score']:3d}/720  ${d['cost']:.4f}")
+
+    chart_main_leaderboard(data, out_dir / "01_leaderboard.png")
+    chart_cost_vs_accuracy(data, out_dir / "02_cost_vs_accuracy.png")
+    chart_subject_heatmap(data, out_dir / "03_subject_heatmap.png")
+    print(f"\ncharts written to {out_dir.resolve()}")
+
+
+if __name__ == "__main__":
+    main()