diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 9e52a25..7d0fce1 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -64,7 +64,15 @@ jobs:
         with:
           key: libkrun   # distinct from the no-default-features `test` job's artifacts
       - name: Install libkrun
-        run: brew tap libkrun/krun && brew install libkrun libkrunfw
+        # Homebrew now gates third-party taps behind HOMEBREW_REQUIRE_TAP_TRUST
+        # (set on the runner image) — `brew install` refuses an untapped-trust
+        # formula ("Refusing to load formula … from untrusted tap"). `brew trust`
+        # the tap between tap+install; non-interactive + forward-compatible for
+        # when tap-trust becomes the Homebrew default.
+        run: |
+          brew tap libkrun/krun
+          brew trust libkrun/krun
+          brew install libkrun libkrunfw
       - name: Clippy (libkrun)
         run: cargo clippy --all-targets --features libkrun -- -D warnings
       - name: Test (libkrun)
diff --git a/.gitignore b/.gitignore
index 91a2ba0..ee22d87 100644
--- a/.gitignore
+++ b/.gitignore
@@ -23,3 +23,12 @@ __pycache__/
 *.pyc
 .venv/
 scripts/eval/toolz-tasks/
+
+# Run logs + ghost scratch outputs — noisy, regenerable, not evidence (all just
+# {task,cond,trial,score,cost} scalars + progress lines; no transcripts/memory).
+# The curated σ̂ result jsonls under scripts/eval/segmentation/results/ ARE tracked
+# (the evidence archive — see its README); only the .log transcripts there are ignored.
+scripts/eval/segmentation/results/*.log
+scripts/ghost/*.jsonl
+scripts/ghost/*.json
+scripts/ghost/*.log
diff --git a/scripts/bench/README.md b/scripts/bench/README.md
new file mode 100644
index 0000000..c8d1930
--- /dev/null
+++ b/scripts/bench/README.md
@@ -0,0 +1,45 @@
+# Startup benchmark
+
+`startup.py` runs pillbox startup cases, reads host-emitted
+`session.started` lifecycle events, and summarizes `startup_ms` plus per-stage
+timings.
+
+## Examples
+
+```sh
+# One Docker PTY case, five measured runs after one warmup.
+scripts/bench/startup.py --case docker-claude --warmup 1 --repeat 5
+
+# Server-mode opencode on Docker and libkrun, with a fixed model.
+# Defaults to the same opencode image used by the smoke/eval scripts:
+# PILLBOX_RUNNER_IMAGE=pillbox-runner:dev.
+scripts/bench/startup.py \
+  --case docker-opencode \
+  --case libkrun-opencode \
+  --model zai-coding-plan/glm-4.5-air \
+  --warmup 1 \
+  --repeat 5
+
+# Show the commands without starting agents.
+scripts/bench/startup.py --all-docker --dry-run
+
+# Emit raw measured runs plus aggregate summaries.
+scripts/bench/startup.py --all-docker --json
+```
+
+## Notes
+
+- Cases require the corresponding agent auth and backend support.
+- opencode cases default to `pillbox-runner:dev`; `libkrun-codex-serve` defaults
+  to `pillbox-runner:dev`, matching `scripts/smoke/`. Override with
+  `--runner-image IMAGE` or `PILLBOX_RUNNER_IMAGE`.
+- Every run uses `--detach --json --ttl` and is removed with
+  `pillbox session rm` after its event is captured.
+- By default, the benchmark creates one temporary empty workspace and reuses it
+  for every run. Use `--workspace PATH` to benchmark a real tree.
+- Commands time out after 300 seconds by default. Override with
+  `--timeout SECONDS` for slow cold-start environments.
+- Use `--json` for raw run records and aggregate summaries, or `--csv` for one
+  row per measured startup stage.
+- `PILLBOX` can point at a specific binary; otherwise the script uses
+  `target/debug/pillbox` when present, then falls back to `pillbox` on `PATH`.
diff --git a/scripts/bench/startup.py b/scripts/bench/startup.py
new file mode 100755
index 0000000..030c88a
--- /dev/null
+++ b/scripts/bench/startup.py
@@ -0,0 +1,594 @@
+#!/usr/bin/env python3
+"""Benchmark pillbox sandbox startup timings.
+
+The script consumes the lifecycle fields emitted by host-side
+`session.started` events:
+
+  - startup_ms
+  - startup_stages: [{name, duration_ms}, ...]
+
+It runs configured cases, captures each run's session id from `pillbox run
+--json`, reads `pillbox session events --json`, and summarizes totals plus
+per-stage p50/p95. The harness deliberately stays outside pillbox proper so the
+measurement surface can be used before and after each optimization PR.
+"""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import json
+import math
+import os
+import shutil
+import subprocess
+import sys
+import tempfile
+import time
+import uuid
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, Optional, Union
+
+
+ROOT = Path(__file__).resolve().parents[2]
+DEFAULT_PILLBOX = ROOT / "target" / "debug" / "pillbox"
+
+
+@dataclass(frozen=True)
+class Case:
+    name: str
+    backend: str
+    agent: str
+    description: str
+    needs_model: bool = False
+    runner_image: Optional[str] = None
+
+
+CASES: dict[str, Case] = {
+    "docker-claude": Case(
+        "docker-claude",
+        "docker",
+        "claude",
+        "Docker backend, Claude PTY session",
+    ),
+    "docker-codex": Case(
+        "docker-codex",
+        "docker",
+        "codex",
+        "Docker backend, Codex PTY session",
+    ),
+    "docker-opencode": Case(
+        "docker-opencode",
+        "docker",
+        "opencode",
+        "Docker backend, opencode server session",
+        needs_model=True,
+        runner_image="pillbox-runner:dev",
+    ),
+    "libkrun-claude": Case(
+        "libkrun-claude",
+        "libkrun",
+        "claude",
+        "libkrun backend, Claude PTY session",
+    ),
+    "libkrun-codex": Case(
+        "libkrun-codex",
+        "libkrun",
+        "codex",
+        "libkrun backend, Codex PTY session",
+    ),
+    "libkrun-opencode": Case(
+        "libkrun-opencode",
+        "libkrun",
+        "opencode",
+        "libkrun backend, opencode server session",
+        needs_model=True,
+        runner_image="pillbox-runner:dev",
+    ),
+    "libkrun-codex-serve": Case(
+        "libkrun-codex-serve",
+        "libkrun",
+        "codex-serve",
+        "libkrun backend, codex app-server session",
+        needs_model=True,
+        runner_image="pillbox-runner:dev",
+    ),
+}
+
+
+def main() -> int:
+    args = parse_args()
+    if args.list_cases:
+        list_cases()
+        return 0
+
+    case_names = selected_cases(args)
+    if not case_names:
+        print(
+            "startup-bench: select at least one --case, --all-docker, --all-libkrun, or --all",
+            file=sys.stderr,
+        )
+        return 2
+
+    pillbox = resolve_pillbox(args.pillbox)
+    workspace_guard = None
+    workspace = Path(args.workspace).resolve() if args.workspace else None
+    if workspace is None:
+        workspace_guard = tempfile.TemporaryDirectory(prefix="pillbox-startup-bench-")
+        workspace = Path(workspace_guard.name)
+        (workspace / "README.md").write_text("startup benchmark workspace\n", encoding="utf-8")
+
+    base_cmd = [str(pillbox)]
+    if args.pillbox_name:
+        base_cmd.extend(["--pillbox", args.pillbox_name])
+
+    runs: list[dict[str, Any]] = []
+    failures: list[dict[str, str]] = []
+    try:
+        for name in case_names:
+            case = CASES[name]
+            total = args.warmup + args.repeat
+            for index in range(total):
+                measured = index >= args.warmup
+                label = f"startup-bench-{case.name}-{uuid.uuid4().hex[:8]}"
+                run = run_case(
+                    args=args,
+                    case=case,
+                    base_cmd=base_cmd,
+                    workspace=workspace,
+                    label=label,
+                    measured=measured,
+                )
+                if run.get("ok"):
+                    if measured:
+                        runs.append(run)
+                else:
+                    failures.append(
+                        {
+                            "case": case.name,
+                            "label": label,
+                            "reason": str(run.get("reason", "unknown failure")),
+                        }
+                    )
+    finally:
+        if workspace_guard is not None:
+            workspace_guard.cleanup()
+
+    if args.dry_run:
+        return 0
+    if args.output_json:
+        print(
+            json.dumps(
+                {"runs": runs, "summary": summarize(runs), "failures": failures},
+                indent=2,
+                sort_keys=True,
+            )
+        )
+    elif args.output_csv:
+        print_csv(runs)
+    else:
+        print_human(runs, failures)
+    return 1 if failures else 0
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Run pillbox startup cases and summarize host session.started startup timings.",
+    )
+    parser.add_argument(
+        "--case",
+        action="append",
+        choices=sorted(CASES),
+        default=[],
+        help="Case to run. Repeatable.",
+    )
+    parser.add_argument("--all-docker", action="store_true", help="Run every Docker case.")
+    parser.add_argument("--all-libkrun", action="store_true", help="Run every libkrun case.")
+    parser.add_argument("--all", action="store_true", help="Run every known case.")
+    parser.add_argument("--list-cases", action="store_true", help="List cases and exit.")
+    parser.add_argument(
+        "--repeat",
+        type=positive_int,
+        default=5,
+        help="Measured runs per case. Default: 5.",
+    )
+    parser.add_argument(
+        "--warmup",
+        type=nonnegative_int,
+        default=1,
+        help="Unmeasured warmup runs per case. Default: 1.",
+    )
+    parser.add_argument("--workspace", help="Workspace to mount. Default: temporary empty workspace.")
+    parser.add_argument(
+        "--workspace-name",
+        default="startup-bench",
+        help="Guest /workspace mount name. Default: startup-bench.",
+    )
+    parser.add_argument("--ttl", default="30m", help="Detached session TTL. Default: 30m.")
+    parser.add_argument("--model", default=os.environ.get("MODEL"), help="Model for server-mode agents.")
+    parser.add_argument(
+        "--runner-image",
+        default=os.environ.get("PILLBOX_RUNNER_IMAGE"),
+        help="Override PILLBOX_RUNNER_IMAGE for every case.",
+    )
+    parser.add_argument(
+        "--pillbox",
+        default=os.environ.get("PILLBOX"),
+        help="Pillbox binary. Default: target/debug/pillbox, then PATH.",
+    )
+    parser.add_argument("--pillbox-name", help="Pass --pillbox NAME to the CLI.")
+    parser.add_argument(
+        "--cwd",
+        default=os.getcwd(),
+        help="Directory where pillbox commands run. Default: current directory.",
+    )
+    parser.add_argument(
+        "--timeout",
+        type=positive_int,
+        default=300,
+        help="Per pillbox command timeout in seconds. Default: 300.",
+    )
+    parser.add_argument("--dry-run", action="store_true", help="Print commands without running them.")
+    parser.add_argument("--json", dest="output_json", action="store_true", help="Print JSON results.")
+    parser.add_argument("--csv", dest="output_csv", action="store_true", help="Print CSV rows for measured runs.")
+    return parser.parse_args()
+
+
+def positive_int(raw: str) -> int:
+    value = int(raw)
+    if value < 1:
+        raise argparse.ArgumentTypeError("must be >= 1")
+    return value
+
+
+def nonnegative_int(raw: str) -> int:
+    value = int(raw)
+    if value < 0:
+        raise argparse.ArgumentTypeError("must be >= 0")
+    return value
+
+
+def resolve_pillbox(raw: Optional[str]) -> Union[Path, str]:
+    if raw:
+        return Path(raw).resolve() if "/" in raw else raw
+    if DEFAULT_PILLBOX.exists():
+        return DEFAULT_PILLBOX
+    found = shutil.which("pillbox")
+    return found if found else "pillbox"
+
+
+def selected_cases(args: argparse.Namespace) -> list[str]:
+    names: list[str] = []
+    if args.all:
+        names.extend(CASES)
+    if args.all_docker:
+        names.extend(name for name, case in CASES.items() if case.backend == "docker")
+    if args.all_libkrun:
+        names.extend(name for name, case in CASES.items() if case.backend == "libkrun")
+    names.extend(args.case)
+    out: list[str] = []
+    seen: set[str] = set()
+    for name in names:
+        if name not in seen:
+            seen.add(name)
+            out.append(name)
+    return out
+
+
+def list_cases() -> None:
+    width = max(len(name) for name in CASES)
+    for name, case in CASES.items():
+        image = f" image={case.runner_image}" if case.runner_image else ""
+        print(f"{name:<{width}}  backend={case.backend:<7} agent={case.agent:<11}{image} {case.description}")
+
+
+def run_case(
+    *,
+    args: argparse.Namespace,
+    case: Case,
+    base_cmd: list[str],
+    workspace: Path,
+    label: str,
+    measured: bool,
+) -> dict[str, Any]:
+    cmd = build_run_cmd(args, case, base_cmd, workspace, label)
+    env = os.environ.copy()
+    if case.backend == "libkrun":
+        env["PILLBOX_BACKEND"] = "libkrun"
+    else:
+        env.pop("PILLBOX_BACKEND", None)
+    runner_image = args.runner_image or case.runner_image
+    if runner_image:
+        env["PILLBOX_RUNNER_IMAGE"] = runner_image
+
+    prefix = "measure" if measured else "warmup"
+    if args.dry_run:
+        print(f"[{prefix}] {case.name}: {env_prefix(case, runner_image)}{shell_join(cmd)}")
+        return {"ok": True, "case": case.name, "dry_run": True}
+
+    started = time.time()
+    try:
+        proc = subprocess.run(
+            cmd,
+            cwd=args.cwd,
+            env=env,
+            text=True,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.PIPE,
+            timeout=args.timeout,
+        )
+    except subprocess.TimeoutExpired as exc:
+        elapsed_ms = int((time.time() - started) * 1000)
+        output = tail_timeout_output(exc)
+        return {
+            "ok": False,
+            "case": case.name,
+            "reason": f"run timed out after {args.timeout}s ({elapsed_ms}ms elapsed): {output}",
+        }
+    elapsed_ms = int((time.time() - started) * 1000)
+    if proc.returncode != 0:
+        return {
+            "ok": False,
+            "case": case.name,
+            "reason": tail(proc.stderr or proc.stdout),
+        }
+
+    session_id = parse_session_id(proc.stdout)
+    if not session_id:
+        return {
+            "ok": False,
+            "case": case.name,
+            "reason": f"no session id in stdout: {tail(proc.stdout)}",
+        }
+
+    try:
+        event = find_started_event(base_cmd, args.cwd, env, session_id, args.timeout)
+        if event is None:
+            return {"ok": False, "case": case.name, "reason": f"no host session.started event for {session_id}"}
+        stages = event.get("startup_stages") or []
+        return {
+            "ok": True,
+            "case": case.name,
+            "backend": event.get("backend") or case.backend,
+            "agent_id": event.get("agent_id") or case.agent,
+            "session_id": session_id,
+            "startup_ms": event.get("startup_ms"),
+            "startup_stages": stages,
+            "run_elapsed_ms": elapsed_ms,
+            "measured": measured,
+        }
+    finally:
+        cleanup_session(base_cmd, args.cwd, env, session_id, args.timeout)
+
+
+def build_run_cmd(
+    args: argparse.Namespace,
+    case: Case,
+    base_cmd: list[str],
+    workspace: Path,
+    label: str,
+) -> list[str]:
+    cmd = [
+        *base_cmd,
+        "run",
+        "--agent",
+        case.agent,
+        "--workspace",
+        str(workspace),
+        "--name",
+        args.workspace_name,
+        "--detach",
+        "--label",
+        label,
+        "--ttl",
+        args.ttl,
+        "--json",
+    ]
+    if case.needs_model and args.model:
+        cmd.extend(["--model", args.model])
+    return cmd
+
+
+def env_prefix(case: Case, runner_image: Optional[str]) -> str:
+    parts = []
+    if case.backend == "libkrun":
+        parts.append("PILLBOX_BACKEND=libkrun")
+    if runner_image:
+        parts.append(f"PILLBOX_RUNNER_IMAGE={shell_quote(runner_image)}")
+    return "".join(f"{part} " for part in parts)
+
+
+def parse_session_id(stdout: str) -> Optional[str]:
+    try:
+        payload = json.loads(stdout)
+    except json.JSONDecodeError:
+        return None
+    session = payload.get("session")
+    if isinstance(session, dict):
+        sid = session.get("id")
+        return sid if isinstance(sid, str) and sid else None
+    return None
+
+
+def find_started_event(
+    base_cmd: list[str],
+    cwd: str,
+    env: dict[str, str],
+    session_id: str,
+    timeout: int,
+) -> Optional[dict[str, Any]]:
+    try:
+        proc = subprocess.run(
+            [*base_cmd, "session", "events", "--json"],
+            cwd=cwd,
+            env=env,
+            text=True,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.PIPE,
+            timeout=timeout,
+        )
+    except subprocess.TimeoutExpired:
+        return None
+    if proc.returncode != 0:
+        return None
+    found: Optional[dict[str, Any]] = None
+    for line in proc.stdout.splitlines():
+        try:
+            event = json.loads(line)
+        except json.JSONDecodeError:
+            continue
+        if (
+            event.get("session_id") == session_id
+            and event.get("event") == "session.started"
+            and event.get("emitter") == "host"
+        ):
+            found = event
+    return found
+
+
+def cleanup_session(
+    base_cmd: list[str],
+    cwd: str,
+    env: dict[str, str],
+    session_id: str,
+    timeout: int,
+) -> None:
+    try:
+        subprocess.run(
+            [*base_cmd, "session", "rm", session_id],
+            cwd=cwd,
+            env=env,
+            stdout=subprocess.DEVNULL,
+            stderr=subprocess.DEVNULL,
+            text=True,
+            timeout=min(timeout, 30),
+        )
+    except subprocess.TimeoutExpired:
+        pass
+
+
+def summarize(runs: list[dict[str, Any]]) -> dict[str, Any]:
+    out: dict[str, Any] = {}
+    for case in sorted({run["case"] for run in runs}):
+        case_runs = [run for run in runs if run["case"] == case]
+        totals = [int(run["startup_ms"]) for run in case_runs if isinstance(run.get("startup_ms"), int)]
+        stages: dict[str, list[int]] = {}
+        for run in case_runs:
+            for stage in run.get("startup_stages") or []:
+                name = stage.get("name")
+                dur = stage.get("duration_ms")
+                if isinstance(name, str) and isinstance(dur, int):
+                    stages.setdefault(name, []).append(dur)
+        out[case] = {
+            "n": len(case_runs),
+            "startup_ms": stats(totals),
+            "stages": {name: stats(values) for name, values in sorted(stages.items())},
+        }
+    return out
+
+
+def stats(values: list[int]) -> dict[str, Optional[int]]:
+    if not values:
+        return {"min": None, "p50": None, "p95": None, "max": None}
+    ordered = sorted(values)
+    return {
+        "min": ordered[0],
+        "p50": percentile(ordered, 50),
+        "p95": percentile(ordered, 95),
+        "max": ordered[-1],
+    }
+
+
+def percentile(ordered: list[int], pct: int) -> int:
+    if not ordered:
+        raise ValueError("percentile needs at least one value")
+    index = max(
+        0,
+        min(len(ordered) - 1, math.ceil((pct / 100.0) * len(ordered)) - 1),
+    )
+    return ordered[index]
+
+
+def print_human(runs: list[dict[str, Any]], failures: list[dict[str, str]]) -> None:
+    summary = summarize(runs)
+    if not summary:
+        print("(no measured startup events captured)")
+    for case, data in summary.items():
+        total = data["startup_ms"]
+        print(
+            f"{case}: n={data['n']} startup_ms "
+            f"min={total['min']} p50={total['p50']} p95={total['p95']} max={total['max']}"
+        )
+        for stage, st in data["stages"].items():
+            print(f"  {stage:<24} min={st['min']} p50={st['p50']} p95={st['p95']} max={st['max']}")
+    if failures:
+        print("\nfailures:", file=sys.stderr)
+        for fail in failures:
+            print(f"  {fail['case']} [{fail['label']}]: {fail['reason']}", file=sys.stderr)
+
+
+def print_csv(runs: list[dict[str, Any]]) -> None:
+    writer = csv.writer(sys.stdout)
+    writer.writerow(
+        ["case", "session_id", "backend", "agent_id", "startup_ms", "stage_name", "stage_duration_ms"]
+    )
+    for run in runs:
+        stages = run.get("startup_stages") or []
+        if not stages:
+            writer.writerow(
+                [
+                    run["case"],
+                    run["session_id"],
+                    run["backend"],
+                    run["agent_id"],
+                    run.get("startup_ms"),
+                    "",
+                    "",
+                ]
+            )
+            continue
+        for stage in stages:
+            writer.writerow(
+                [
+                    run["case"],
+                    run["session_id"],
+                    run["backend"],
+                    run["agent_id"],
+                    run.get("startup_ms"),
+                    stage.get("name"),
+                    stage.get("duration_ms"),
+                ]
+            )
+
+
+def tail(text: str, limit: int = 1200) -> str:
+    text = text.strip()
+    return text[-limit:] if len(text) > limit else text
+
+
+def tail_timeout_output(exc: subprocess.TimeoutExpired, limit: int = 1200) -> str:
+    parts = []
+    for value in (exc.stderr, exc.stdout):
+        if isinstance(value, bytes):
+            value = value.decode("utf-8", "replace")
+        if isinstance(value, str) and value.strip():
+            parts.append(value)
+    return tail("\n".join(parts), limit) if parts else "no output"
+
+
+def shell_join(cmd: list[str]) -> str:
+    return " ".join(shell_quote(part) for part in cmd)
+
+
+def shell_quote(part: str) -> str:
+    if not part:
+        return "''"
+    safe = set("abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789_+-=.,:/")
+    if all(ch in safe for ch in part):
+        return part
+    return "'" + part.replace("'", "'\\''") + "'"
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/scripts/eval/segmentation/results/README.md b/scripts/eval/segmentation/results/README.md
index 98b74f3..a4bc097 100644
--- a/scripts/eval/segmentation/results/README.md
+++ b/scripts/eval/segmentation/results/README.md
@@ -12,6 +12,16 @@ consumes them lives in [`docs/optimization-gate.md`](../../../../docs/optimizati
 | `h2-segretries0-glm51-n10.jsonl` | H2 retry-isolation, 2026-06-15 | 10 | zai-coding-plan/glm-5.1 | same 3 tasks, `SEG_RETRIES=0`; σ̂ 0.251 → 0.037, lift +0.25 [0.13, 0.33] still excludes zero → retry not the driver |
 | `h4-horizon-reset-glm51-n10.jsonl` | H4 reset-vs-scope, 2026-06-15 | 10 | zai-coding-plan/glm-5.1 | 3 arms (monolithic/chained/segmented); σ̂ 0.198/0.055/0.052; scope alone +0.52 [0.23,0.69], horizon-reset-on-top −0.025 [−0.075,0.0] → **scope is the mechanism, session reset adds nothing** |
 | `enum-control-3task-glm51-n10.jsonl` | enumerated-monolithic control (`ENUM_MONO=1`), 2026-06-19 | 10 | zai-coding-plan/glm-5.1 | 4 arms; σ̂ mono 0.305 / enum 0.109 / chained 0.045 / seg 0.057; prompt-decomp (enum−mono) +0.17, Δσ̂ −0.196; checkpoint-gating (chained−enum) +0.18, Δσ̂ −0.064, pass 11→19/30 → **gating is a real, separable lever — "just a better prompt" refuted** |
+| `control-serial-glm52-pov-n10.jsonl` | H5-inversion refutation, **serial** driver, 2026-06-19 | 10 | zai-coding-plan/glm-5.2 | 3 arms, ap_pov; σ̂ mono 0.292 / chained 0.000 / seg 0.000; mean 0.13 → 1.00. **Refutes the H5 "inversion"**: run serially, glm-5.2 chained σ̂ = 0.000 (== H4) — the cross-model campaign's chained σ̂ 0.317 was a **parallel-driver artifact** ($1.92) |
+| `h5-opencode-go_deepseek-v4-pro-n10.jsonl` | H5 cross-model ⚠️ parallel driver, 2026-06-19 | 10 | deepseek-v4-pro (opencode-go) | 3 tasks × 3 arms; mono 0.50/σ̂0.29 → seg 0.91/σ̂0.15; mean lift **+0.41** ($10.32) |
+| `h5-opencode-go_kimi-k2.7-code-n10.jsonl` | H5 cross-model ⚠️ parallel driver, 2026-06-19 | 10 | kimi-k2.7-code (opencode-go) | 3 tasks × 3 arms; mono 0.80/σ̂0.28 → seg 0.90/σ̂0.17; mean lift **+0.11** (high mono baseline) ($13.79) |
+| `h5-zai-coding-plan_glm-5.2-n10.jsonl` | H5 cross-model ⚠️ parallel driver, 2026-06-19 | 10 | zai-coding-plan/glm-5.2 | 3 tasks × 3 arms; mono 0.47/σ̂0.38 → seg 0.88/σ̂0.19; mean lift **+0.41** ($7.12) |
+
+> ⚠️ **The `h5-*` cross-model files came from the 5-models-concurrent driver.** The
+> concurrency inflated σ̂ via VM boot/drive contention, so treat their per-model **σ̂
+> as an upper bound** — the clean σ̂ is the serial `control-serial-*` run (chained
+> 0.000). The cross-model **mean lift is robust** to contention and holds across all
+> three models (mono → seg: +0.41 / +0.11 / +0.41). Run σ̂ campaigns **serially**.
 
 Re-derive the stats from any file:
 
diff --git a/scripts/eval/segmentation/results/control-serial-glm52-pov-n10.jsonl b/scripts/eval/segmentation/results/control-serial-glm52-pov-n10.jsonl
new file mode 100644
index 0000000..d5b3d4f
--- /dev/null
+++ b/scripts/eval/segmentation/results/control-serial-glm52-pov-n10.jsonl
@@ -0,0 +1,30 @@
+{"task": "ap_pov", "cond": "monolithic", "trial": 1, "score": 0.0, "cost": 0.032649}
+{"task": "ap_pov", "cond": "chained", "trial": 1, "score": 1.0, "cost": 0.091435}
+{"task": "ap_pov", "cond": "segmented", "trial": 1, "score": 1.0, "cost": 0.051192999999999995}
+{"task": "ap_pov", "cond": "monolithic", "trial": 2, "score": 0.0, "cost": 0.013125}
+{"task": "ap_pov", "cond": "chained", "trial": 2, "score": 1.0, "cost": 0.072287}
+{"task": "ap_pov", "cond": "segmented", "trial": 2, "score": 1.0, "cost": 0.087895}
+{"task": "ap_pov", "cond": "monolithic", "trial": 3, "score": 0.933, "cost": 0.090671}
+{"task": "ap_pov", "cond": "chained", "trial": 3, "score": 1.0, "cost": 0.111443}
+{"task": "ap_pov", "cond": "segmented", "trial": 3, "score": 1.0, "cost": 0.067679}
+{"task": "ap_pov", "cond": "monolithic", "trial": 4, "score": 0.0, "cost": 0.005309}
+{"task": "ap_pov", "cond": "chained", "trial": 4, "score": 1.0, "cost": 0.125006}
+{"task": "ap_pov", "cond": "segmented", "trial": 4, "score": 1.0, "cost": 0.06986}
+{"task": "ap_pov", "cond": "monolithic", "trial": 5, "score": 0.0, "cost": 0.013431}
+{"task": "ap_pov", "cond": "chained", "trial": 5, "score": 1.0, "cost": 0.093992}
+{"task": "ap_pov", "cond": "segmented", "trial": 5, "score": 1.0, "cost": 0.066916}
+{"task": "ap_pov", "cond": "monolithic", "trial": 6, "score": 0.0, "cost": 0.018271}
+{"task": "ap_pov", "cond": "chained", "trial": 6, "score": 1.0, "cost": 0.09103}
+{"task": "ap_pov", "cond": "segmented", "trial": 6, "score": 1.0, "cost": 0.117451}
+{"task": "ap_pov", "cond": "monolithic", "trial": 7, "score": 0.0, "cost": 0.012999}
+{"task": "ap_pov", "cond": "chained", "trial": 7, "score": 1.0, "cost": 0.0}
+{"task": "ap_pov", "cond": "segmented", "trial": 7, "score": 1.0, "cost": 0.085784}
+{"task": "ap_pov", "cond": "monolithic", "trial": 8, "score": 0.0, "cost": 0.013296}
+{"task": "ap_pov", "cond": "chained", "trial": 8, "score": 1.0, "cost": 0.07658}
+{"task": "ap_pov", "cond": "segmented", "trial": 8, "score": 1.0, "cost": 0.080613}
+{"task": "ap_pov", "cond": "monolithic", "trial": 9, "score": 0.4, "cost": 0.128307}
+{"task": "ap_pov", "cond": "chained", "trial": 9, "score": 1.0, "cost": 0.105255}
+{"task": "ap_pov", "cond": "segmented", "trial": 9, "score": 1.0, "cost": 0.068423}
+{"task": "ap_pov", "cond": "monolithic", "trial": 10, "score": 0.0, "cost": 0.012558}
+{"task": "ap_pov", "cond": "chained", "trial": 10, "score": 1.0, "cost": 0.056784}
+{"task": "ap_pov", "cond": "segmented", "trial": 10, "score": 1.0, "cost": 0.059369}
diff --git a/scripts/eval/segmentation/results/h5-opencode-go_deepseek-v4-pro-n10.jsonl b/scripts/eval/segmentation/results/h5-opencode-go_deepseek-v4-pro-n10.jsonl
new file mode 100644
index 0000000..5aed3f1
--- /dev/null
+++ b/scripts/eval/segmentation/results/h5-opencode-go_deepseek-v4-pro-n10.jsonl
@@ -0,0 +1,90 @@
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 1, "score": 0.0, "cost": 0.032344}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 1, "score": 0.917, "cost": 0.141349}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 1, "score": 0.667, "cost": 0.147177}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 2, "score": 0.0, "cost": 0.03633}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 2, "score": 0.667, "cost": 0.146986}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 2, "score": 0.667, "cost": 0.16428}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 3, "score": 0.583, "cost": 0.086237}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 3, "score": 0.667, "cost": 0.125122}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 3, "score": 0.667, "cost": 0.154767}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 4, "score": 0.5, "cost": 0.093657}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 4, "score": 0.917, "cost": 0.159142}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 4, "score": 0.667, "cost": 0.142023}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 5, "score": 0.583, "cost": 0.124264}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 5, "score": 0.667, "cost": 0.161917}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 5, "score": 0.5, "cost": 0.135798}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 6, "score": 0.583, "cost": 0.072098}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 6, "score": 0.917, "cost": 0.161697}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 6, "score": 0.917, "cost": 0.150611}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 7, "score": 0.083, "cost": 0.247415}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 7, "score": 0.917, "cost": 0.164992}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 7, "score": 0.667, "cost": 0.216306}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 8, "score": 0.583, "cost": 0.086125}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 8, "score": 0.167, "cost": 0.0}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 8, "score": 0.917, "cost": 0.162578}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 9, "score": 0.083, "cost": 0.102048}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 9, "score": 0.667, "cost": 0.213085}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 9, "score": 0.917, "cost": 0.154304}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 10, "score": 0.583, "cost": 0.091732}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 10, "score": 0.667, "cost": 0.154709}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 10, "score": 0.75, "cost": 0.168482}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 1, "score": 0.75, "cost": 0.111367}
+{"task": "ap_grade_school", "cond": "chained", "trial": 1, "score": 1.0, "cost": 0.110381}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 1, "score": 1.0, "cost": 0.099389}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 2, "score": 0.5, "cost": 0.060658}
+{"task": "ap_grade_school", "cond": "chained", "trial": 2, "score": 1.0, "cost": 0.091303}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 2, "score": 1.0, "cost": 0.119384}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 3, "score": 0.5, "cost": 0.06832}
+{"task": "ap_grade_school", "cond": "chained", "trial": 3, "score": 1.0, "cost": 0.098258}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 3, "score": 1.0, "cost": 0.11217099999999999}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 4, "score": 0.75, "cost": 0.070945}
+{"task": "ap_grade_school", "cond": "chained", "trial": 4, "score": 1.0, "cost": 0.083209}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 4, "score": 1.0, "cost": 0.114643}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 5, "score": 0.5, "cost": 0.072215}
+{"task": "ap_grade_school", "cond": "chained", "trial": 5, "score": 1.0, "cost": 0.072289}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 5, "score": 1.0, "cost": 0.11388499999999999}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 6, "score": 0.5, "cost": 0.049899}
+{"task": "ap_grade_school", "cond": "chained", "trial": 6, "score": 1.0, "cost": 0.095387}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 6, "score": 1.0, "cost": 0.120114}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 7, "score": 0.5, "cost": 0.082006}
+{"task": "ap_grade_school", "cond": "chained", "trial": 7, "score": 1.0, "cost": 0.112602}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 7, "score": 1.0, "cost": 0.115014}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 8, "score": 0.75, "cost": 0.069136}
+{"task": "ap_grade_school", "cond": "chained", "trial": 8, "score": 1.0, "cost": 0.123614}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 8, "score": 1.0, "cost": 0.116203}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 9, "score": 0.5, "cost": 0.096931}
+{"task": "ap_grade_school", "cond": "chained", "trial": 9, "score": 1.0, "cost": 0.100424}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 9, "score": 1.0, "cost": 0.13287100000000002}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 10, "score": 0.5, "cost": 0.050032}
+{"task": "ap_grade_school", "cond": "chained", "trial": 10, "score": 1.0, "cost": 0.085625}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 10, "score": 1.0, "cost": 0.113237}
+{"task": "ap_pov", "cond": "monolithic", "trial": 1, "score": 0.733, "cost": 0.122212}
+{"task": "ap_pov", "cond": "chained", "trial": 1, "score": 1.0, "cost": 0.131951}
+{"task": "ap_pov", "cond": "segmented", "trial": 1, "score": 0.867, "cost": 0.14813900000000002}
+{"task": "ap_pov", "cond": "monolithic", "trial": 2, "score": 0.933, "cost": 0.050135}
+{"task": "ap_pov", "cond": "chained", "trial": 2, "score": 0.533, "cost": 0.09224}
+{"task": "ap_pov", "cond": "segmented", "trial": 2, "score": 1.0, "cost": 0.19605100000000003}
+{"task": "ap_pov", "cond": "monolithic", "trial": 3, "score": 0.0, "cost": 0.035189}
+{"task": "ap_pov", "cond": "chained", "trial": 3, "score": 0.4, "cost": 0.065955}
+{"task": "ap_pov", "cond": "segmented", "trial": 3, "score": 1.0, "cost": 0.153869}
+{"task": "ap_pov", "cond": "monolithic", "trial": 4, "score": 0.0, "cost": 0.035865}
+{"task": "ap_pov", "cond": "chained", "trial": 4, "score": 0.4, "cost": 0.074975}
+{"task": "ap_pov", "cond": "segmented", "trial": 4, "score": 1.0, "cost": 0.15659499999999998}
+{"task": "ap_pov", "cond": "monolithic", "trial": 5, "score": 0.0, "cost": 0.031009}
+{"task": "ap_pov", "cond": "chained", "trial": 5, "score": 1.0, "cost": 0.141648}
+{"task": "ap_pov", "cond": "segmented", "trial": 5, "score": 1.0, "cost": 0.128922}
+{"task": "ap_pov", "cond": "monolithic", "trial": 6, "score": 0.733, "cost": 0.15045}
+{"task": "ap_pov", "cond": "chained", "trial": 6, "score": 1.0, "cost": 0.155336}
+{"task": "ap_pov", "cond": "segmented", "trial": 6, "score": 1.0, "cost": 0.19031599999999999}
+{"task": "ap_pov", "cond": "monolithic", "trial": 7, "score": 0.733, "cost": 0.112585}
+{"task": "ap_pov", "cond": "chained", "trial": 7, "score": 0.533, "cost": 0.082022}
+{"task": "ap_pov", "cond": "segmented", "trial": 7, "score": 1.0, "cost": 0.15221}
+{"task": "ap_pov", "cond": "monolithic", "trial": 8, "score": 0.733, "cost": 0.087855}
+{"task": "ap_pov", "cond": "chained", "trial": 8, "score": 1.0, "cost": 0.146665}
+{"task": "ap_pov", "cond": "segmented", "trial": 8, "score": 1.0, "cost": 0.135879}
+{"task": "ap_pov", "cond": "monolithic", "trial": 9, "score": 0.933, "cost": 0.106641}
+{"task": "ap_pov", "cond": "chained", "trial": 9, "score": 0.4, "cost": 0.114157}
+{"task": "ap_pov", "cond": "segmented", "trial": 9, "score": 1.0, "cost": 0.133425}
+{"task": "ap_pov", "cond": "monolithic", "trial": 10, "score": 0.733, "cost": 0.113005}
+{"task": "ap_pov", "cond": "chained", "trial": 10, "score": 0.4, "cost": 0.06951}
+{"task": "ap_pov", "cond": "segmented", "trial": 10, "score": 1.0, "cost": 0.149174}
diff --git a/scripts/eval/segmentation/results/h5-opencode-go_kimi-k2.7-code-n10.jsonl b/scripts/eval/segmentation/results/h5-opencode-go_kimi-k2.7-code-n10.jsonl
new file mode 100644
index 0000000..4be70bb
--- /dev/null
+++ b/scripts/eval/segmentation/results/h5-opencode-go_kimi-k2.7-code-n10.jsonl
@@ -0,0 +1,90 @@
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 1, "score": 0.583, "cost": 0.263714}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 1, "score": 0.667, "cost": 0.239858}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 1, "score": 0.667, "cost": 0.43297800000000003}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 2, "score": 0.583, "cost": 0.134228}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 2, "score": 0.5, "cost": 0.127307}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 2, "score": 0.917, "cost": 0.133161}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 3, "score": 0.583, "cost": 0.194482}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 3, "score": 0.917, "cost": 0.179029}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 3, "score": 0.667, "cost": 0.240832}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 4, "score": 0.5, "cost": 0.255856}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 4, "score": 0.917, "cost": 0.549721}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 4, "score": 0.667, "cost": 0.37438299999999997}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 5, "score": 0.583, "cost": 0.216631}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 5, "score": 0.917, "cost": 0.338349}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 5, "score": 0.667, "cost": 0.547085}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 6, "score": 0.583, "cost": 0.119545}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 6, "score": 0.917, "cost": 0.225548}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 6, "score": 0.917, "cost": 0.18648800000000001}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 7, "score": 0.0, "cost": 0.011179}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 7, "score": 0.917, "cost": 0.212241}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 7, "score": 0.917, "cost": 0.215362}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 8, "score": 0.583, "cost": 0.175848}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 8, "score": 0.917, "cost": 0.168271}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 8, "score": 0.5, "cost": 0.102829}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 9, "score": 0.583, "cost": 0.079859}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 9, "score": 0.5, "cost": 0.170653}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 9, "score": 0.667, "cost": 0.42117499999999997}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 10, "score": 0.0, "cost": 0.023852}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 10, "score": 0.917, "cost": 0.197244}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 10, "score": 0.417, "cost": 0.269322}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 1, "score": 1.0, "cost": 0.053168}
+{"task": "ap_grade_school", "cond": "chained", "trial": 1, "score": 1.0, "cost": 0.089482}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 1, "score": 1.0, "cost": 0.100163}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 2, "score": 1.0, "cost": 0.061648}
+{"task": "ap_grade_school", "cond": "chained", "trial": 2, "score": 1.0, "cost": 0.089717}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 2, "score": 1.0, "cost": 0.089521}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 3, "score": 1.0, "cost": 0.045618}
+{"task": "ap_grade_school", "cond": "chained", "trial": 3, "score": 1.0, "cost": 0.066207}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 3, "score": 1.0, "cost": 0.10169900000000001}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 4, "score": 1.0, "cost": 0.042883}
+{"task": "ap_grade_school", "cond": "chained", "trial": 4, "score": 1.0, "cost": 0.113449}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 4, "score": 1.0, "cost": 0.10062}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 5, "score": 1.0, "cost": 0.051494}
+{"task": "ap_grade_school", "cond": "chained", "trial": 5, "score": 1.0, "cost": 0.074355}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 5, "score": 1.0, "cost": 0.086843}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 6, "score": 1.0, "cost": 0.051516}
+{"task": "ap_grade_school", "cond": "chained", "trial": 6, "score": 1.0, "cost": 0.047979}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 6, "score": 1.0, "cost": 0.10292100000000001}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 7, "score": 1.0, "cost": 0.051104}
+{"task": "ap_grade_school", "cond": "chained", "trial": 7, "score": 1.0, "cost": 0.096186}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 7, "score": 1.0, "cost": 0.099065}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 8, "score": 1.0, "cost": 0.042686}
+{"task": "ap_grade_school", "cond": "chained", "trial": 8, "score": 1.0, "cost": 0.061836}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 8, "score": 1.0, "cost": 0.057517}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 9, "score": 1.0, "cost": 0.049021}
+{"task": "ap_grade_school", "cond": "chained", "trial": 9, "score": 1.0, "cost": 0.077453}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 9, "score": 1.0, "cost": 0.075845}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 10, "score": 1.0, "cost": 0.074998}
+{"task": "ap_grade_school", "cond": "chained", "trial": 10, "score": 1.0, "cost": 0.052932}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 10, "score": 1.0, "cost": 0.069306}
+{"task": "ap_pov", "cond": "monolithic", "trial": 1, "score": 0.933, "cost": 0.185651}
+{"task": "ap_pov", "cond": "chained", "trial": 1, "score": 0.533, "cost": 0.106338}
+{"task": "ap_pov", "cond": "segmented", "trial": 1, "score": 1.0, "cost": 0.272141}
+{"task": "ap_pov", "cond": "monolithic", "trial": 2, "score": 0.933, "cost": 0.180872}
+{"task": "ap_pov", "cond": "chained", "trial": 2, "score": 1.0, "cost": 0.291128}
+{"task": "ap_pov", "cond": "segmented", "trial": 2, "score": 1.0, "cost": 0.156616}
+{"task": "ap_pov", "cond": "monolithic", "trial": 3, "score": 0.933, "cost": 0.113233}
+{"task": "ap_pov", "cond": "chained", "trial": 3, "score": 1.0, "cost": 0.17191}
+{"task": "ap_pov", "cond": "segmented", "trial": 3, "score": 1.0, "cost": 0.147053}
+{"task": "ap_pov", "cond": "monolithic", "trial": 4, "score": 0.933, "cost": 0.136742}
+{"task": "ap_pov", "cond": "chained", "trial": 4, "score": 1.0, "cost": 0.129588}
+{"task": "ap_pov", "cond": "segmented", "trial": 4, "score": 1.0, "cost": 0.254353}
+{"task": "ap_pov", "cond": "monolithic", "trial": 5, "score": 0.933, "cost": 0.110044}
+{"task": "ap_pov", "cond": "chained", "trial": 5, "score": 1.0, "cost": 0.127067}
+{"task": "ap_pov", "cond": "segmented", "trial": 5, "score": 1.0, "cost": 0.120614}
+{"task": "ap_pov", "cond": "monolithic", "trial": 6, "score": 0.933, "cost": 0.159007}
+{"task": "ap_pov", "cond": "chained", "trial": 6, "score": 1.0, "cost": 0.132126}
+{"task": "ap_pov", "cond": "segmented", "trial": 6, "score": 1.0, "cost": 0.163681}
+{"task": "ap_pov", "cond": "monolithic", "trial": 7, "score": 0.933, "cost": 0.315791}
+{"task": "ap_pov", "cond": "chained", "trial": 7, "score": 1.0, "cost": 0.157985}
+{"task": "ap_pov", "cond": "segmented", "trial": 7, "score": 1.0, "cost": 0.109949}
+{"task": "ap_pov", "cond": "monolithic", "trial": 8, "score": 0.933, "cost": 0.155004}
+{"task": "ap_pov", "cond": "chained", "trial": 8, "score": 1.0, "cost": 0.184958}
+{"task": "ap_pov", "cond": "segmented", "trial": 8, "score": 1.0, "cost": 0.10896600000000001}
+{"task": "ap_pov", "cond": "monolithic", "trial": 9, "score": 0.867, "cost": 0.109895}
+{"task": "ap_pov", "cond": "chained", "trial": 9, "score": 1.0, "cost": 0.148656}
+{"task": "ap_pov", "cond": "segmented", "trial": 9, "score": 1.0, "cost": 0.167844}
+{"task": "ap_pov", "cond": "monolithic", "trial": 10, "score": 0.933, "cost": 0.092343}
+{"task": "ap_pov", "cond": "chained", "trial": 10, "score": 1.0, "cost": 0.155014}
+{"task": "ap_pov", "cond": "segmented", "trial": 10, "score": 1.0, "cost": 0.14584}
diff --git a/scripts/eval/segmentation/results/h5-zai-coding-plan_glm-5.2-n10.jsonl b/scripts/eval/segmentation/results/h5-zai-coding-plan_glm-5.2-n10.jsonl
new file mode 100644
index 0000000..4cbae40
--- /dev/null
+++ b/scripts/eval/segmentation/results/h5-zai-coding-plan_glm-5.2-n10.jsonl
@@ -0,0 +1,90 @@
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 1, "score": 0.0, "cost": 0.112423}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 1, "score": 0.667, "cost": 0.058102}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 1, "score": 0.667, "cost": 0.109179}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 2, "score": 0.083, "cost": 0.172395}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 2, "score": 0.667, "cost": 0.154526}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 2, "score": 0.667, "cost": 0.119821}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 3, "score": 0.083, "cost": 0.224435}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 3, "score": 0.667, "cost": 0.116891}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 3, "score": 0.75, "cost": 0.138405}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 4, "score": 0.083, "cost": 0.076814}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 4, "score": 0.667, "cost": 0.153942}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 4, "score": 0.5, "cost": 0.089349}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 5, "score": 0.083, "cost": 0.053499}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 5, "score": 0.75, "cost": 0.100873}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 5, "score": 0.5, "cost": 0.025668}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 6, "score": 0.0, "cost": 0.017692}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 6, "score": 1.0, "cost": 0.151507}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 6, "score": 0.917, "cost": 0.212633}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 7, "score": 0.083, "cost": 0.067372}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 7, "score": 0.5, "cost": 0.111056}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 7, "score": 0.667, "cost": 0.151573}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 8, "score": 0.083, "cost": 0.049285}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 8, "score": 0.75, "cost": 0.060129}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 8, "score": 0.5, "cost": 0.131219}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 9, "score": 0.083, "cost": 0.111088}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 9, "score": 0.75, "cost": 0.128139}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 9, "score": 0.5, "cost": 0.1076}
+{"task": "ap_dot_dsl", "cond": "monolithic", "trial": 10, "score": 0.083, "cost": 0.139247}
+{"task": "ap_dot_dsl", "cond": "chained", "trial": 10, "score": 0.75, "cost": 0.148684}
+{"task": "ap_dot_dsl", "cond": "segmented", "trial": 10, "score": 0.667, "cost": 0.10191499999999999}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 1, "score": 0.65, "cost": 0.027309}
+{"task": "ap_grade_school", "cond": "chained", "trial": 1, "score": 1.0, "cost": 0.061694}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 1, "score": 1.0, "cost": 0.047206000000000005}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 2, "score": 1.0, "cost": 0.023457}
+{"task": "ap_grade_school", "cond": "chained", "trial": 2, "score": 1.0, "cost": 0.051382}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 2, "score": 1.0, "cost": 0.056610999999999995}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 3, "score": 0.75, "cost": 0.025809}
+{"task": "ap_grade_school", "cond": "chained", "trial": 3, "score": 1.0, "cost": 0.081511}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 3, "score": 1.0, "cost": 0.062561}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 4, "score": 0.75, "cost": 0.029446}
+{"task": "ap_grade_school", "cond": "chained", "trial": 4, "score": 1.0, "cost": 0.075512}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 4, "score": 1.0, "cost": 0.063412}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 5, "score": 0.75, "cost": 0.026068}
+{"task": "ap_grade_school", "cond": "chained", "trial": 5, "score": 1.0, "cost": 0.072312}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 5, "score": 1.0, "cost": 0.070007}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 6, "score": 0.75, "cost": 0.034361}
+{"task": "ap_grade_school", "cond": "chained", "trial": 6, "score": 1.0, "cost": 0.062863}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 6, "score": 1.0, "cost": 0.06772}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 7, "score": 0.75, "cost": 0.037742}
+{"task": "ap_grade_school", "cond": "chained", "trial": 7, "score": 1.0, "cost": 0.052679}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 7, "score": 1.0, "cost": 0.057592000000000004}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 8, "score": 0.75, "cost": 0.028229}
+{"task": "ap_grade_school", "cond": "chained", "trial": 8, "score": 1.0, "cost": 0.083011}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 8, "score": 1.0, "cost": 0.10345399999999999}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 9, "score": 0.75, "cost": 0.028483}
+{"task": "ap_grade_school", "cond": "chained", "trial": 9, "score": 1.0, "cost": 0.072627}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 9, "score": 1.0, "cost": 0.06532}
+{"task": "ap_grade_school", "cond": "monolithic", "trial": 10, "score": 0.75, "cost": 0.033972}
+{"task": "ap_grade_school", "cond": "chained", "trial": 10, "score": 1.0, "cost": 0.068537}
+{"task": "ap_grade_school", "cond": "segmented", "trial": 10, "score": 1.0, "cost": 0.048803}
+{"task": "ap_pov", "cond": "monolithic", "trial": 1, "score": 0.0, "cost": 0.024281}
+{"task": "ap_pov", "cond": "chained", "trial": 1, "score": 1.0, "cost": 0.120206}
+{"task": "ap_pov", "cond": "segmented", "trial": 1, "score": 1.0, "cost": 0.067585}
+{"task": "ap_pov", "cond": "monolithic", "trial": 2, "score": 0.8, "cost": 0.039488}
+{"task": "ap_pov", "cond": "chained", "trial": 2, "score": 0.533, "cost": 0.056386}
+{"task": "ap_pov", "cond": "segmented", "trial": 2, "score": 1.0, "cost": 0.096887}
+{"task": "ap_pov", "cond": "monolithic", "trial": 3, "score": 0.933, "cost": 0.068128}
+{"task": "ap_pov", "cond": "chained", "trial": 3, "score": 0.0, "cost": 0.068052}
+{"task": "ap_pov", "cond": "segmented", "trial": 3, "score": 1.0, "cost": 0.056244999999999996}
+{"task": "ap_pov", "cond": "monolithic", "trial": 4, "score": 0.0, "cost": 0.012771}
+{"task": "ap_pov", "cond": "chained", "trial": 4, "score": 1.0, "cost": 0.102978}
+{"task": "ap_pov", "cond": "segmented", "trial": 4, "score": 1.0, "cost": 0.07569200000000001}
+{"task": "ap_pov", "cond": "monolithic", "trial": 5, "score": 0.933, "cost": 0.054598}
+{"task": "ap_pov", "cond": "chained", "trial": 5, "score": 1.0, "cost": 0.077884}
+{"task": "ap_pov", "cond": "segmented", "trial": 5, "score": 1.0, "cost": 0.12985}
+{"task": "ap_pov", "cond": "monolithic", "trial": 6, "score": 0.0, "cost": 0.012669}
+{"task": "ap_pov", "cond": "chained", "trial": 6, "score": 1.0, "cost": 0.110582}
+{"task": "ap_pov", "cond": "segmented", "trial": 6, "score": 1.0, "cost": 0.057485}
+{"task": "ap_pov", "cond": "monolithic", "trial": 7, "score": 0.933, "cost": 0.055636}
+{"task": "ap_pov", "cond": "chained", "trial": 7, "score": 1.0, "cost": 0.07025}
+{"task": "ap_pov", "cond": "segmented", "trial": 7, "score": 1.0, "cost": 0.06523999999999999}
+{"task": "ap_pov", "cond": "monolithic", "trial": 8, "score": 0.467, "cost": 0.037643}
+{"task": "ap_pov", "cond": "chained", "trial": 8, "score": 1.0, "cost": 0.074312}
+{"task": "ap_pov", "cond": "segmented", "trial": 8, "score": 1.0, "cost": 0.075598}
+{"task": "ap_pov", "cond": "monolithic", "trial": 9, "score": 0.933, "cost": 0.070199}
+{"task": "ap_pov", "cond": "chained", "trial": 9, "score": 1.0, "cost": 0.093794}
+{"task": "ap_pov", "cond": "segmented", "trial": 9, "score": 1.0, "cost": 0.08077999999999999}
+{"task": "ap_pov", "cond": "monolithic", "trial": 10, "score": 0.8, "cost": 0.107426}
+{"task": "ap_pov", "cond": "chained", "trial": 10, "score": 1.0, "cost": 0.068282}
+{"task": "ap_pov", "cond": "segmented", "trial": 10, "score": 1.0, "cost": 0.070141}