Blame for auto-impl/analysis/make_tables.py

#!/usr/bin/env python3
"""Render the summary into the markdown tables the report embeds.

Kept separate from `run_analysis.py` so the numbers and their presentation
can be regenerated independently: re-run this after any change to the
feature extractors and the report's tables update without hand-editing.

    python analysis/make_tables.py    # per-agent tables into analysis/out/ and docs/results/
"""

from __future__ import annotations

import os
import sys

sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))

from run_analysis import AGENT_ROOTS, OUT_DIR, ROOT, build, summarize  # noqa: E402
from trace_analysis.probes import probe_cohort  # noqa: E402
from trace_analysis.stats import within_task_z  # noqa: E402


DOC_NAMES = {
    "claude": ("claude_code_agent_trace_tables_2026-10.md", "Claude Code (claude-opus-5)", "claude_code_agent_trace_analysis_2026-10.md"),
    "codex": ("codex_sdk_agent_trace_tables_2026-10.md", "Codex SDK (gpt-5.6-sol)", "codex_sdk_agent_trace_analysis_2026-10.md"),
}

DOC_HEADER = """# 附录:{title} Trace 逐项数据表(2026-10)

本文件由 `python analysis/make_tables.py` 生成,配合
[`{doc}`]({doc}) 阅读。
修改特征抽取后重跑该脚本即可刷新,不要手工编辑表格内容。

## 高层要点 / Takeaways

1. **这是附录,不是结论。** 这里只有逐 run、逐特征的**全量数值**,
   解读一律在 [`{doc}`]({doc}) 里,两边不重复。
2. **要查一个具体的 run 发生了什么**,从这里的表定位到 run id,
   再去 `analysis/out/*/narratives/` 读那一份的可读流水账。
3. **跨 benchmark 不要直接比绝对分** —— 用 `z in task` 那一列(同 benchmark 内的 z-score),
   本项目所有相关性分析都基于它。
4. **`cohort` 列决定一行算不算数**:只有 `primary` 和 `recovered` 进入分析,
   其余是被我们自己的基础设施掐断的,不是 agent 放弃。

> This file is a generated appendix: per-run, per-feature raw numbers only. All interpretation
> lives in the analysis document it accompanies. Compare across benchmarks using the `z in task`
> column, never the raw accuracy; and note that only `primary` and `recovered` rows enter the
> analysis — the rest were cut short by our own infrastructure, not by the agent.

列的含义:
- **cohort** — `primary` 为 agent 自己跑完并被打分;`recovered` 为研究循环完整、仅收尾消息被我们的网关掐断。
- **z in task** — 最终准确率在**同一 benchmark 内部**的 z-score。跨 benchmark 的绝对分不可比,所有相关性分析都基于这一列。
- **versions** — 按 `evaluate.py --model-path` 去重的独立模型版本数,即真实迭代轮数。
- **idle h** — 没有任何前台或后台任务在跑的墙钟时长。

---

"""


def table(rows: list[dict], cols: list[tuple[str, str]]) -> str:
    out = ["| " + " | ".join(h for _, h in cols) + " |", "|" + "---|" * len(cols)]
    for r in rows:
        cells = []
        for k, _ in cols:
            v = r.get(k)
            cells.append("" if v is None else (f"{v:.3f}" if isinstance(v, float) else str(v)))
        out.append("| " + " | ".join(cells) + " |")
    return "\n".join(out)


def build_tables(agent: str, results_root: str) -> None:
    runs, feats = build(results_root)
    s = summarize(feats)
    inc = [f for f in feats if f.cohort in ("primary", "recovered")]
    keep = {f.run_id for f in inc}
    s["prose_probes"] = probe_cohort([r for r in runs if r.run_id in keep])
    z = within_task_z(inc)

    parts: list[str] = ["<!-- generated by analysis/make_tables.py; do not hand-edit -->", ""]

    parts += ["## cohort", "", table(
        [{"k": k, "v": v} for k, v in s["cohort_counts"].items()]
        + [{"k": f"excluded: {k}", "v": v} for k, v in s["exclusions"].items()],
        [("k", "bucket"), ("v", "traces")],
    ), ""]

    parts += ["## per benchmark", "", table(
        s["per_task"],
        [
            ("task", "benchmark"),
            ("n_runs", "runs"),
            ("mean_accuracy", "accuracy"),
            ("mean_turns", "turns"),
            ("mean_eval_calls", "official evals"),
            ("mean_model_versions", "model versions"),
            ("mean_dead_air_h", "idle h"),
            ("mean_tail_idle_h", "h after best"),
            ("mean_oom_hits", "OOM hits"),
            ("mean_cost_usd", "cost $"),
        ],
    ), ""]

    parts += ["## per run", "", table(
        sorted(
            [
                {
                    "run": f.run_id.rsplit("_", 1)[0],
                    "when": f.run_id.rsplit("_", 1)[1],
                    "cohort": f.cohort,
                    "acc": f.accuracy,
                    "z": round(z.get(f.run_id, 0.0), 2),
                    "turns": f.turns,
                    "vers": f.n_distinct_model_versions,
                    "evals": f.n_eval_calls,
                    "oom": f.error_signals.get("oom", 0),
                    "errs": f.tool_errors,
                    "idle_h": round(f.dead_air_h or 0, 2),
                    "cost": round(f.cost_usd, 2),
                }
                for f in inc
            ],
            key=lambda r: (r["run"], r["when"]),
        ),
        [
            ("run", "task / base model"),
            ("when", "stamp"),
            ("cohort", "cohort"),
            ("acc", "accuracy"),
            ("z", "z in task"),
            ("turns", "turns"),
            ("vers", "versions"),
            ("evals", "evals"),
            ("oom", "OOM"),
            ("errs", "tool errors"),
            ("idle_h", "idle h"),
            ("cost", "cost $"),
        ],
    ), ""]

    parts += ["## techniques (runs with evidence, of %d)" % len(inc), "", table(
        [{"t": k, "n": v, "pct": round(100 * v / len(inc), 1)} for k, v in s["technique_run_counts"].items()],
        [("t", "technique"), ("n", "runs"), ("pct", "%")],
    ), ""]

    parts += ["## technique vs relative outcome", "", table(
        s["technique_effects"],
        [("technique", "technique"), ("n_with", "with"), ("n_without", "without"),
         ("mean_z_with", "mean z with"), ("mean_z_without", "mean z without"), ("delta_z", "Δz")],
    ), ""]

    parts += ["## feature vs relative outcome (Spearman)", "", table(
        s["feature_correlations"],
        [("feature", "feature"), ("rho_vs_within_task_z", "rho"), ("n", "n")],
    ), ""]

    parts += ["## rerun spread (same cell, two runs)", "", table(
        s["rerun_variance"],
        [("task", "benchmark"), ("base_model", "base model"), ("min", "worse"), ("max", "better"), ("spread", "spread")],
    ), ""]

    parts += ["## prose habits", "", table(
        [{"probe": k, **v} for k, v in s["prose_probes"].items() if not k.startswith("anywhere_")],
        [("probe", "habit"), ("runs_with_evidence", "runs"), ("of", "of"), ("pct", "%"), ("surface", "where")],
    ), ""]

    parts += ["## datasets used (top 25)", "", table(
        [{"d": k, "n": v} for k, v in list(s["dataset_run_counts"].items())[:25]],
        [("d", "dataset"), ("n", "runs")],
    ), ""]

    path = os.path.join(OUT_DIR, agent, "tables.md")
    os.makedirs(os.path.dirname(path), exist_ok=True)
    with open(path, "w") as fh:
        fh.write("\n".join(parts))
    print(f"wrote {path}")

    fname, title, analysis_doc = DOC_NAMES[agent]
    doc = os.path.join(ROOT, "docs", "results", fname)
    with open(doc, "w") as fh:
        fh.write(DOC_HEADER.format(title=title, doc=analysis_doc) + "\n".join(parts[1:]))
    print(f"wrote {doc}")


def main() -> None:
    for agent, root in AGENT_ROOTS.items():
        build_tables(agent, root)


if __name__ == "__main__":
    main()