analysis/code/charts.py

"""Grafici della corsa di conferma: PNG + un HTML autonomo (SVG in linea, nessuna dipendenza esterna).
Uso: ~/.hermes/cache/scratch/mpl-venv/bin/python analysis/charts.py   (matplotlib solo per disegnare)"""
import io
import json
from pathlib import Path

import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt  # noqa: E402

import stats as S  # noqa: E402

OUT = Path(__file__).resolve().parents[1] / "charts"
OUT.mkdir(exist_ok=True)
A = Path(__file__).resolve().parent
summ = json.load(open(A / "summary.json"))
xsum = json.load(open(A / "extra.json"))
POLS = ["A", "D", "B", "T"]
NAMES = {"A": "full window", "D": "mnemonic-proxy", "B": "summary (LLM compaction)", "T": "truncation"}
COL = {"A": "#7f7f7f", "D": "#1f77b4", "B": "#ff7f0e", "T": "#2ca02c"}
SVGS = {}
LAB = {"unite": "64k + 128k pooled"}


def save(fig, name):
    fig.savefig(OUT / (name + ".png"), dpi=130, bbox_inches="tight")
    buf = io.StringIO()
    fig.savefig(buf, format="svg", bbox_inches="tight")
    s = buf.getvalue()
    SVGS[name] = s[s.index("<svg"):]
    plt.close(fig)


# 1. corrette per politica e lunghezza, IC 95% Wilson
def bars(data, title, name, groups, ylabel="fully correct (official score = 1), %"):
    fig, ax = plt.subplots(figsize=(8, 4.2))
    w = 0.2
    for k, p in enumerate(POLS):
        xs, ys, lo, hi = [], [], [], []
        for g, key in enumerate(groups):
            x, n = data[key][p]
            l, u = S.wilson(x, n)
            xs.append(g + (k - 1.5) * w)
            ys.append(100 * x / n)
            lo.append(100 * (x / n - l))
            hi.append(100 * (u - x / n))
        ax.bar(xs, ys, w, color=COL[p], label=NAMES[p], yerr=[lo, hi], capsize=3, error_kw={"lw": 1})
        for xx, yy in zip(xs, ys):
            ax.text(xx, 2, f"{yy:.0f}", ha="center", va="bottom", fontsize=7, color="white")
    ax.set_xticks(range(len(groups)))
    ax.set_xticklabels([f"{LAB.get(g, g)} (n={data[g]['A'][1]})" for g in groups])
    ax.set_ylim(0, 100)
    ax.set_ylabel(ylabel)
    ax.set_title(title, fontsize=10)
    ax.legend(fontsize=8, ncol=4, loc="upper center", bbox_to_anchor=(0.5, -0.1))
    ax.grid(axis="y", alpha=0.3)
    save(fig, name)


bars(summ["correct"]["off"], "Confirmation run. AgentLongBench kf-c tool_response, Qwen3.8 Flash Next IQ3_XXS,\n"
     "window 32K (full window: 80K / 160K). Error bars: 95% Wilson", "correct_by_policy", ["64k", "128k", "unite"])

# 1b. corsa principale (domanda ambigua) contro conferma (domanda con la nota), 64k + 128k
mvc = xsum["main_vs_confirm"]["unite"]
data = {"main run (ambiguous question)": {p: [mvc[p]["main"], mvc[p]["n"]] for p in POLS},
        "confirmation run (question + note)": {p: [mvc[p]["confirm"], mvc[p]["n"]] for p in POLS}}
bars(data, "Same 100 questions, same engine and settings. Only change: one sentence added to the question\n"
     "(64k + 128k pooled, error bars: 95% Wilson)", "main_vs_confirm",
     ["main run (ambiguous question)", "confirmation run (question + note)"])

# 2. tempo mediano per domanda
fig, ax = plt.subplots(figsize=(8, 4.2))
w = 0.2
for k, p in enumerate(POLS):
    for g, L in enumerate(["64k", "128k"]):
        c = summ["cost"][f"{L}/{p}"]
        x = g + (k - 1.5) * w
        ax.bar(x, c["prep_med"], w, color=COL[p], alpha=0.45, hatch="//", edgecolor="white")
        ax.bar(x, c["q_med"], w, bottom=c["prep_med"], color=COL[p], label=NAMES[p] if g == 0 else None)
        ax.text(x, c["prep_med"] + c["q_med"] + 3, f"{c['wall_med']:.0f}s", ha="center", fontsize=7)
from matplotlib.patches import Patch  # noqa: E402
ax.set_xticks([0, 1])
ax.set_xticklabels(["64k (~74K-token history)", "128k (~150K-token history)"])
ax.set_ylim(0, 330)
h, l = ax.get_legend_handles_labels()
h += [Patch(facecolor="#bbbbbb", hatch="//", edgecolor="white"), Patch(facecolor="#555555")]
l += ["preparation", "at question time"]
ax.set_ylabel("seconds per question (median)")
ax.set_title("Median time per question. Hatched = preparation (proxy: history replay through the proxy;\n"
             "summary: compactions). Solid = work at question time. Label = median end-to-end wall time.", fontsize=9)
ax.legend(h, l, fontsize=8, ncol=3, loc="upper center", bbox_to_anchor=(0.5, -0.1))
ax.grid(axis="y", alpha=0.3)
save(fig, "time_per_question")

# 3. matrice delle discordanze D/B
fig, axs = plt.subplots(1, 3, figsize=(10, 3.4), constrained_layout=True)
for ax, key in zip(axs, ["64k", "128k", "unite"]):
    c = summ["compare"]["main"][f"D-B/{key}"]
    m = [[c["a"], c["b"]], [c["c"], c["d"]]]
    for i in range(2):
        for j in range(2):
            ax.text(j, i, m[i][j], ha="center", va="center", fontsize=14,
                    color="white" if m[i][j] >= 20 else "black")
    ax.imshow(m, cmap="Blues", vmin=0, vmax=45)
    ax.set_xticks([0, 1])
    ax.set_xticklabels(["summary right", "summary wrong"])
    ax.set_yticks([0, 1])
    ax.set_yticklabels(["proxy right", "proxy wrong"])
    ax.set_title(f"{LAB.get(key, key)}: n = {c['n']}, McNemar exact p = {c['p_mcnemar']:.2g}", fontsize=9)
fig.suptitle("Primary comparison (confirmation run): mnemonic-proxy vs summary, paired", fontsize=10)
save(fig, "discordance_D_B")

# HTML autonomo
html = ["<!doctype html><html><head><meta charset='utf-8'><title>AgentLongBench confirmation run — mnemonic-proxy</title>",
        "<style>body{font-family:sans-serif;max-width:980px;margin:2em auto;color:#222}svg{max-width:100%;height:auto}"
        "table{border-collapse:collapse;font-size:13px}td,th{border:1px solid #ccc;padding:3px 7px;text-align:right}"
        "</style></head><body>",
        "<h1>AgentLongBench confirmation run (window 32K, 64k + 128k)</h1>",
        "<p>Qwen3.8 Flash Next IQ3_XXS on Strata 0.1.40, RTX 4070 Ti 12 GB, KV int8, seed 1. kf-c tool_response, "
        "Count Frequency + Find Duplicates, 25 + 25 questions per length, max_tokens 4,096. Full window = 81,920 / 163,840; "
        "mnemonic-proxy 0.3.5 (b8d4920), summary (LLM compaction, at most 2,048 tokens) and truncation at 32,768. "
        "Same questions as the main run plus one note that says which tool output \u201cround N\u201d refers to; "
        "official answers and scorer unchanged.</p>"]
html.append("<table><tr><th>set</th>" + "".join(f"<th>{NAMES[p]}</th>" for p in POLS) + "</tr>")
for key in ["64k", "128k", "unite"]:
    html.append(f"<tr><td>{LAB.get(key, key)}</td>" + "".join(
        f"<td>{summ['correct']['off'][key][p][0]}/{summ['correct']['off'][key][p][1]}</td>" for p in POLS) + "</tr>")
html.append("</table>")
for name, cap in (("correct_by_policy", "Fully correct answers per method and length."),
                  ("main_vs_confirm", "Main run vs confirmation run."),
                  ("discordance_D_B", "Primary comparison: mnemonic-proxy vs summary, paired."),
                  ("time_per_question", "Median time per question.")):
    html.append(f"<h2>{cap}</h2>")
    html.append(SVGS[name])
html.append("</body></html>")
(OUT / "confirm-charts.html").write_text("\n".join(html))
print("ok", sorted(p.name for p in OUT.iterdir()))