Recalibrate public SQS and exposure-aware tracking

5 months ago · c646303423
parent 9ec0b26e10
commit c646303423
13 changed files with 2129 additions and 386 deletions
--- a/README.md
+++ b/README.md
@ -491,41 +491,50 @@ journal/
 ```
 1. 실험 config 생성       configs/experiments/my_experiment.json
 2. 3-split 백테스트 실행   for split in train valid test; do ... done
-3. Journal에 기록          python -m apps.tracker.cli record ...
+3. Journal에 기록          fithia2 rec -e my_experiment ...
-4. Leaderboard 확인        python -m apps.tracker.cli leaderboard ...
+4. Leaderboard 확인        fithia2 lb
 5. 다음 실험 계획          verdict + SQS breakdown 기반
-6. 중복 실험 확인          python -m apps.tracker.cli check-duplicate ...
+6. 중복 실험 확인          fithia2 dup -e my_experiment
 7. 1번부터 반복
 ```
 ### 8.5 Tracker CLI
 `pip install -e .` 후 `fithia2` 명령어로 실행. `--journal-dir` / `--runs-dir` 생략 시 기본값(`journal/`, `runs/`) 사용.
 ```bash
 # 도움말
 fithia2
 # 리더보드 출력 (top 10)
 fithia2 lb
 # 리더보드 — 상위 N개
 fithia2 lb -n 20
 # 실험 결과 기록
-python -m apps.tracker.cli record \
+fithia2 rec \
-  --journal-dir journal/ \
+  -e pead_midcap_step14_score65 \
-  --runs-dir runs/midcap_steps/ \
+  -H "Score threshold 0.60→0.65" \
-  --experiment pead_midcap_step14_score65 \
+  -b pead_midcap_step13_best \
-  --hypothesis "Score threshold 0.60→0.65" \
+  -v better \
-  --baseline pead_midcap_step13_best \
+  -r "Test SQS 64.2 > 57.7, Return +0.73% > +0.43%" \
-  --verdict better \
+  -n "Test exit tuning on top of score 0.65"
  --reasoning "Test SQS 64.2 > 57.7, Return +0.73% > +0.43%" \
  --next "Test exit tuning on top of score 0.65"
 # 리더보드 출력
 python -m apps.tracker.cli leaderboard --journal-dir journal/
 # 엔트리 상세 조회
-python -m apps.tracker.cli show --journal-dir journal/ IMP-0015
+fithia2 s IMP-0015
 # 중복 실험 확인
-python -m apps.tracker.cli check-duplicate \
+fithia2 dup -e my_experiment_name
  --journal-dir journal/ --experiment my_experiment_name
 # 레지스트리 재생성
 python -m apps.tracker.cli rebuild-registry --journal-dir journal/
 ```
 | 명령 | alias | 설명 |
 |------|-------|------|
 | `leaderboard` | `lb` | SQS 순위표 출력 및 LEADERBOARD.md 재생성 |
 | `record` | `rec` | 실험 결과를 저널에 기록 |
 | `show` | `s` | 특정 저널 항목 상세 조회 |
 | `check-duplicate` | `dup` | 동일 실험명 중복 여부 확인 |
 ### 8.6 핵심 발견사항
 21개 실험을 통해 얻은 인사이트:
@ -695,18 +704,16 @@ done
 ```bash
 # Journal에 기록
-python -m apps.tracker.cli record \
+fithia2 rec \
-  --journal-dir journal/ \
+  -e pead_midcap_step14_score65 \
-  --runs-dir runs/midcap_steps/ \
+  -H "Score threshold 0.60→0.65" \
-  --experiment pead_midcap_step14_score65 \
+  -b pead_midcap_step13_best \
-  --hypothesis "Score threshold 0.60→0.65" \
+  -v better \
-  --baseline pead_midcap_step13_best \
+  -r "Test SQS 64.2 > 57.7" \
-  --verdict better \
+  -n "Exit tuning on top of score 0.65"
  --reasoning "Test SQS 64.2 > 57.7" \
  --next "Exit tuning on top of score 0.65"
 # 리더보드 확인
-python -m apps.tracker.cli leaderboard --journal-dir journal/
+fithia2 lb
 ```
 ### 테스트
--- a/apps/backtester/run.py
+++ b/apps/backtester/run.py
@ -93,6 +93,22 @@ class BacktestRunner:
        self._kill_switch_cooldown_remaining = 0
        self._engine_daily_new_risk_used: dict[str, float] = defaultdict(float)
    def _compute_portfolio_exposure(self, date: dt.date) -> tuple[float, float]:
        """Return (gross, net) exposure using current close notional when available."""
        gross = 0.0
        net = 0.0
        for pos in self._open_positions:
            bar = self.store.get_bar(pos.plan.candidate.symbol, date)
            close = (
                float(bar["close"])
                if bar and bar.get("close") is not None and float(bar["close"]) > 0
                else pos.entry_price
            )
            notional = close * pos.shares_open
            gross += abs(notional)
            net += -notional if pos.plan.candidate.trade_direction == "short" else notional
        return gross, net
    def run(self, output_root: str | Path | None = None) -> ExperimentResult:
        """Execute the full simulation. Returns ExperimentResult."""
        started_at = utc_now()
@ -369,17 +385,14 @@ class BacktestRunner:
            if self._peak_equity > 0
            else 0.0
        )
        gross_exposure, net_exposure = self._compute_portfolio_exposure(date)
        self._equity_curve.append(
            DailyPortfolioState(
                date=date,
                equity=self._equity,
                cash_available=self._cash,
-                gross_exposure=sum(
+                gross_exposure=gross_exposure,
-                    p.entry_price * p.shares_open for p in self._open_positions
+                net_exposure=net_exposure,
                ),
                net_exposure=sum(
                    p.entry_price * p.shares_open for p in self._open_positions
                ),
                reserved_risk_budget=self._daily_new_risk_used,
                unrealized_pnl=unrealized_final,
                realized_pnl=self._realized_pnl,
@ -563,16 +576,13 @@ class BacktestRunner:
        drawdown_pct: float,
        unrealized: float,
    ) -> DailyPortfolioState:
        gross_exposure, net_exposure = self._compute_portfolio_exposure(date)
        return DailyPortfolioState(
            date=date,
            equity=self._equity,
            cash_available=self._cash,
-            gross_exposure=sum(
+            gross_exposure=gross_exposure,
-                p.entry_price * p.shares_open for p in self._open_positions
+            net_exposure=net_exposure,
            ),
            net_exposure=sum(
                p.entry_price * p.shares_open for p in self._open_positions
            ),
            reserved_risk_budget=self._daily_new_risk_used,
            unrealized_pnl=unrealized,
            realized_pnl=self._realized_pnl,
--- a/apps/tracker/cli.py
+++ b/apps/tracker/cli.py
@ -5,6 +5,14 @@ import argparse
 import sys
 from pathlib import Path
 from rich import box
 from rich.console import Console
 from rich.panel import Panel
 from rich.table import Table
 from rich.text import Text
 _console = Console(width=140)
 from libs.backtest.domain import (
    ConfigDelta,
    JournalEntry,
@ -15,7 +23,11 @@ from libs.backtest.tracker import (
    append_journal_entry,
    build_split_result,
    check_duplicate,
    compute_public_sqs,
    compute_promotion_score,
    compute_sqs,
    compute_sqs_v2,
    compute_unified_score,
    get_next_entry_id,
    load_journal,
    rebuild_registry,
@ -54,17 +66,41 @@ def cmd_record(args: argparse.Namespace) -> None:
    for split_name, (run_id, metrics) in split_runs.items():
        results[split_name] = build_split_result(split_name, run_id, metrics)
-    # Compute SQS from test split (or best available)
+    # Compute public SQS from valid+test when available, or test split fallback.
    test_metrics = None
    for preferred in ["test", "valid", "train"]:
        if preferred in split_runs:
            _, test_metrics = split_runs[preferred]
            break
-    sqs_score = 0.0
+    sqs_score = None
    sqs_breakdown: dict[str, float] = {}
    sqs_v2_score = None
    sqs_v2_breakdown: dict[str, float] = {}
    promotion_score = None
    promotion_breakdown: dict[str, float] = {}
    unified_score = None
    unified_breakdown: dict[str, float] = {}
    if test_metrics:
-        sqs_score, sqs_breakdown = compute_sqs(test_metrics)
+        legacy_sqs_score, legacy_sqs_breakdown = compute_sqs(test_metrics)
        sqs_v2_score, sqs_v2_breakdown = compute_sqs_v2(test_metrics)
        sqs_score = legacy_sqs_score
        sqs_breakdown = legacy_sqs_breakdown
    promotion_score, promotion_breakdown = compute_promotion_score(
        results.get("test"),
        results.get("valid"),
    )
    public_sqs, public_breakdown, public_source = compute_public_sqs(
        results.get("test"),
        results.get("valid"),
    )
    if public_sqs is not None:
        sqs_score = public_sqs
        sqs_breakdown = public_breakdown
    unified_score, unified_breakdown = compute_unified_score(
        results.get("test"),
        results.get("valid"),
    )
    # Build config delta
    config_delta = None
@ -84,6 +120,12 @@ def cmd_record(args: argparse.Namespace) -> None:
        results=results,
        sqs_score=sqs_score,
        sqs_breakdown=sqs_breakdown,
        sqs_v2_score=sqs_v2_score,
        sqs_v2_breakdown=sqs_v2_breakdown,
        promotion_score=promotion_score,
        promotion_breakdown=promotion_breakdown,
        unified_score=unified_score,
        unified_breakdown=unified_breakdown,
        verdict=args.verdict or "unknown",
        verdict_reasoning=args.reasoning or "",
        next_direction=args.next or "",
@ -91,7 +133,8 @@ def cmd_record(args: argparse.Namespace) -> None:
    )
    append_journal_entry(journal_path, entry)
-    print(f"Recorded {entry_id}: {args.experiment} (SQS={sqs_score})")
+    source_label = f", source={public_source}" if public_source else ""
    print(f"Recorded {entry_id}: {args.experiment} (SQS={sqs_score}{source_label})")
    # Show splits found
    for split_name, sr in results.items():
@ -104,6 +147,15 @@ def cmd_record(args: argparse.Namespace) -> None:
    print(f"Leaderboard updated: {leaderboard_path}")
 _DEFAULT_JOURNAL_DIR = "journal"
 _COL_NAME = 38  # max experiment name width before truncation
 def _fmt(val: float | None, fmt: str) -> str:
    return format(val, fmt) if val is not None else "-"
 def cmd_leaderboard(args: argparse.Namespace) -> None:
    """Show or regenerate the leaderboard."""
    journal_dir = Path(args.journal_dir)
@ -117,13 +169,66 @@ def cmd_leaderboard(args: argparse.Namespace) -> None:
    registry = rebuild_registry(journal_path, registry_path, leaderboard_path)
-    # Print to console
+    top_n = getattr(args, "top", 10)
-    print(f"\n{'#':>3} {'Experiment':<40} {'SQS':>5} {'PF':>5} {'Ret%':>6} {'Trades':>6}")
+    sort_by = getattr(args, "sort", "sqs")
-    print("-" * 70)
+    if sort_by == "promotion":
-    for rank, e in enumerate(registry.entries, 1):
+        ranked_entries = sorted(
-        pf = f"{e.profit_factor:.2f}" if e.profit_factor is not None else "-"
+            registry.entries,
-        ret = f"{e.total_return_pct:+.1f}" if e.total_return_pct is not None else "-"
+            key=lambda e: (e.promotion_score is None, -(e.promotion_score or 0.0), -e.sqs_score),
-        print(f"{rank:>3} {e.experiment_name:<40} {e.sqs_score:>5.1f} {pf:>5} {ret:>6} {e.trade_count:>6}")
+        )
        title_mode = "Promotion 기준 내림차순"
    elif sort_by in {"sqs", "unified"}:
        ranked_entries = sorted(registry.entries, key=lambda e: -e.sqs_score)
        title_mode = "SQS 기준 내림차순"
    else:
        ranked_entries = sorted(registry.entries, key=lambda e: -e.sqs_score)
        title_mode = "SQS 기준 내림차순"
    total = len(ranked_entries)
    tbl = Table(
        box=box.SIMPLE_HEAD,
        show_header=True,
        header_style="bold yellow",
        row_styles=["", "dim"],
        padding=(0, 1),
        title=f"[bold cyan]Top {top_n} / {total}[/]  [dim]· {title_mode} · T=test  V=valid[/]",
        title_justify="left",
        expand=False,
    )
    # Columns: #, Experiment, SQS | Test: PF / Ret% / WR / DD% / N | Valid: PF / Ret% / WR / N
    tbl.add_column("#",              justify="right", style="bold",      no_wrap=True, min_width=3)
    tbl.add_column("Experiment",                                          no_wrap=True, min_width=30)
    tbl.add_column("SQS",            justify="right", style="bold cyan", no_wrap=True, min_width=5)
    tbl.add_column("T.PF",           justify="right",                    no_wrap=True, min_width=5)
    tbl.add_column("T.Ret%",         justify="right",                    no_wrap=True, min_width=6)
    tbl.add_column("T.WR%",          justify="right",                    no_wrap=True, min_width=5)
    tbl.add_column("T.DD%",          justify="right",                    no_wrap=True, min_width=5)
    tbl.add_column("T.N",            justify="right",                    no_wrap=True, min_width=4)
    tbl.add_column("V.PF",           justify="right", style="green",    no_wrap=True, min_width=5)
    tbl.add_column("V.Ret%",         justify="right", style="green",    no_wrap=True, min_width=6)
    tbl.add_column("V.WR%",          justify="right", style="green",    no_wrap=True, min_width=5)
    tbl.add_column("V.N",            justify="right", style="green",    no_wrap=True, min_width=4)
    for rank, e in enumerate(ranked_entries[:top_n], 1):
        name = e.experiment_name
        if len(name) > _COL_NAME:
            name = name[: _COL_NAME - 1] + "…"
        pf   = _fmt(e.profit_factor, ".2f")
        ret  = _fmt(e.total_return_pct, "+.1f")
        wr   = _fmt(e.win_rate * 100 if e.win_rate is not None else None, ".0f")
        dd   = _fmt(e.max_drawdown_pct, ".1f")
        vpf  = _fmt(e.valid_profit_factor, ".2f")
        vret = _fmt(e.valid_total_return_pct, "+.1f")
        vwr  = _fmt(e.valid_win_rate * 100 if e.valid_win_rate is not None else None, ".0f")
        tbl.add_row(
            str(rank), name, f"{e.sqs_score:.1f}",
            pf, ret, wr, dd, str(e.trade_count),
            vpf, vret, vwr, str(e.valid_trade_count) if e.valid_trade_count else "-",
        )
    _console.print()
    _console.print(tbl)
    _console.print(f"  [dim]LEADERBOARD.md → {leaderboard_path}[/]\n")
 def cmd_show(args: argparse.Namespace) -> None:
@ -144,11 +249,28 @@ def cmd_show(args: argparse.Namespace) -> None:
        sys.exit(1)
    for e in found:
        promotion_score = e.promotion_score
        promotion_breakdown = e.promotion_breakdown
        if promotion_score is None:
            promotion_score, promotion_breakdown = compute_promotion_score(
                e.results.get("test"),
                e.results.get("valid"),
            )
        public_sqs, public_breakdown, public_source = compute_public_sqs(
            e.results.get("test"),
            e.results.get("valid"),
        )
        if public_sqs is None:
            public_sqs = e.sqs_score
            public_breakdown = e.sqs_breakdown
            public_source = "stored"
        print(f"\n{e.entry_id} — {e.experiment_name}")
        print(f"  Timestamp:  {e.timestamp}")
        print(f"  Hypothesis: {e.hypothesis}")
        print(f"  Verdict:    {e.verdict}")
-        print(f"  SQS:        {e.sqs_score} {e.sqs_breakdown}")
+        print(f"  SQS:        {public_sqs} {public_breakdown} [{public_source}]")
        if promotion_score is not None:
            print(f"  Promotion:  {promotion_score} {promotion_breakdown} [internal]")
        if e.config_delta:
            print(f"  Baseline:   {e.config_delta.base_experiment}")
            for k, v in e.config_delta.changes.items():
@ -177,43 +299,107 @@ def cmd_check_duplicate(args: argparse.Namespace) -> None:
        print(f"No existing entries for '{args.experiment}'.")
 def _print_help() -> None:
    _console.print()
    _console.print(Panel(
        "[bold cyan]fithia2[/] — ACE-F Strategy Improvement Tracker\n"
        "[dim]백테스트 실험을 기록하고 전략 품질 점수(SQS)로 순위를 매깁니다.[/]",
        border_style="cyan",
        padding=(0, 2),
    ))
    t = Table(box=box.SIMPLE, show_header=True, header_style="bold yellow", padding=(0, 2))
    t.add_column("Command", style="bold green", no_wrap=True)
    t.add_column("Description")
    t.add_column("Key Options", style="dim")
    t.add_row(
        "leaderboard  [dim]lb[/]",
        "SQS 순위표 출력 및 LEADERBOARD.md 재생성",
        "-n N  --sort sqs|promotion",
    )
    t.add_row(
        "record  [dim]rec[/]",
        "실험 결과를 저널에 기록",
        "-e NAME  -H TEXT  -v better|worse|neutral  -b BASELINE",
    )
    t.add_row(
        "show  [dim]s[/]",
        "특정 저널 항목 상세 조회",
        "ENTRY_ID  (예: IMP-0001 또는 실험명)",
    )
    t.add_row(
        "check-duplicate  [dim]dup[/]",
        "동일 실험명이 이미 기록됐는지 확인",
        "-e NAME",
    )
    _console.print(t)
    _console.print(
        "  [dim]공통 옵션:[/]  [bold]--journal-dir[/] [dim](기본: journal/)[/]"
        "  [bold]--runs-dir[/] [dim](기본: runs/)[/]\n"
    )
    _console.print("  [bold]예시[/]")
    _console.print("    [green]fithia2 leaderboard[/]")
    _console.print("    [green]fithia2 leaderboard --top 20[/]")
    _console.print("    [green]fithia2 leaderboard --sort sqs[/]")
    _console.print("    [green]fithia2 leaderboard --sort promotion[/]")
    _console.print("    [green]fithia2 record --experiment pead_v2 --hypothesis '...' --verdict better[/]")
    _console.print("    [green]fithia2 show IMP-0007[/]")
    _console.print()
 def main() -> None:
-    parser = argparse.ArgumentParser(description="Strategy Improvement Tracker")
+    if len(sys.argv) == 1:
        _print_help()
        sys.exit(0)
    parser = argparse.ArgumentParser(description="Strategy Improvement Tracker", add_help=True)
    sub = parser.add_subparsers(dest="command", required=True)
-    # record
+    _jdir_kwargs = {"default": _DEFAULT_JOURNAL_DIR, "help": f"Path to journal/ directory (default: {_DEFAULT_JOURNAL_DIR})"}
-    p_record = sub.add_parser("record", help="Record an experiment to the journal")
+
-    p_record.add_argument("--journal-dir", required=True, help="Path to journal/ directory")
+    # record  (alias: rec)
-    p_record.add_argument("--runs-dir", required=True, help="Path to runs output directory")
+    for name in ("record", "rec"):
-    p_record.add_argument("--experiment", required=True, help="Experiment name (matches manifest)")
+        p = sub.add_parser(name, help="Record an experiment to the journal")
-    p_record.add_argument("--hypothesis", help="What you expected this change to do")
+        p.add_argument("--journal-dir", **_jdir_kwargs)
-    p_record.add_argument("--baseline", help="Baseline experiment name for comparison")
+        p.add_argument("--runs-dir", default="runs", help="Path to runs output directory (default: runs)")
-    p_record.add_argument("--verdict", choices=["better", "worse", "neutral", "unknown"], default="unknown")
+        p.add_argument("--experiment", "-e", required=True, help="Experiment name (matches manifest)")
-    p_record.add_argument("--reasoning", help="Why this verdict")
+        p.add_argument("--hypothesis", "-H", help="What you expected this change to do")
-    p_record.add_argument("--next", help="Next experiment direction")
+        p.add_argument("--baseline", "-b", help="Baseline experiment name for comparison")
-    p_record.add_argument("--force", action="store_true", help="Allow duplicate experiment names")
+        p.add_argument("--verdict", "-v", choices=["better", "worse", "neutral", "unknown"], default="unknown")
-
+        p.add_argument("--reasoning", "-r", help="Why this verdict")
-    # leaderboard
+        p.add_argument("--next", "-n", help="Next experiment direction")
-    p_lb = sub.add_parser("leaderboard", help="Show/regenerate the leaderboard")
+        p.add_argument("--force", "-f", action="store_true", help="Allow duplicate experiment names")
-    p_lb.add_argument("--journal-dir", required=True, help="Path to journal/ directory")
+
-
+    # leaderboard  (alias: lb)
-    # show
+    for name in ("leaderboard", "lb"):
-    p_show = sub.add_parser("show", help="Show details of a journal entry")
+        p = sub.add_parser(name, help="Show/regenerate the leaderboard")
-    p_show.add_argument("--journal-dir", required=True, help="Path to journal/ directory")
+        p.add_argument("--journal-dir", **_jdir_kwargs)
-    p_show.add_argument("entry_id", help="Entry ID (IMP-0001) or experiment name")
+        p.add_argument("--top", "-n", type=int, default=10, help="Show top N entries (default: 10)")
-
+        p.add_argument("--sort", choices=["sqs", "promotion", "unified"], default="sqs", help="Sort leaderboard by public SQS (default) or internal promotion score")
-    # check-duplicate
+
-    p_dup = sub.add_parser("check-duplicate", help="Check if experiment already recorded")
+    # show  (alias: s)
-    p_dup.add_argument("--journal-dir", required=True, help="Path to journal/ directory")
+    for name in ("show", "s"):
-    p_dup.add_argument("--experiment", required=True, help="Experiment name to check")
+        p = sub.add_parser(name, help="Show details of a journal entry")
        p.add_argument("--journal-dir", **_jdir_kwargs)
        p.add_argument("entry_id", help="Entry ID (IMP-0001) or experiment name")
    # check-duplicate  (alias: dup)
    for name in ("check-duplicate", "dup"):
        p = sub.add_parser(name, help="Check if experiment already recorded")
        p.add_argument("--journal-dir", **_jdir_kwargs)
        p.add_argument("--experiment", "-e", required=True, help="Experiment name to check")
    args = parser.parse_args()
    dispatch = {
-        "record": cmd_record,
+        "record": cmd_record,   "rec": cmd_record,
-        "leaderboard": cmd_leaderboard,
+        "leaderboard": cmd_leaderboard, "lb": cmd_leaderboard,
-        "show": cmd_show,
+        "show": cmd_show,       "s": cmd_show,
-        "check-duplicate": cmd_check_duplicate,
+        "check-duplicate": cmd_check_duplicate, "dup": cmd_check_duplicate,
    }
    dispatch[args.command](args)
--- a/journal/LEADERBOARD.md
+++ b/journal/LEADERBOARD.md
@ -1,84 +1,84 @@
 # Strategy Improvement Leaderboard
-_Updated: 2026-03-17T07:57:00.128792+00:00_
+_Updated: 2026-03-17T09:17:27.726326+00:00_
-| # | Experiment | SQS | PF | Ret% | WR | Sharpe | DD% | Trades | Date |
+| # | Experiment | SQS | [T]PF | [T]Ret% | [T]WR | [T]Sharpe | [T]DD% | [T]N | [T]Gross% | [T]Net% | [T]DIM% | [V]PF | [V]Ret% | [V]WR | [V]Sharpe | [V]DD% | [V]N | [V]Gross% | [V]Net% | [V]DIM% | Date |
-|---|-----------|-----|-----|------|-----|--------|-----|--------|------|
+|---|-----------|-----|-------|---------|-------|-----------|--------|------|-----------|---------|---------|-------|---------|-------|-----------|--------|------|-----------|---------|---------|------|
-| 1 | pead_midcap_step35_balanced_sleeves_nofrac_aclong12 | 90.4 | 2.01 | +2.2 | 61% | 4.2 | 0.4 | 56 | 2026-03-17 |
+| 1 | pead_midcap_step52_short_core_macro_block_crashcap_gap10 | 49.8 | 3.35 | +1.0 | 77% | 3.0 | 0.2 | 22 | 2.9 | -2.2 | 53.2 | 5.66 | +1.8 | 72% | 4.8 | 0.2 | 25 | 3.6 | -2.6 | 56.1 | 2026-03-17 |
-| 2 | pead_midcap_step36_balanced_sleeves_nofrac_aclong12_sdlong12 | 90.3 | 2.06 | +2.1 | 61% | 4.5 | 0.4 | 54 | 2026-03-17 |
+| 2 | pead_midcap_step48_short_core_macro_block_nolong | 46.9 | 3.40 | +0.8 | 80% | 2.4 | 0.4 | 20 | 2.6 | +2.6 | 52.2 | 5.73 | +1.3 | 70% | 3.4 | 0.3 | 20 | 3.3 | +3.3 | 51.8 | 2026-03-17 |
-| 3 | pead_midcap_step44_short_core_macro50 | 87.9 | 2.01 | +1.3 | 57% | 2.6 | 0.7 | 58 | 2026-03-17 |
+| 3 | pead_midcap_step46_short_core_macro_block_acshort12 | 45.8 | 4.52 | +1.3 | 71% | 3.6 | 0.3 | 21 | 2.1 | +2.1 | 40.4 | 2.44 | +1.3 | 69% | 3.2 | 0.4 | 26 | 4.0 | +4.0 | 57.9 | 2026-03-17 |
-| 4 | pead_midcap_step47_short_core_macro_block_sdlong25 | 87.0 | 3.43 | +1.3 | 68% | 3.3 | 0.3 | 25 | 2026-03-17 |
+| 4 | pead_midcap_step36_balanced_sleeves_nofrac_aclong12_sdlong12 | 44.5 | 2.06 | +2.1 | 61% | 4.5 | 0.4 | 54 | 6.0 | +6.0 | 76.6 | - | - | - | - | - | 0 | - | - | - | 2026-03-17 |
-| 5 | pead_midcap_step45_short_core_macro_block | 86.7 | 3.78 | +1.2 | 70% | 3.4 | 0.3 | 23 | 2026-03-17 |
+| 5 | pead_midcap_step51_short_core_macro_block_crashcap | 44.4 | 4.19 | +1.2 | 73% | 3.6 | 0.2 | 22 | 2.3 | -0.8 | 42.6 | 2.31 | +1.3 | 68% | 3.0 | 0.4 | 28 | 4.1 | -1.9 | 57.9 | 2026-03-17 |
-| 6 | pead_midcap_step51_short_core_macro_block_crashcap | 86.7 | 4.19 | +1.2 | 73% | 3.6 | 0.2 | 22 | 2026-03-17 |
+| 6 | pead_midcap_step45_short_core_macro_block | 44.3 | 3.78 | +1.2 | 70% | 3.4 | 0.3 | 23 | 2.4 | +2.4 | 42.6 | 2.31 | +1.3 | 68% | 3.0 | 0.4 | 28 | 4.2 | +4.2 | 57.9 | 2026-03-17 |
-| 7 | pead_midcap_step46_short_core_macro_block_acshort12 | 86.6 | 4.52 | +1.3 | 71% | 3.6 | 0.3 | 21 | 2026-03-17 |
+| 7 | pead_midcap_step53_short_core_macro_block_crashcap_gap14 | 39.5 | 4.60 | +1.1 | 84% | 3.5 | 0.2 | 19 | 2.3 | +2.3 | 53.2 | 8.21 | +2.0 | 76% | 5.3 | 0.2 | 25 | 3.6 | +3.6 | 56.1 | 2026-03-17 |
-| 8 | pead_midcap_step52_short_core_macro_block_crashcap_gap10 | 86.3 | 3.35 | +1.0 | 77% | 3.0 | 0.2 | 22 | 2026-03-17 |
+| 8 | pead_midcap_step50_same_day_short_long_macro_block | 36.2 | 3.21 | +0.8 | 60% | 2.7 | 0.4 | 15 | 1.6 | +1.6 | 38.3 | 2.21 | +1.0 | 65% | 2.7 | 0.3 | 20 | 3.4 | +3.4 | 43.9 | 2026-03-17 |
-| 9 | pead_midcap_step48_short_core_macro_block_nolong | 85.7 | 3.40 | +0.8 | 80% | 2.4 | 0.4 | 20 | 2026-03-17 |
+| 9 | pead_midcap_step30_balanced_sleeves_nofrac | 35.4 | 1.41 | +1.4 | 54% | 2.8 | 0.8 | 67 | 8.0 | +8.0 | 76.6 | - | - | - | - | - | 0 | - | - | - | 2026-03-17 |
-| 10 | pead_midcap_step40_short_core_sdlong12 | 82.1 | 1.77 | +1.5 | 58% | 2.2 | 1.1 | 55 | 2026-03-17 |
+| 10 | pead_midcap_step47_short_core_macro_block_sdlong25 | 35.2 | 3.43 | +1.3 | 68% | 3.3 | 0.3 | 25 | 2.9 | +2.9 | 48.9 | 1.77 | +1.0 | 65% | 2.2 | 0.4 | 31 | 4.6 | +4.6 | 57.9 | 2026-03-17 |
-| 11 | pead_midcap_step34_balanced_sleeves_nofrac_aclong25 | 81.3 | 1.62 | +1.8 | 56% | 3.8 | 0.6 | 62 | 2026-03-17 |
+| 11 | pead_midcap_step44_short_core_macro50 | 35.0 | 2.01 | +1.3 | 57% | 2.6 | 0.7 | 58 | 4.5 | -1.6 | 72.3 | 1.78 | +1.1 | 55% | 2.5 | 0.4 | 40 | 4.8 | -2.0 | 73.7 | 2026-03-17 |
-| 12 | pead_midcap_step43_short_core_sdlong25 | 78.9 | 1.66 | +1.5 | 57% | 2.0 | 1.3 | 58 | 2026-03-17 |
+| 12 | pead_midcap_step33_balanced_sleeves_nofrac_acshort6 | 35.0 | 1.47 | +1.2 | 52% | 2.3 | 0.7 | 48 | 6.2 | +6.2 | 70.2 | - | - | - | - | - | 0 | - | - | - | 2026-03-17 |
-| 13 | pead_midcap_step31_balanced_sleeves_nofrac_acshort12 | 77.9 | 1.49 | +1.6 | 55% | 3.2 | 0.7 | 64 | 2026-03-17 |
+| 13 | pead_midcap_step43_short_core_sdlong25 | 31.7 | 1.66 | +1.5 | 57% | 2.0 | 1.3 | 58 | 7.1 | +7.1 | 72.3 | 1.46 | +1.0 | 56% | 1.9 | 0.6 | 45 | 6.2 | +6.2 | 73.7 | 2026-03-17 |
-| 14 | pead_midcap_step39_balanced_sleeves_sdlong12 | 77.9 | 1.50 | +1.5 | 55% | 3.4 | 0.5 | 62 | 2026-03-17 |
+| 14 | pead_midcap_step41_short_core_sdlong12_acshort50 | 30.7 | 1.53 | +1.2 | 56% | 1.6 | 1.3 | 59 | 6.8 | +6.8 | 72.3 | 1.48 | +0.9 | 55% | 1.8 | 0.6 | 40 | 5.6 | +5.6 | 73.7 | 2026-03-17 |
-| 15 | pead_midcap_step30_balanced_sleeves_nofrac | 75.2 | 1.41 | +1.4 | 54% | 2.8 | 0.8 | 67 | 2026-03-17 |
+| 15 | pead_midcap_step37_balanced_sleeves_aclong_vol3 | 30.6 | 1.38 | +1.2 | 54% | 2.5 | 0.8 | 63 | 7.5 | +7.5 | 76.6 | 1.38 | +1.1 | 51% | 1.8 | 0.9 | 57 | 7.9 | +7.9 | 77.2 | 2026-03-17 |
-| 16 | pead_midcap_step33_balanced_sleeves_nofrac_acshort6 | 73.9 | 1.47 | +1.2 | 52% | 2.3 | 0.7 | 48 | 2026-03-17 |
+| 16 | pead_midcap_step42_short_core_only | 30.5 | 1.44 | +0.8 | 63% | 1.2 | 1.2 | 46 | 5.5 | +5.5 | 71.7 | 2.55 | +1.2 | 58% | 2.8 | 0.5 | 31 | 4.5 | +4.5 | 71.4 | 2026-03-17 |
-| 17 | pead_midcap_step37_balanced_sleeves_aclong_vol3 | 73.6 | 1.38 | +1.2 | 54% | 2.5 | 0.8 | 63 | 2026-03-17 |
+| 17 | pead_midcap_step14_score65 | 29.9 | 1.22 | +0.7 | 57% | 1.3 | 0.9 | 72 | - | - | - | 1.40 | +1.0 | 56% | 1.6 | 0.7 | 66 | - | - | - | 2026-03-17 |
-| 18 | pead_midcap_step41_short_core_sdlong12_acshort50 | 72.6 | 1.53 | +1.2 | 56% | 1.6 | 1.3 | 59 | 2026-03-17 |
+| 18 | pead_midcap_step18_nofrac | 29.4 | 1.23 | +0.8 | 52% | 1.4 | 0.9 | 64 | - | - | - | 1.46 | +1.2 | 47% | 1.9 | 0.6 | 55 | - | - | - | 2026-03-17 |
-| 19 | pead_midcap_step42_short_core_only | 66.3 | 1.44 | +0.8 | 63% | 1.2 | 1.2 | 46 | 2026-03-17 |
+| 19 | pead_midcap_step31_balanced_sleeves_nofrac_acshort12 | 29.3 | 1.49 | +1.6 | 55% | 3.2 | 0.7 | 64 | 7.6 | +7.6 | 76.6 | 1.38 | +1.1 | 52% | 1.8 | 0.9 | 58 | 8.0 | +8.0 | 77.2 | 2026-03-17 |
-| 20 | pead_midcap_step14_score65 | 64.2 | 1.22 | +0.7 | 57% | 1.3 | 0.9 | 72 | 2026-03-17 |
+| 20 | pead_midcap_step19_hold5 | 29.2 | 1.20 | +0.7 | 57% | 1.2 | 0.9 | 72 | - | - | - | 1.55 | +1.4 | 57% | 2.2 | 0.7 | 68 | - | - | - | 2026-03-17 |
-| 21 | pead_midcap_step18_nofrac | 63.2 | 1.23 | +0.8 | 52% | 1.4 | 0.9 | 64 | 2026-03-17 |
+| 21 | pead_midcap_step49_same_day_short_macro_block | 29.1 | 2.40 | +0.4 | 70% | 1.6 | 0.4 | 10 | 1.0 | +1.0 | 26.1 | 25.09 | +1.2 | 75% | 3.2 | 0.3 | 12 | 2.6 | +2.6 | 32.1 | 2026-03-17 |
-| 22 | pead_midcap_step19_hold5 | 62.9 | 1.20 | +0.7 | 57% | 1.2 | 0.9 | 72 | 2026-03-17 |
+| 22 | pead_midcap_step20_best3 | 28.7 | 1.22 | +0.7 | 52% | 1.3 | 0.9 | 64 | - | - | - | 1.61 | +1.6 | 48% | 2.5 | 0.6 | 56 | - | - | - | 2026-03-17 |
-| 23 | pead_midcap_step20_best3 | 62.0 | 1.22 | +0.7 | 52% | 1.3 | 0.9 | 64 | 2026-03-17 |
+| 23 | pead_midcap_step40_short_core_sdlong12 | 28.6 | 1.77 | +1.5 | 58% | 2.2 | 1.1 | 55 | 6.4 | +6.4 | 72.3 | 1.48 | +0.9 | 55% | 1.8 | 0.6 | 40 | 5.6 | +5.6 | 73.7 | 2026-03-17 |
-| 24 | pead_midcap_step17_target2 | 59.4 | 1.18 | +0.6 | 53% | 1.1 | 0.8 | 66 | 2026-03-17 |
+| 24 | pead_midcap_step17_target2 | 27.3 | 1.18 | +0.6 | 53% | 1.1 | 0.8 | 66 | - | - | - | 1.38 | +1.0 | 51% | 1.5 | 0.7 | 59 | - | - | - | 2026-03-17 |
-| 25 | pead_midcap_step27_sdlong_close7_budget25 | 57.9 | 1.17 | +0.6 | 54% | 0.9 | 1.3 | 68 | 2026-03-17 |
+| 25 | pead_midcap_step39_balanced_sleeves_sdlong12 | 27.0 | 1.50 | +1.5 | 55% | 3.4 | 0.5 | 62 | 7.4 | +7.4 | 76.6 | 1.38 | +1.0 | 51% | 1.7 | 0.9 | 53 | 7.5 | +7.5 | 77.2 | 2026-03-17 |
-| 26 | pead_midcap_step13_best | 57.7 | 1.12 | +0.4 | 55% | 0.8 | 0.9 | 75 | 2026-03-16 |
+| 26 | pead_midcap_step27_sdlong_close7_budget25 | 26.4 | 1.17 | +0.6 | 54% | 0.9 | 1.3 | 68 | 8.5 | +8.5 | 72.3 | - | - | - | - | - | 0 | - | - | - | 2026-03-17 |
-| 27 | pead_midcap_step23_sdlong_close7 | 55.1 | 1.13 | +0.5 | 54% | 0.7 | 1.4 | 69 | 2026-03-17 |
+| 27 | pead_midcap_step13_best | 26.3 | 1.12 | +0.4 | 55% | 0.8 | 0.9 | 75 | - | - | - | 1.61 | +1.6 | 59% | 2.2 | 0.7 | 70 | - | - | - | 2026-03-16 |
-| 28 | pead_midcap_step38_balanced_sleeves_aclong_vol4 | 54.2 | 1.12 | +0.4 | 51% | 0.8 | 0.9 | 61 | 2026-03-17 |
+| 28 | pead_midcap_step23_sdlong_close7 | 24.9 | 1.13 | +0.5 | 54% | 0.7 | 1.4 | 69 | 8.6 | +8.6 | 72.3 | - | - | - | - | - | 0 | - | - | - | 2026-03-17 |
-| 29 | pead_midcap_step16_react7_score65 | 53.2 | 0.97 | -0.1 | 56% | -0.2 | 1.6 | 89 | 2026-03-17 |
+| 29 | pead_midcap_step34_balanced_sleeves_nofrac_aclong25 | 23.9 | 1.62 | +1.8 | 56% | 3.8 | 0.6 | 62 | 7.4 | +7.4 | 76.6 | 1.30 | +0.9 | 51% | 1.4 | 0.9 | 57 | 7.9 | +7.9 | 77.2 | 2026-03-17 |
-| 30 | pead_midcap_step5_maxcand3 | 52.5 | 0.95 | -0.2 | 54% | -0.4 | 1.4 | 96 | 2026-03-16 |
+| 30 | pead_midcap_step16_react7_score65 | 23.8 | 0.97 | -0.1 | 56% | -0.2 | 1.6 | 89 | - | - | - | 2.01 | +2.8 | 63% | 3.7 | 0.8 | 83 | - | - | - | 2026-03-17 |
-| 31 | pead_midcap_portfolio_v2 | 52.0 | 1.08 | +0.3 | 51% | 0.5 | 1.8 | 70 | 2026-03-17 |
+| 31 | pead_midcap_step5_maxcand3 | 23.4 | 0.95 | -0.2 | 54% | -0.4 | 1.4 | 96 | - | - | - | 2.08 | +3.3 | 65% | 4.0 | 1.1 | 89 | - | - | - | 2026-03-16 |
-| 32 | pead_midcap_step15_react7 | 51.7 | 0.94 | -0.3 | 55% | -0.5 | 1.6 | 91 | 2026-03-17 |
+| 32 | pead_midcap_step38_balanced_sleeves_aclong_vol4 | 23.3 | 1.12 | +0.4 | 51% | 0.8 | 0.9 | 61 | 7.4 | +7.4 | 76.6 | 1.29 | +0.8 | 53% | 1.4 | 0.9 | 53 | 7.6 | +7.6 | 77.2 | 2026-03-17 |
-| 33 | pead_midcap_step11_score60 | 50.2 | 1.02 | +0.1 | 52% | 0.2 | 0.9 | 77 | 2026-03-16 |
+| 33 | pead_midcap_step15_react7 | 23.0 | 0.94 | -0.3 | 55% | -0.5 | 1.6 | 91 | - | - | - | 1.89 | +2.6 | 62% | 3.5 | 0.9 | 84 | - | - | - | 2026-03-17 |
-| 34 | pead_midcap_step12_vol2x | 50.2 | 1.02 | +0.1 | 52% | 0.1 | 0.9 | 77 | 2026-03-16 |
+| 34 | pead_midcap_portfolio_v2 | 23.0 | 1.08 | +0.3 | 51% | 0.5 | 1.8 | 70 | 7.9 | +7.9 | 72.3 | - | - | - | - | - | 0 | - | - | - | 2026-03-17 |
-| 35 | pead_midcap_step3_10pct | 49.8 | 1.00 | +0.0 | 50% | 0.0 | 1.4 | 98 | 2026-03-16 |
+| 35 | pead_midcap_step11_score60 | 22.1 | 1.02 | +0.1 | 52% | 0.2 | 0.9 | 77 | - | - | - | 1.91 | +2.3 | 62% | 3.0 | 0.7 | 72 | - | - | - | 2026-03-16 |
-| 36 | pead_midcap_step10_short | 49.2 | 1.00 | -0.0 | 52% | -0.0 | 0.9 | 79 | 2026-03-16 |
+| 36 | pead_midcap_step12_vol2x | 22.1 | 1.02 | +0.1 | 52% | 0.1 | 0.9 | 77 | - | - | - | 1.91 | +2.3 | 62% | 3.0 | 0.7 | 72 | - | - | - | 2026-03-16 |
-| 37 | pead_midcap_step53_short_core_macro_block_crashcap_gap14 | 43.1 | 4.60 | +1.1 | 84% | 3.5 | 0.2 | 19 | 2026-03-17 |
+| 37 | pead_midcap_step3_10pct | 21.9 | 1.00 | +0.0 | 50% | 0.0 | 1.4 | 98 | - | - | - | 1.66 | +2.0 | 60% | 2.9 | 0.7 | 78 | - | - | - | 2026-03-16 |
-| 38 | pead_midcap_step50_same_day_short_long_macro_block | 41.9 | 3.21 | +0.8 | 60% | 2.7 | 0.4 | 15 | 2026-03-17 |
+| 38 | pead_midcap_step10_short | 21.5 | 1.00 | -0.0 | 52% | -0.0 | 0.9 | 79 | - | - | - | 1.61 | +1.6 | 59% | 2.2 | 0.7 | 70 | - | - | - | 2026-03-16 |
-| 39 | pead_midcap_step2_notrail | 41.4 | 0.93 | -0.3 | 67% | -0.4 | 1.7 | 54 | 2026-03-16 |
+| 39 | pead_midcap_step35_balanced_sleeves_nofrac_aclong12 | 21.2 | 2.01 | +2.2 | 61% | 4.2 | 0.4 | 56 | 6.2 | +1.1 | 76.6 | 1.24 | +0.7 | 51% | 1.2 | 0.9 | 53 | 7.5 | +1.6 | 77.2 | 2026-03-17 |
-| 40 | pead_midcap_step1_fixedr | 39.7 | 0.91 | -0.5 | 47% | -0.7 | 1.5 | 95 | 2026-03-16 |
+| 40 | pead_midcap_step2_notrail | 17.2 | 0.93 | -0.3 | 67% | -0.4 | 1.7 | 54 | - | - | - | 1.07 | +0.4 | 73% | 0.4 | 1.7 | 62 | - | - | - | 2026-03-16 |
-| 41 | pead_midcap_combo_10pct_maxcand3 | 39.2 | 0.97 | -0.1 | 51% | -0.2 | 0.9 | 79 | 2026-03-16 |
+| 41 | pead_midcap_step1_fixedr | 16.2 | 0.91 | -0.5 | 47% | -0.7 | 1.5 | 95 | - | - | - | 1.77 | +2.8 | 49% | 3.4 | 1.6 | 69 | - | - | - | 2026-03-16 |
-| 42 | pead_midcap_step49_same_day_short_macro_block | 38.9 | 2.40 | +0.4 | 70% | 1.6 | 0.4 | 10 | 2026-03-17 |
+| 42 | pead_midcap_combo_10pct_maxcand3 | 15.9 | 0.97 | -0.1 | 51% | -0.2 | 0.9 | 79 | - | - | - | 1.91 | +2.3 | 62% | 3.0 | 0.7 | 72 | - | - | - | 2026-03-16 |
-| 43 | pead_midcap_step6_drift | 37.0 | 0.86 | -0.9 | 43% | -1.6 | 1.7 | 100 | 2026-03-16 |
+| 43 | pead_midcap_step6_drift | 14.7 | 0.86 | -0.9 | 43% | -1.6 | 1.7 | 100 | - | - | - | 1.70 | +2.7 | 51% | 3.7 | 1.0 | 75 | - | - | - | 2026-03-16 |
-| 44 | pead_midcap_step7_fixedr | 35.4 | 0.94 | -0.2 | 45% | -0.4 | 1.0 | 71 | 2026-03-16 |
+| 44 | pead_midcap_step7_fixedr | 13.8 | 0.94 | -0.2 | 45% | -0.4 | 1.0 | 71 | - | - | - | 2.00 | +2.4 | 53% | 3.1 | 0.7 | 53 | - | - | - | 2026-03-16 |
-| 45 | pead_midcap_step4_longonly | 32.6 | 0.84 | -0.8 | 45% | -1.2 | 1.4 | 78 | 2026-03-16 |
+| 45 | pead_midcap_step4_longonly | 12.2 | 0.84 | -0.8 | 45% | -1.2 | 1.4 | 78 | - | - | - | 1.43 | +1.5 | 61% | 1.7 | 1.6 | 76 | - | - | - | 2026-03-16 |
-| 46 | pead_midcap_step8_nft | 31.3 | 0.65 | -1.8 | 41% | -3.0 | 2.2 | 71 | 2026-03-16 |
+| 46 | pead_midcap_step8_nft | 11.5 | 0.65 | -1.8 | 41% | -3.0 | 2.2 | 71 | - | - | - | 1.55 | +1.8 | 46% | 2.4 | 1.0 | 57 | - | - | - | 2026-03-16 |
-| 47 | pead_midcap_step9_stop2 | 31.1 | 0.77 | -1.7 | 45% | -1.8 | 2.5 | 71 | 2026-03-16 |
+| 47 | pead_midcap_step9_stop2 | 11.4 | 0.77 | -1.7 | 45% | -1.8 | 2.5 | 71 | - | - | - | 1.81 | +3.2 | 53% | 2.8 | 1.2 | 53 | - | - | - | 2026-03-16 |
 ## Recent Entries
 ### IMP-0047 (2026-03-17) — pead_midcap_step53_short_core_macro_block_crashcap_gap14
 Hypothesis: A stricter same-day long gap filter might further concentrate the overlay into only the strongest continuation setups.
-Verdict: **WORSE** (SQS 43.1)
+Verdict: **WORSE** (SQS 39.5)
 Reasoning: The stricter gap filter over-concentrated the overlay, dropped total trade count below a healthy level, and cratered test SQS.
 Next: Use moderate overlay filters only; the strict version is too sparse.
 ### IMP-0046 (2026-03-17) — pead_midcap_step52_short_core_macro_block_crashcap_gap10
 Hypothesis: The same-day long overlay may work better when restricted to larger reaction-day gap moves.
-Verdict: **NEUTRAL** (SQS 86.3)
+Verdict: **NEUTRAL** (SQS 49.8)
 Reasoning: A 10% gap filter made train and valid much stronger but gave back some test performance, so this is a balanced alternative rather than a clear new leader.
 Next: If optimizing for robustness across splits, keep exploring overlay quality gates around this variant.
 ### IMP-0045 (2026-03-17) — pead_midcap_step51_short_core_macro_block_crashcap
 Hypothesis: Extreme one-day crash continuations are too stretched for the same-day short sleeve and should be excluded.
-Verdict: **BETTER** (SQS 86.7)
+Verdict: **BETTER** (SQS 44.4)
 Reasoning: Capping same-day shorts at -45% reaction preserved train and valid while modestly improving test return, PF, drawdown, and Sharpe versus step45.
 Next: Combine the crash cap with a quality filter on the same-day long overlay.
 ### IMP-0044 (2026-03-17) — pead_midcap_step50_same_day_short_long_macro_block
 Hypothesis: The same-day long overlay may matter, but the after-close short sleeve may be removable.
-Verdict: **WORSE** (SQS 41.9)
+Verdict: **WORSE** (SQS 36.2)
 Reasoning: Dropping the after-close short sleeve reduced both valid and test performance, so step45 still benefits from carrying all three active sleeves.
 Next: Refine sleeve quality rather than deleting sleeves wholesale.
 ### IMP-0043 (2026-03-17) — pead_midcap_step49_same_day_short_macro_block
 Hypothesis: The pure same-day short engine might dominate the portfolio and make other sleeves unnecessary.
-Verdict: **WORSE** (SQS 38.9)
+Verdict: **WORSE** (SQS 29.1)
 Reasoning: The single-sleeve version collapsed SQS because trade count and robustness fell too far, even though the kept trades were profitable.
 Next: Keep the supporting sleeves and test smaller structural adjustments instead.
--- a/journal/experiment_registry.json
+++ b/journal/experiment_registry.json
--- a/journal/improvement_journal.jsonl
+++ b/journal/improvement_journal.jsonl
@ -26,7 +26,7 @@
 {"entry_id":"IMP-0026","timestamp":"2026-03-17T06:59:09.255218+00:00","experiment_name":"pead_midcap_step31_balanced_sleeves_nofrac_acshort12","hypothesis":"Shrink the after-close short sleeve again while keeping the rest of the balanced step30 structure intact.","config_delta":{"base_experiment":"pead_midcap_step30_balanced_sleeves_nofrac","changes":{}},"results":{"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317065814_ca861015","trade_count":64,"profit_factor":1.4929664814770336,"total_return_pct":1.5563811867822805,"win_rate":0.546875,"max_drawdown_pct":0.6813257982524192,"sharpe_ratio":3.235108878683093,"monthly_win_rate":1.0,"equity_curve_r_squared":0.8144638886979153},"train":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317065750_ca861015","trade_count":426,"profit_factor":1.0908713237524914,"total_return_pct":2.45066812206927,"win_rate":0.45539906103286387,"max_drawdown_pct":2.6845497224397294,"sharpe_ratio":0.3357364835430949,"monthly_win_rate":0.5142857142857142,"equity_curve_r_squared":0.6011932815228401},"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317065810_ca861015","trade_count":58,"profit_factor":1.37572962559925,"total_return_pct":1.122807593766629,"win_rate":0.5172413793103449,"max_drawdown_pct":0.8884863869932943,"sharpe_ratio":1.8162493466149443,"monthly_win_rate":0.75,"equity_curve_r_squared":0.3292893146712864}},"sqs_score":77.9,"sqs_breakdown":{"profitability":60.9,"risk":100.0,"consistency":82.8,"robustness":80.0},"verdict":"better","verdict_reasoning":"New best. Train/valid/test all held up, with test SQS 77.9, return +1.56%, PF 1.66, Sharpe 2.39, and max drawdown 0.32%. This beat both step30 and step14 by a wide margin.","next_direction":"Use step31 as the new base and only explore very local refinements around sleeve weights or execution if further gains are needed.","tags":["pead","midcap","step31","balanced","sleeves","nofrac","acshort12"]}
 {"entry_id":"IMP-0027","timestamp":"2026-03-17T06:59:09.622681+00:00","experiment_name":"pead_midcap_step33_balanced_sleeves_nofrac_acshort6","hypothesis":"Cut the after-close short sleeve even further to see if the balanced portfolio still benefits from the bucket at very low size.","config_delta":{"base_experiment":"pead_midcap_step31_balanced_sleeves_nofrac_acshort12","changes":{}},"results":{"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317065700_765096af","trade_count":48,"profit_factor":1.4726487821009215,"total_return_pct":1.2304910166146,"win_rate":0.5208333333333334,"max_drawdown_pct":0.7081594843000599,"sharpe_ratio":2.3026797742321596,"monthly_win_rate":0.6666666666666666,"equity_curve_r_squared":0.7956775968800425}},"sqs_score":73.9,"sqs_breakdown":{"profitability":58.6,"risk":100.0,"consistency":74.3,"robustness":70.8},"verdict":"worse","verdict_reasoning":"Test-only run stayed strong but slipped versus step31: SQS 73.9 vs 77.9 and return +1.23% vs +1.56%. The smaller sleeve gave up too much trade-count support.","next_direction":"Keep the 12.5% after-close short sleeve from step31.","tags":["pead","midcap","step33","balanced","sleeves","nofrac","acshort6"]}
 {"entry_id":"IMP-0028","timestamp":"2026-03-17T07:05:41.969503+00:00","experiment_name":"pead_midcap_step34_balanced_sleeves_nofrac_aclong25","hypothesis":"Cap the after-close long sleeve to two trades per day so only the best-ranked overnight continuation names survive.","config_delta":{"base_experiment":"pead_midcap_step31_balanced_sleeves_nofrac_acshort12","changes":{}},"results":{"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317070326_6bef3bf5","trade_count":62,"profit_factor":1.6229787581193362,"total_return_pct":1.809973740790665,"win_rate":0.5645161290322581,"max_drawdown_pct":0.5514994557958487,"sharpe_ratio":3.784062042365747,"monthly_win_rate":1.0,"equity_curve_r_squared":0.8527995261422007},"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317070447_6bef3bf5","trade_count":57,"profit_factor":1.2977962107392036,"total_return_pct":0.888003662238436,"win_rate":0.5087719298245614,"max_drawdown_pct":0.884279658549619,"sharpe_ratio":1.43172550959332,"monthly_win_rate":0.75,"equity_curve_r_squared":0.26161939040716076}},"sqs_score":81.3,"sqs_breakdown":{"profitability":68.4,"risk":100.0,"consistency":85.8,"robustness":78.9},"verdict":"worse","verdict_reasoning":"Test improved to SQS 81.3 and +1.81% return, but valid slipped to SQS 63.7 and +0.89% versus step31 valid SQS 68.2 and +1.12%. The top-two cap on after-close longs was not robust across splits.","next_direction":"Keep step31 as the robust base. If after-close long needs filtering, prefer a quality gate rather than a blunt daily-cap reduction.","tags":["pead","midcap","step34","balanced","sleeves","nofrac","aclong25"]}
-{"entry_id":"IMP-0029","timestamp":"2026-03-17T07:05:47.363036+00:00","experiment_name":"pead_midcap_step35_balanced_sleeves_nofrac_aclong12","hypothesis":"Cap the after-close long sleeve to one trade per day so the portfolio fully leans into the strongest overnight continuation name only.","config_delta":{"base_experiment":"pead_midcap_step31_balanced_sleeves_nofrac_acshort12","changes":{}},"results":{"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317070326_01aa6342","trade_count":56,"profit_factor":2.009368125746884,"total_return_pct":2.181331690538893,"win_rate":0.6071428571428571,"max_drawdown_pct":0.4424438123900949,"sharpe_ratio":4.216651506745797,"monthly_win_rate":1.0,"equity_curve_r_squared":0.8888944531705611},"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317070357_01aa6342","trade_count":53,"profit_factor":1.2430238868359724,"total_return_pct":0.7082451550234983,"win_rate":0.5094339622641509,"max_drawdown_pct":0.8845604127450726,"sharpe_ratio":1.1796633240683954,"monthly_win_rate":0.75,"equity_curve_r_squared":0.14646025414461133},"train":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317070358_01aa6342","trade_count":393,"profit_factor":1.1135008933147288,"total_return_pct":2.755962139586409,"win_rate":0.46055979643765904,"max_drawdown_pct":2.418955602136075,"sharpe_ratio":0.38778263039611355,"monthly_win_rate":0.4857142857142857,"equity_curve_r_squared":0.6345243468091162}},"sqs_score":90.4,"sqs_breakdown":{"profitability":88.7,"risk":100.0,"consistency":92.9,"robustness":75.6},"verdict":"worse","verdict_reasoning":"This became a new test-only high water mark at SQS 90.4 and +2.18%, but valid deteriorated to SQS 59.8 and +0.71%, materially below step31. The tighter top-one cap overfit to the test window.","next_direction":"Avoid promoting step35. Explore engine-specific quality filters for after-close longs instead of hard caps that reshuffle trade selection too aggressively.","tags":["pead","midcap","step35","balanced","sleeves","nofrac","aclong12"]}
+{"entry_id":"IMP-0029","timestamp":"2026-03-17T07:05:47.363036+00:00","experiment_name":"pead_midcap_step35_balanced_sleeves_nofrac_aclong12","hypothesis":"Cap the after-close long sleeve to one trade per day so the portfolio fully leans into the strongest overnight continuation name only.","config_delta":{"base_experiment":"pead_midcap_step31_balanced_sleeves_nofrac_acshort12","changes":{}},"results":{"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317081806944455_48768969","trade_count":56,"profit_factor":2.009368125746884,"total_return_pct":2.181331690538893,"win_rate":0.6071428571428571,"max_drawdown_pct":0.4424438123900949,"sharpe_ratio":4.216651506745797,"monthly_win_rate":1.0,"equity_curve_r_squared":0.8888944531705611,"avg_gross_exposure_pct":6.227596335660352,"avg_net_exposure_pct":1.1262657538133194,"days_in_market_pct":76.59574468085107},"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317081817695254_48768969","trade_count":53,"profit_factor":1.2430238868359724,"total_return_pct":0.7082451550234983,"win_rate":0.5094339622641509,"max_drawdown_pct":0.8845604127450726,"sharpe_ratio":1.1796633240683954,"monthly_win_rate":0.75,"equity_curve_r_squared":0.14646025414461133,"avg_gross_exposure_pct":7.5108896612986165,"avg_net_exposure_pct":1.5673421201773916,"days_in_market_pct":77.19298245614034},"train":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317070358_01aa6342","trade_count":393,"profit_factor":1.1135008933147288,"total_return_pct":2.755962139586409,"win_rate":0.46055979643765904,"max_drawdown_pct":2.418955602136075,"sharpe_ratio":0.38778263039611355,"monthly_win_rate":0.4857142857142857,"equity_curve_r_squared":0.6345243468091162}},"sqs_score":90.4,"sqs_breakdown":{"profitability":88.7,"risk":100.0,"consistency":92.9,"robustness":75.6},"verdict":"worse","verdict_reasoning":"This became a new test-only high water mark at SQS 90.4 and +2.18%, but valid deteriorated to SQS 59.8 and +0.71%, materially below step31. The tighter top-one cap overfit to the test window.","next_direction":"Avoid promoting step35. Explore engine-specific quality filters for after-close longs instead of hard caps that reshuffle trade selection too aggressively.","tags":["pead","midcap","step35","balanced","sleeves","nofrac","aclong12"],"sqs_v2_score":90.1,"sqs_v2_breakdown":{"profitability":88.7,"risk":100.0,"consistency":92.9,"robustness":75.6,"capital_efficiency":79.0}}
 {"entry_id":"IMP-0030","timestamp":"2026-03-17T07:05:52.980782+00:00","experiment_name":"pead_midcap_step36_balanced_sleeves_nofrac_aclong12_sdlong12","hypothesis":"Keep only one same-day long close trade per day alongside the top-one after-close long sleeve, concentrating the portfolio into the single best long continuation setups.","config_delta":{"base_experiment":"pead_midcap_step35_balanced_sleeves_nofrac_aclong12","changes":{}},"results":{"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317070326_18804e7c","trade_count":54,"profit_factor":2.061648751658133,"total_return_pct":2.1478646359501146,"win_rate":0.6111111111111112,"max_drawdown_pct":0.3579823467213594,"sharpe_ratio":4.4957307480895325,"monthly_win_rate":1.0,"equity_curve_r_squared":0.9122676472502861}},"sqs_score":90.3,"sqs_breakdown":{"profitability":88.6,"risk":100.0,"consistency":93.5,"robustness":74.4},"verdict":"worse","verdict_reasoning":"Test stayed extremely strong at SQS 90.3 and +2.15%, but it did not exceed step35 on test and reduced trade support further. Without valid/train confirmation, it is not a better promotion candidate than step31.","next_direction":"Keep the same-day long sleeve at two trades per day if using this family, and focus next on smarter after-close long quality filtering.","tags":["pead","midcap","step36","balanced","sleeves","nofrac","aclong12","sdlong12"]}
 {"entry_id":"IMP-0031","timestamp":"2026-03-17T07:11:20.119143+00:00","experiment_name":"pead_midcap_step37_balanced_sleeves_aclong_vol3","hypothesis":"Require stronger volume confirmation for after-close long signals only, while leaving the rest of the step31 sleeve mix unchanged.","config_delta":{"base_experiment":"pead_midcap_step31_balanced_sleeves_nofrac_acshort12","changes":{}},"results":{"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317071110746395_1ae4b33a","trade_count":63,"profit_factor":1.377310183342628,"total_return_pct":1.1862161229211343,"win_rate":0.5396825396825397,"max_drawdown_pct":0.8243191281564817,"sharpe_ratio":2.4626680484273273,"monthly_win_rate":1.0,"equity_curve_r_squared":0.6885324256339993},"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317071110780542_1ae4b33a","trade_count":57,"profit_factor":1.375336580561267,"total_return_pct":1.1216330418461293,"win_rate":0.5087719298245614,"max_drawdown_pct":0.8896384382974281,"sharpe_ratio":1.7701058914354275,"monthly_win_rate":0.75,"equity_curve_r_squared":0.3302488231788083}},"sqs_score":73.6,"sqs_breakdown":{"profitability":53.6,"risk":100.0,"consistency":81.6,"robustness":72.5},"verdict":"worse","verdict_reasoning":"Engine-specific volume gating on after-close longs did not help. Test fell to SQS 73.6 and valid to 67.6, both below the step31 base. The extra volume filter removed too much breadth without improving robustness.","next_direction":"Do not tighten after-close long volume gates further. Keep after-close long breadth and search elsewhere if more robustness is needed.","tags":["pead","midcap","step37","balanced","sleeves","aclong","vol3"]}
 {"entry_id":"IMP-0032","timestamp":"2026-03-17T07:11:26.965107+00:00","experiment_name":"pead_midcap_step38_balanced_sleeves_aclong_vol4","hypothesis":"Push the after-close long sleeve to an even stricter volume gate so only the highest-conviction overnight reactions remain.","config_delta":{"base_experiment":"pead_midcap_step37_balanced_sleeves_aclong_vol3","changes":{}},"results":{"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317071110780988_d9020d34","trade_count":61,"profit_factor":1.1204890397898135,"total_return_pct":0.388000418802214,"win_rate":0.5081967213114754,"max_drawdown_pct":0.9399793760032171,"sharpe_ratio":0.8136845994700913,"monthly_win_rate":0.6666666666666666,"equity_curve_r_squared":0.043822307069783864},"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317071110695904_d9020d34","trade_count":53,"profit_factor":1.2930052787721988,"total_return_pct":0.8333477612284769,"win_rate":0.5283018867924528,"max_drawdown_pct":0.8783421675410786,"sharpe_ratio":1.4376847128011023,"monthly_win_rate":0.75,"equity_curve_r_squared":0.19294412573010883}},"sqs_score":54.2,"sqs_breakdown":{"profitability":37.6,"risk":80.2,"consistency":72.2,"robustness":31.1},"verdict":"worse","verdict_reasoning":"The stricter after-close long filter clearly broke the portfolio. Test dropped to SQS 54.2 and valid to 63.2, confirming that this sleeve cannot be improved by simply tightening volume thresholds.","next_direction":"Abandon the after-close long volume-threshold path. If that sleeve is revisited, it needs a different filter than raw PEAD volume.","tags":["pead","midcap","step38","balanced","sleeves","aclong","vol4"]}
@ -35,13 +35,13 @@
 {"entry_id":"IMP-0035","timestamp":"2026-03-17T07:20:29.978594+00:00","experiment_name":"pead_midcap_step41_short_core_sdlong12_acshort50","hypothesis":"Lean harder into the after-close short sleeve inside the new short-core portfolio.","config_delta":{"base_experiment":"pead_midcap_step40_short_core_sdlong12","changes":{}},"results":{"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317071756232509_4c9fd4ce","trade_count":40,"profit_factor":1.483920817012022,"total_return_pct":0.8662763872782817,"win_rate":0.55,"max_drawdown_pct":0.6476993822793542,"sharpe_ratio":1.8482432920991685,"monthly_win_rate":0.75,"equity_curve_r_squared":0.19435203451054603},"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317071756491852_4c9fd4ce","trade_count":59,"profit_factor":1.5287187084320328,"total_return_pct":1.231827071365813,"win_rate":0.559322033898305,"max_drawdown_pct":1.2759660172387226,"sharpe_ratio":1.6310365384096348,"monthly_win_rate":1.0,"equity_curve_r_squared":0.4227182249282274}},"sqs_score":72.6,"sqs_breakdown":{"profitability":61.4,"risk":92.3,"consistency":84.9,"robustness":53.6},"verdict":"worse","verdict_reasoning":"Increasing after-close short capacity weakened the portfolio: test dropped from SQS 82.1 to 72.6 and valid stayed flat at 68.4. The short core benefits from the bucket, but not at this larger size.","next_direction":"Keep the after-close short sleeve at 25% inside the short-core family.","tags":["pead","midcap","step41","short","core","sdlong12","acshort50"]}
 {"entry_id":"IMP-0037","timestamp":"2026-03-17T07:20:29.978670+00:00","experiment_name":"pead_midcap_step43_short_core_sdlong25","hypothesis":"Restore a larger same-day long close sleeve after removing after-close longs, to see if breadth improves the short-core portfolio.","config_delta":{"base_experiment":"pead_midcap_step40_short_core_sdlong12","changes":{}},"results":{"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317071957079395_9046f612","trade_count":45,"profit_factor":1.4604102163883825,"total_return_pct":0.9859009158709378,"win_rate":0.5555555555555556,"max_drawdown_pct":0.6469365320312458,"sharpe_ratio":1.9269062309154923,"monthly_win_rate":0.75,"equity_curve_r_squared":0.2568675902061873},"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317071957079892_9046f612","trade_count":58,"profit_factor":1.6628306811204845,"total_return_pct":1.4636686220428092,"win_rate":0.5689655172413793,"max_drawdown_pct":1.2630043081731899,"sharpe_ratio":2.0112049109753753,"monthly_win_rate":1.0,"equity_curve_r_squared":0.57291038217223}},"sqs_score":78.9,"sqs_breakdown":{"profitability":69.0,"risk":98.5,"consistency":86.5,"robustness":62.5},"verdict":"worse","verdict_reasoning":"Restoring more same-day long breadth weakened both splits versus step40: valid moved from SQS 68.4 to 69.7 but test fell from 82.1 to 78.9 and profitability dropped. The smaller 12.5% sleeve remains the better balance.","next_direction":"Keep the same-day long overlay small inside step40.","tags":["pead","midcap","step43","short","core","sdlong25"]}
 {"entry_id":"IMP-0034","timestamp":"2026-03-17T07:20:29.978842+00:00","experiment_name":"pead_midcap_step40_short_core_sdlong12","hypothesis":"Drop the unstable after-close long sleeve and reallocate the portfolio to same-day shorts, after-close shorts, and a small same-day close-entry long overlay.","config_delta":{"base_experiment":"pead_midcap_step31_balanced_sleeves_nofrac_acshort12","changes":{}},"results":{"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317071756453145_9eb08c8d","trade_count":40,"profit_factor":1.483920817012022,"total_return_pct":0.8662763872782817,"win_rate":0.55,"max_drawdown_pct":0.6476993822793542,"sharpe_ratio":1.8482432920991685,"monthly_win_rate":0.75,"equity_curve_r_squared":0.19435203451054603},"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317071756505329_9eb08c8d","trade_count":55,"profit_factor":1.7695290624299977,"total_return_pct":1.52020169384827,"win_rate":0.5818181818181818,"max_drawdown_pct":1.128135882565315,"sharpe_ratio":2.2087400904317267,"monthly_win_rate":1.0,"equity_curve_r_squared":0.632820774620957},"train":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317071814970965_9eb08c8d","trade_count":306,"profit_factor":1.2169887676968005,"total_return_pct":3.708684730821959,"win_rate":0.4869281045751634,"max_drawdown_pct":2.5507043144769854,"sharpe_ratio":0.5720232080317771,"monthly_win_rate":0.625,"equity_curve_r_squared":0.39497202444939644}},"sqs_score":82.1,"sqs_breakdown":{"profitability":74.6,"risk":99.3,"consistency":88.6,"robustness":64.6},"verdict":"better","verdict_reasoning":"New robust leader. Train improved from SQS 55.5 to 63.1 and return +2.45% to +3.71%. Valid edged up from SQS 68.2 to 68.4 with lower drawdown, and test improved from SQS 77.9 to 82.1 with PF 1.77. Removing after-close longs fixed the biggest unstable sleeve without giving up the same-day long upside.","next_direction":"Use step40 as the new base. Only explore local refinements around the short-core structure if needed.","tags":["pead","midcap","step40","short","core","sdlong12"]}
-{"entry_id":"IMP-0038","timestamp":"2026-03-17T07:54:25.486815+00:00","experiment_name":"pead_midcap_step44_short_core_macro50","hypothesis":"Scaling entries down in weak macro regimes will keep the short-core structure while cutting drawdowns.","config_delta":{"base_experiment":"pead_midcap_step40_short_core_sdlong12","changes":{}},"results":{"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317073536976331_9c3139ba","trade_count":40,"profit_factor":1.7786229517773076,"total_return_pct":1.0842382260887244,"win_rate":0.55,"max_drawdown_pct":0.4065245518278289,"sharpe_ratio":2.4891782517624543,"monthly_win_rate":0.75,"equity_curve_r_squared":0.34564834918762316},"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317073537015426_9c3139ba","trade_count":58,"profit_factor":2.0052763103391356,"total_return_pct":1.2986802302195721,"win_rate":0.5689655172413793,"max_drawdown_pct":0.7045335236824514,"sharpe_ratio":2.6217356101467053,"monthly_win_rate":1.0,"equity_curve_r_squared":0.7991432493330419},"train":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317073555963588_9c3139ba","trade_count":307,"profit_factor":1.2319887960741855,"total_return_pct":3.4232792009141852,"win_rate":0.4820846905537459,"max_drawdown_pct":1.97377752972113,"sharpe_ratio":0.6203457661608955,"monthly_win_rate":0.625,"equity_curve_r_squared":0.4148680874417292}},"sqs_score":87.9,"sqs_breakdown":{"profitability":85.2,"risk":100.0,"consistency":86.5,"robustness":76.6},"verdict":"better","verdict_reasoning":"Half-size macro scaling materially improved valid and test risk-adjusted performance versus step40, confirming that SPY-below-SMA exposure was a real drag.","next_direction":"Try a full macro block to see whether removing weak-regime entries entirely is even cleaner.","tags":["pead","midcap","step44","short","core","macro50"]}
+{"entry_id":"IMP-0038","timestamp":"2026-03-17T07:54:25.486815+00:00","experiment_name":"pead_midcap_step44_short_core_macro50","hypothesis":"Scaling entries down in weak macro regimes will keep the short-core structure while cutting drawdowns.","config_delta":{"base_experiment":"pead_midcap_step40_short_core_sdlong12","changes":{}},"results":{"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317081817694940_42f8ea0e","trade_count":40,"profit_factor":1.7786229517773076,"total_return_pct":1.0842382260887244,"win_rate":0.55,"max_drawdown_pct":0.4065245518278289,"sharpe_ratio":2.4891782517624543,"monthly_win_rate":0.75,"equity_curve_r_squared":0.34564834918762316,"avg_gross_exposure_pct":4.823763728554508,"avg_net_exposure_pct":-1.952624774981557,"days_in_market_pct":73.68421052631578},"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317081807057141_42f8ea0e","trade_count":58,"profit_factor":2.0052763103391356,"total_return_pct":1.2986802302195721,"win_rate":0.5689655172413793,"max_drawdown_pct":0.7045335236824514,"sharpe_ratio":2.6217356101467053,"monthly_win_rate":1.0,"equity_curve_r_squared":0.7991432493330419,"avg_gross_exposure_pct":4.453478837236966,"avg_net_exposure_pct":-1.5519032472252936,"days_in_market_pct":72.3404255319149},"train":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317073555963588_9c3139ba","trade_count":307,"profit_factor":1.2319887960741855,"total_return_pct":3.4232792009141852,"win_rate":0.4820846905537459,"max_drawdown_pct":1.97377752972113,"sharpe_ratio":0.6203457661608955,"monthly_win_rate":0.625,"equity_curve_r_squared":0.4148680874417292}},"sqs_score":87.9,"sqs_breakdown":{"profitability":85.2,"risk":100.0,"consistency":86.5,"robustness":76.6},"verdict":"better","verdict_reasoning":"Half-size macro scaling materially improved valid and test risk-adjusted performance versus step40, confirming that SPY-below-SMA exposure was a real drag.","next_direction":"Try a full macro block to see whether removing weak-regime entries entirely is even cleaner.","tags":["pead","midcap","step44","short","core","macro50"],"sqs_v2_score":86.9,"sqs_v2_breakdown":{"profitability":85.2,"risk":100.0,"consistency":86.5,"robustness":76.6,"capital_efficiency":70.8}}
 {"entry_id":"IMP-0039","timestamp":"2026-03-17T07:54:25.866482+00:00","experiment_name":"pead_midcap_step45_short_core_macro_block","hypothesis":"If weak-regime entries are mostly noise, hard-blocking them should outperform simple size scaling.","config_delta":{"base_experiment":"pead_midcap_step44_short_core_macro50","changes":{}},"results":{"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317073536967478_c3a3615c","trade_count":28,"profit_factor":2.3084684930008983,"total_return_pct":1.30245295115927,"win_rate":0.6785714285714286,"max_drawdown_pct":0.37472483014430374,"sharpe_ratio":3.0487382658615467,"monthly_win_rate":0.75,"equity_curve_r_squared":0.48762905493056535},"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317073537015511_c3a3615c","trade_count":23,"profit_factor":3.7827916861128097,"total_return_pct":1.2021184808416436,"win_rate":0.6956521739130435,"max_drawdown_pct":0.2608177180219861,"sharpe_ratio":3.3795660622862123,"monthly_win_rate":0.6666666666666666,"equity_curve_r_squared":0.9110928059216074},"train":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317073556022527_c3a3615c","trade_count":220,"profit_factor":1.2595187982631553,"total_return_pct":3.200633818982489,"win_rate":0.5045454545454545,"max_drawdown_pct":1.389607178327899,"sharpe_ratio":0.6336391727317187,"monthly_win_rate":0.6551724137931034,"equity_curve_r_squared":0.39958317262208615}},"sqs_score":86.7,"sqs_breakdown":{"profitability":84.8,"risk":100.0,"consistency":95.8,"robustness":57.2},"verdict":"better","verdict_reasoning":"The hard macro block improved valid and test again, with sharper PF and much lower drawdown than the 50% scaler version.","next_direction":"Stress the sleeve mix around the new macro-blocked core.","tags":["pead","midcap","step45","short","core","macro","block"]}
 {"entry_id":"IMP-0040","timestamp":"2026-03-17T07:54:26.245073+00:00","experiment_name":"pead_midcap_step46_short_core_macro_block_acshort12","hypothesis":"The after-close short sleeve may be oversized after the macro block and could improve if reduced.","config_delta":{"base_experiment":"pead_midcap_step45_short_core_macro_block","changes":{}},"results":{"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317073936915792_e3adb886","trade_count":26,"profit_factor":2.4445635979938296,"total_return_pct":1.339314216731771,"win_rate":0.6923076923076923,"max_drawdown_pct":0.37458871743417604,"sharpe_ratio":3.1583041871985076,"monthly_win_rate":0.75,"equity_curve_r_squared":0.48286315791941375},"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317073939814431_e3adb886","trade_count":21,"profit_factor":4.522043594902001,"total_return_pct":1.2518263219734362,"win_rate":0.7142857142857143,"max_drawdown_pct":0.2783619920052574,"sharpe_ratio":3.6389410132883055,"monthly_win_rate":0.6666666666666666,"equity_curve_r_squared":0.9244229299732519}},"sqs_score":86.6,"sqs_breakdown":{"profitability":85.0,"risk":100.0,"consistency":95.8,"robustness":56.1},"verdict":"worse","verdict_reasoning":"Shrinking the after-close short sleeve slightly degraded both valid and test, so the step45 25% sleeve was not the problem.","next_direction":"Test whether the same-day long sleeve or the short-only core is the real source of edge.","tags":["pead","midcap","step46","short","core","macro","block","acshort12"]}
 {"entry_id":"IMP-0041","timestamp":"2026-03-17T07:54:26.612610+00:00","experiment_name":"pead_midcap_step47_short_core_macro_block_sdlong25","hypothesis":"A larger same-day reaction-close long overlay might scale once the macro block removes bad regimes.","config_delta":{"base_experiment":"pead_midcap_step45_short_core_macro_block","changes":{}},"results":{"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317073936910739_a38d69e4","trade_count":31,"profit_factor":1.7661839600751525,"total_return_pct":1.03177071332802,"win_rate":0.6451612903225806,"max_drawdown_pct":0.42599257212123065,"sharpe_ratio":2.218739891960781,"monthly_win_rate":0.75,"equity_curve_r_squared":0.4214549671036421},"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317073939814560_a38d69e4","trade_count":25,"profit_factor":3.429404855633068,"total_return_pct":1.2833562667310616,"win_rate":0.68,"max_drawdown_pct":0.2608177180219861,"sharpe_ratio":3.32348066596022,"monthly_win_rate":0.6666666666666666,"equity_curve_r_squared":0.889835319415119}},"sqs_score":87.0,"sqs_breakdown":{"profitability":85.1,"risk":100.0,"consistency":95.8,"robustness":58.3},"verdict":"worse","verdict_reasoning":"Increasing the same-day long sleeve hurt valid materially and did not produce a cleaner overall profile than step45.","next_direction":"Try sleeve removal experiments instead of scaling overlays up.","tags":["pead","midcap","step47","short","core","macro","block","sdlong25"]}
 {"entry_id":"IMP-0042","timestamp":"2026-03-17T07:54:26.967747+00:00","experiment_name":"pead_midcap_step48_short_core_macro_block_nolong","hypothesis":"The macro-blocked short core may be strong enough without the same-day long overlay.","config_delta":{"base_experiment":"pead_midcap_step45_short_core_macro_block","changes":{}},"results":{"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317074234188388_d9b28f40","trade_count":20,"profit_factor":5.726207009868067,"total_return_pct":1.2584927752282966,"win_rate":0.7,"max_drawdown_pct":0.29982153086364316,"sharpe_ratio":3.3529664036078217,"monthly_win_rate":1.0,"equity_curve_r_squared":0.4654938970631658},"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317074234267904_d9b28f40","trade_count":20,"profit_factor":3.402559985815088,"total_return_pct":0.8451446297187069,"win_rate":0.8,"max_drawdown_pct":0.4239007555767844,"sharpe_ratio":2.4413021645869604,"monthly_win_rate":0.6666666666666666,"equity_curve_r_squared":0.7861833255279326}},"sqs_score":85.7,"sqs_breakdown":{"profitability":83.4,"risk":100.0,"consistency":95.8,"robustness":54.7},"verdict":"worse","verdict_reasoning":"Removing the long sleeve weakened both valid and test relative to step45, so the small same-day long overlay still adds useful diversification.","next_direction":"Test whether the same-day short sleeve can stand alone or whether the after-close short sleeve is also required.","tags":["pead","midcap","step48","short","core","macro","block","nolong"]}
 {"entry_id":"IMP-0043","timestamp":"2026-03-17T07:54:27.322093+00:00","experiment_name":"pead_midcap_step49_same_day_short_macro_block","hypothesis":"The pure same-day short engine might dominate the portfolio and make other sleeves unnecessary.","config_delta":{"base_experiment":"pead_midcap_step45_short_core_macro_block","changes":{}},"results":{"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317074233919071_b5cae2ca","trade_count":12,"profit_factor":25.094599474518425,"total_return_pct":1.1648266582814248,"win_rate":0.75,"max_drawdown_pct":0.33436592553948014,"sharpe_ratio":3.227317987298049,"monthly_win_rate":1.0,"equity_curve_r_squared":0.3679821567954865},"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317074234357252_b5cae2ca","trade_count":10,"profit_factor":2.4025420210691006,"total_return_pct":0.40093684044296973,"win_rate":0.7,"max_drawdown_pct":0.3693400619520335,"sharpe_ratio":1.5580313035787354,"monthly_win_rate":0.6666666666666666,"equity_curve_r_squared":0.3032092765977471}},"sqs_score":38.9,"sqs_breakdown":{"profitability":81.6,"risk":92.6,"consistency":95.8,"robustness":19.0},"verdict":"worse","verdict_reasoning":"The single-sleeve version collapsed SQS because trade count and robustness fell too far, even though the kept trades were profitable.","next_direction":"Keep the supporting sleeves and test smaller structural adjustments instead.","tags":["pead","midcap","step49","same","day","short","macro","block"]}
 {"entry_id":"IMP-0044","timestamp":"2026-03-17T07:54:27.691772+00:00","experiment_name":"pead_midcap_step50_same_day_short_long_macro_block","hypothesis":"The same-day long overlay may matter, but the after-close short sleeve may be removable.","config_delta":{"base_experiment":"pead_midcap_step45_short_core_macro_block","changes":{}},"results":{"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317074234172662_a01fa9b8","trade_count":20,"profit_factor":2.2051399980357798,"total_return_pct":0.9530265960178512,"win_rate":0.65,"max_drawdown_pct":0.3444323377542605,"sharpe_ratio":2.687864694846263,"monthly_win_rate":0.75,"equity_curve_r_squared":0.41775889730721794},"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317074234263930_a01fa9b8","trade_count":15,"profit_factor":3.2060142950020296,"total_return_pct":0.805535692316771,"win_rate":0.6,"max_drawdown_pct":0.42315703199236054,"sharpe_ratio":2.6871013818111624,"monthly_win_rate":0.6666666666666666,"equity_curve_r_squared":0.8242465440616173}},"sqs_score":41.9,"sqs_breakdown":{"profitability":83.2,"risk":100.0,"consistency":87.5,"robustness":52.8},"verdict":"worse","verdict_reasoning":"Dropping the after-close short sleeve reduced both valid and test performance, so step45 still benefits from carrying all three active sleeves.","next_direction":"Refine sleeve quality rather than deleting sleeves wholesale.","tags":["pead","midcap","step50","same","day","short","long","macro","block"]}
-{"entry_id":"IMP-0045","timestamp":"2026-03-17T07:54:28.043618+00:00","experiment_name":"pead_midcap_step51_short_core_macro_block_crashcap","hypothesis":"Extreme one-day crash continuations are too stretched for the same-day short sleeve and should be excluded.","config_delta":{"base_experiment":"pead_midcap_step45_short_core_macro_block","changes":{}},"results":{"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317074801780500_1585063f","trade_count":28,"profit_factor":2.3084684930008983,"total_return_pct":1.30245295115927,"win_rate":0.6785714285714286,"max_drawdown_pct":0.37472483014430374,"sharpe_ratio":3.0487382658615467,"monthly_win_rate":0.75,"equity_curve_r_squared":0.48762905493056535},"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317074801781828_1585063f","trade_count":22,"profit_factor":4.190184861108528,"total_return_pct":1.2441182731003355,"win_rate":0.7272727272727273,"max_drawdown_pct":0.22380328257556925,"sharpe_ratio":3.5956143566120704,"monthly_win_rate":0.6666666666666666,"equity_curve_r_squared":0.918016321908289},"train":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317074932001383_1585063f","trade_count":220,"profit_factor":1.2595187982631553,"total_return_pct":3.200633818982489,"win_rate":0.5045454545454545,"max_drawdown_pct":1.389607178327899,"sharpe_ratio":0.6336391727317187,"monthly_win_rate":0.6551724137931034,"equity_curve_r_squared":0.39958317262208615}},"sqs_score":86.7,"sqs_breakdown":{"profitability":85.0,"risk":100.0,"consistency":95.8,"robustness":56.7},"verdict":"better","verdict_reasoning":"Capping same-day shorts at -45% reaction preserved train and valid while modestly improving test return, PF, drawdown, and Sharpe versus step45.","next_direction":"Combine the crash cap with a quality filter on the same-day long overlay.","tags":["pead","midcap","step51","short","core","macro","block","crashcap"]}
+{"entry_id":"IMP-0045","timestamp":"2026-03-17T07:54:28.043618+00:00","experiment_name":"pead_midcap_step51_short_core_macro_block_crashcap","hypothesis":"Extreme one-day crash continuations are too stretched for the same-day short sleeve and should be excluded.","config_delta":{"base_experiment":"pead_midcap_step45_short_core_macro_block","changes":{}},"results":{"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317081833666987_5039805d","trade_count":28,"profit_factor":2.3084684930008983,"total_return_pct":1.30245295115927,"win_rate":0.6785714285714286,"max_drawdown_pct":0.37472483014430374,"sharpe_ratio":3.0487382658615467,"monthly_win_rate":0.75,"equity_curve_r_squared":0.48762905493056535,"avg_gross_exposure_pct":4.09489843643577,"avg_net_exposure_pct":-1.9273432556417505,"days_in_market_pct":57.89473684210527},"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317081834927778_5039805d","trade_count":22,"profit_factor":4.190184861108528,"total_return_pct":1.2441182731003355,"win_rate":0.7272727272727273,"max_drawdown_pct":0.22380328257556925,"sharpe_ratio":3.5956143566120704,"monthly_win_rate":0.6666666666666666,"equity_curve_r_squared":0.918016321908289,"avg_gross_exposure_pct":2.2928950819181733,"avg_net_exposure_pct":-0.8189811228558067,"days_in_market_pct":42.5531914893617},"train":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317074932001383_1585063f","trade_count":220,"profit_factor":1.2595187982631553,"total_return_pct":3.200633818982489,"win_rate":0.5045454545454545,"max_drawdown_pct":1.389607178327899,"sharpe_ratio":0.6336391727317187,"monthly_win_rate":0.6551724137931034,"equity_curve_r_squared":0.39958317262208615}},"sqs_score":86.7,"sqs_breakdown":{"profitability":85.0,"risk":100.0,"consistency":95.8,"robustness":56.7},"verdict":"better","verdict_reasoning":"Capping same-day shorts at -45% reaction preserved train and valid while modestly improving test return, PF, drawdown, and Sharpe versus step45.","next_direction":"Combine the crash cap with a quality filter on the same-day long overlay.","tags":["pead","midcap","step51","short","core","macro","block","crashcap"],"sqs_v2_score":89.6,"sqs_v2_breakdown":{"profitability":85.0,"risk":100.0,"consistency":95.8,"robustness":56.7,"capital_efficiency":100.0}}
-{"entry_id":"IMP-0046","timestamp":"2026-03-17T07:54:28.401055+00:00","experiment_name":"pead_midcap_step52_short_core_macro_block_crashcap_gap10","hypothesis":"The same-day long overlay may work better when restricted to larger reaction-day gap moves.","config_delta":{"base_experiment":"pead_midcap_step51_short_core_macro_block_crashcap","changes":{}},"results":{"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317075123185892_7592b1dc","trade_count":25,"profit_factor":5.660464878695176,"total_return_pct":1.818984312375629,"win_rate":0.72,"max_drawdown_pct":0.21695852738272095,"sharpe_ratio":4.755652871139783,"monthly_win_rate":1.0,"equity_curve_r_squared":0.6608034205824956},"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317075123226004_7592b1dc","trade_count":22,"profit_factor":3.351800408128271,"total_return_pct":1.0111883417758072,"win_rate":0.7727272727272727,"max_drawdown_pct":0.24257912061402945,"sharpe_ratio":3.031867507475822,"monthly_win_rate":0.6666666666666666,"equity_curve_r_squared":0.8468990955221466},"train":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317075225071209_7592b1dc","trade_count":201,"profit_factor":1.3153549627707726,"total_return_pct":3.2988485556152156,"win_rate":0.4975124378109453,"max_drawdown_pct":1.5599163185209133,"sharpe_ratio":0.683194016870928,"monthly_win_rate":0.6923076923076923,"equity_curve_r_squared":0.7186047701175057}},"sqs_score":86.3,"sqs_breakdown":{"profitability":84.0,"risk":100.0,"consistency":95.8,"robustness":56.7},"verdict":"neutral","verdict_reasoning":"A 10% gap filter made train and valid much stronger but gave back some test performance, so this is a balanced alternative rather than a clear new leader.","next_direction":"If optimizing for robustness across splits, keep exploring overlay quality gates around this variant.","tags":["pead","midcap","step52","short","core","macro","block","crashcap","gap10"]}
+{"entry_id":"IMP-0046","timestamp":"2026-03-17T07:54:28.401055+00:00","experiment_name":"pead_midcap_step52_short_core_macro_block_crashcap_gap10","hypothesis":"The same-day long overlay may work better when restricted to larger reaction-day gap moves.","config_delta":{"base_experiment":"pead_midcap_step51_short_core_macro_block_crashcap","changes":{}},"results":{"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317081834927839_7592b1dc","trade_count":25,"profit_factor":5.660464878695176,"total_return_pct":1.818984312375629,"win_rate":0.72,"max_drawdown_pct":0.21695852738272095,"sharpe_ratio":4.755652871139783,"monthly_win_rate":1.0,"equity_curve_r_squared":0.6608034205824956,"avg_gross_exposure_pct":3.5838805580978437,"avg_net_exposure_pct":-2.593156236925373,"days_in_market_pct":56.14035087719298},"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317081833658080_7592b1dc","trade_count":22,"profit_factor":3.351800408128271,"total_return_pct":1.0111883417758072,"win_rate":0.7727272727272727,"max_drawdown_pct":0.24257912061402945,"sharpe_ratio":3.031867507475822,"monthly_win_rate":0.6666666666666666,"equity_curve_r_squared":0.8468990955221466,"avg_gross_exposure_pct":2.858350680305373,"avg_net_exposure_pct":-2.222595187645864,"days_in_market_pct":53.191489361702125},"train":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317075225071209_7592b1dc","trade_count":201,"profit_factor":1.3153549627707726,"total_return_pct":3.2988485556152156,"win_rate":0.4975124378109453,"max_drawdown_pct":1.5599163185209133,"sharpe_ratio":0.683194016870928,"monthly_win_rate":0.6923076923076923,"equity_curve_r_squared":0.7186047701175057}},"sqs_score":86.3,"sqs_breakdown":{"profitability":84.0,"risk":100.0,"consistency":95.8,"robustness":56.7},"verdict":"neutral","verdict_reasoning":"A 10% gap filter made train and valid much stronger but gave back some test performance, so this is a balanced alternative rather than a clear new leader.","next_direction":"If optimizing for robustness across splits, keep exploring overlay quality gates around this variant.","tags":["pead","midcap","step52","short","core","macro","block","crashcap","gap10"],"sqs_v2_score":87.2,"sqs_v2_breakdown":{"profitability":84.0,"risk":100.0,"consistency":95.8,"robustness":56.7,"capital_efficiency":79.5}}
 {"entry_id":"IMP-0047","timestamp":"2026-03-17T07:54:28.762757+00:00","experiment_name":"pead_midcap_step53_short_core_macro_block_crashcap_gap14","hypothesis":"A stricter same-day long gap filter might further concentrate the overlay into only the strongest continuation setups.","config_delta":{"base_experiment":"pead_midcap_step52_short_core_macro_block_crashcap_gap10","changes":{}},"results":{"valid":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317075123208169_66d09993","trade_count":25,"profit_factor":8.207760543782888,"total_return_pct":2.0180349316014032,"win_rate":0.76,"max_drawdown_pct":0.2179161262122437,"sharpe_ratio":5.288977955005757,"monthly_win_rate":1.0,"equity_curve_r_squared":0.6385560108285373},"test":{"run_id":"bt_baseline_swing_v1_midcap-filte_20260317075123225576_66d09993","trade_count":19,"profit_factor":4.599773676047558,"total_return_pct":1.1191525844285641,"win_rate":0.8421052631578947,"max_drawdown_pct":0.21947031826165894,"sharpe_ratio":3.463700507402551,"monthly_win_rate":0.6666666666666666,"equity_curve_r_squared":0.8852420626052896}},"sqs_score":43.1,"sqs_breakdown":{"profitability":84.5,"risk":100.0,"consistency":95.8,"robustness":55.0},"verdict":"worse","verdict_reasoning":"The stricter gap filter over-concentrated the overlay, dropped total trade count below a healthy level, and cratered test SQS.","next_direction":"Use moderate overlay filters only; the strict version is too sparse.","tags":["pead","midcap","step53","short","core","macro","block","crashcap","gap14"]}
--- a/libs/backtest/domain.py
+++ b/libs/backtest/domain.py
@ -165,6 +165,9 @@ class MetricsBundle(BaseModel):
    sortino_ratio: float | None = None
    avg_daily_pnl: float | None = None
    avg_positions_held: float | None = None
    avg_gross_exposure_pct: float | None = None
    avg_net_exposure_pct: float | None = None
    days_in_market_pct: float | None = None
    # Stability metrics (4)
    trade_skewness: float | None = None
@ -364,6 +367,40 @@ class SQSWeights(BaseModel):
    low_trade_penalty_factor: float = 0.5
 class SQSv2Weights(BaseModel):
    """Weights for Strategy Quality Score v2 with capital efficiency."""
    profitability: float = 0.35
    risk: float = 0.25
    consistency: float = 0.20
    robustness: float = 0.10
    capital_efficiency: float = 0.10
    low_trade_penalty_threshold: int = 20
    low_trade_penalty_factor: float = 0.5
 class PromotionScoreWeights(BaseModel):
    """Weights for promotion scoring across valid/test splits."""
    valid_quality: float = 0.55
    test_quality: float = 0.15
    floor_quality: float = 0.30
 class UnifiedScoreWeights(BaseModel):
    """Weights for a stricter single ranking score across valid/test splits."""
    split_profitability: float = 0.25
    split_risk: float = 0.20
    split_consistency: float = 0.15
    split_robustness: float = 0.20
    split_capital_efficiency: float = 0.20
    valid_quality: float = 0.45
    test_quality: float = 0.20
    floor_quality: float = 0.20
    gap_quality: float = 0.15
 class SplitResult(BaseModel):
    """Metrics for a single backtest split (train/valid/test)."""
@ -376,6 +413,9 @@ class SplitResult(BaseModel):
    sharpe_ratio: float | None = None
    monthly_win_rate: float | None = None
    equity_curve_r_squared: float | None = None
    avg_gross_exposure_pct: float | None = None
    avg_net_exposure_pct: float | None = None
    days_in_market_pct: float | None = None
 class ConfigDelta(BaseModel):
@ -396,6 +436,12 @@ class JournalEntry(BaseModel):
    results: dict[str, SplitResult] = Field(default_factory=dict)  # split_name → SplitResult
    sqs_score: float | None = None
    sqs_breakdown: dict[str, float] = Field(default_factory=dict)
    sqs_v2_score: float | None = None
    sqs_v2_breakdown: dict[str, float] = Field(default_factory=dict)
    promotion_score: float | None = None
    promotion_breakdown: dict[str, float] = Field(default_factory=dict)
    unified_score: float | None = None
    unified_breakdown: dict[str, float] = Field(default_factory=dict)
    verdict: str = "unknown"  # better / worse / neutral / unknown
    verdict_reasoning: str = ""
    next_direction: str = ""
@ -408,12 +454,29 @@ class RegistryEntry(BaseModel):
    entry_id: str
    experiment_name: str
    sqs_score: float
    sqs_v2_score: float | None = None
    promotion_score: float | None = None
    unified_score: float | None = None
    # test split metrics
    profit_factor: float | None = None
    total_return_pct: float | None = None
    win_rate: float | None = None
    sharpe_ratio: float | None = None
    max_drawdown_pct: float | None = None
    trade_count: int = 0
    avg_gross_exposure_pct: float | None = None
    avg_net_exposure_pct: float | None = None
    days_in_market_pct: float | None = None
    # valid split metrics
    valid_profit_factor: float | None = None
    valid_total_return_pct: float | None = None
    valid_win_rate: float | None = None
    valid_sharpe_ratio: float | None = None
    valid_max_drawdown_pct: float | None = None
    valid_trade_count: int = 0
    valid_avg_gross_exposure_pct: float | None = None
    valid_avg_net_exposure_pct: float | None = None
    valid_days_in_market_pct: float | None = None
    timestamp: str = ""
--- a/libs/backtest/metrics.py
+++ b/libs/backtest/metrics.py
@ -180,6 +180,45 @@ def compute_avg_positions_held(equity_curve: list[DailyPortfolioState]) -> float
    return statistics.mean(len(s.open_positions) for s in equity_curve)
 def compute_avg_gross_exposure_pct(
    equity_curve: list[DailyPortfolioState],
 ) -> float | None:
    if not equity_curve:
        return None
    exposure_pcts = [
        (state.gross_exposure / state.equity) * 100.0
        for state in equity_curve
        if state.equity > 0
    ]
    if not exposure_pcts:
        return None
    return statistics.mean(exposure_pcts)
 def compute_avg_net_exposure_pct(
    equity_curve: list[DailyPortfolioState],
 ) -> float | None:
    if not equity_curve:
        return None
    exposure_pcts = [
        (state.net_exposure / state.equity) * 100.0
        for state in equity_curve
        if state.equity > 0
    ]
    if not exposure_pcts:
        return None
    return statistics.mean(exposure_pcts)
 def compute_days_in_market_pct(
    equity_curve: list[DailyPortfolioState],
 ) -> float | None:
    if not equity_curve:
        return None
    days_in_market = sum(1 for state in equity_curve if state.gross_exposure > 0)
    return days_in_market / len(equity_curve) * 100.0
 # ---------------------------------------------------------------------------
 # Stability metrics (4)
 # ---------------------------------------------------------------------------
@ -409,6 +448,9 @@ def build_metrics_bundle(
        sortino_ratio=compute_sortino_ratio(equity_curve),
        avg_daily_pnl=compute_avg_daily_pnl(equity_curve),
        avg_positions_held=compute_avg_positions_held(equity_curve),
        avg_gross_exposure_pct=compute_avg_gross_exposure_pct(equity_curve),
        avg_net_exposure_pct=compute_avg_net_exposure_pct(equity_curve),
        days_in_market_pct=compute_days_in_market_pct(equity_curve),
        # Stability
        trade_skewness=compute_trade_skewness(trades),
        trade_kurtosis=compute_trade_kurtosis(trades),
--- a/libs/backtest/tracker.py
+++ b/libs/backtest/tracker.py
@ -1,6 +1,7 @@
 """Strategy improvement tracker: SQS computation, journal I/O, leaderboard."""
 from __future__ import annotations
 import functools
 import json
 from pathlib import Path
 from typing import Any
@ -10,9 +11,12 @@ from libs.backtest.domain import (
    ExperimentRegistry,
    JournalEntry,
    MetricsBundle,
    PromotionScoreWeights,
    RegistryEntry,
    SplitResult,
    SQSWeights,
    SQSv2Weights,
    UnifiedScoreWeights,
 )
 from libs.common.logging import get_logger
 from libs.common.time_utils import utc_now
@ -20,6 +24,9 @@ from libs.common.time_utils import utc_now
 logger = get_logger(__name__)
 _DEFAULT_WEIGHTS = SQSWeights()
 _DEFAULT_V2_WEIGHTS = SQSv2Weights()
 _DEFAULT_PROMOTION_WEIGHTS = PromotionScoreWeights()
 _DEFAULT_UNIFIED_WEIGHTS = UnifiedScoreWeights()
 # ---------------------------------------------------------------------------
@ -49,6 +56,39 @@ def _normalize_inverse(value: float | None, low: float, high: float) -> float:
    return max(0.0, min(100.0, score))
 def _normalize_band(
    value: float | None,
    low_bad: float,
    low_good: float,
    high_good: float,
    high_bad: float,
 ) -> float:
    """Score 0-100 with an optimal middle band."""
    if value is None:
        return 0.0
    if value <= low_bad or value >= high_bad:
        return 0.0
    if low_good <= value <= high_good:
        return 100.0
    if value < low_good:
        return (value - low_bad) / (low_good - low_bad) * 100.0
    return (high_bad - value) / (high_bad - high_good) * 100.0
 def _calibrate_sqs(value: float | None) -> float | None:
    """Compress raw integrated scores into a harsher absolute-looking range."""
    if value is None:
        return None
    return round(max(0.0, value * 0.70 - 7.5), 1)
 def _apply_single_split_penalty(value: float | None) -> float | None:
    """Discount scores that have no valid/test confirmation pair."""
    if value is None:
        return None
    return round(value * 0.80, 1)
 def compute_sqs(
    metrics: MetricsBundle,
    weights: SQSWeights | None = None,
@ -101,6 +141,67 @@ def compute_sqs(
    return sqs, breakdown
 def compute_sqs_v2(
    metrics: MetricsBundle,
    weights: SQSv2Weights | None = None,
 ) -> tuple[float | None, dict[str, float]]:
    """Compute SQS v2, adding a capital-efficiency sleeve to the score."""
    if metrics.avg_gross_exposure_pct is None or metrics.days_in_market_pct is None:
        return None, {}
    w = weights or _DEFAULT_V2_WEIGHTS
    pf_score = _normalize(metrics.profit_factor, low=0.8, high=2.0)
    ret_score = _normalize(metrics.total_return_pct, low=-5.0, high=5.0)
    profitability = pf_score * 0.6 + ret_score * 0.4
    dd_score = _normalize_inverse(metrics.max_drawdown_pct, low=10.0, high=1.0)
    sharpe_score = _normalize(metrics.sharpe_ratio, low=-1.0, high=2.0)
    risk = dd_score * 0.5 + sharpe_score * 0.5
    wr_score = _normalize(metrics.win_rate, low=0.35, high=0.65)
    mwr_score = _normalize(metrics.monthly_win_rate, low=0.30, high=0.70)
    consistency = wr_score * 0.5 + mwr_score * 0.5
    r2_score = _normalize(metrics.equity_curve_r_squared, low=0.0, high=0.80)
    tc_score = _normalize(float(metrics.trade_count), low=10.0, high=100.0)
    robustness = r2_score * 0.5 + tc_score * 0.5
    return_on_gross_exposure = None
    if metrics.total_return_pct is not None and metrics.avg_gross_exposure_pct > 0:
        return_on_gross_exposure = metrics.total_return_pct / metrics.avg_gross_exposure_pct
    roe_score = _normalize(return_on_gross_exposure, low=0.0, high=0.50)
    dim_score = _normalize_band(
        metrics.days_in_market_pct,
        low_bad=10.0,
        low_good=40.0,
        high_good=80.0,
        high_bad=100.0,
    )
    capital_efficiency = roe_score * 0.7 + dim_score * 0.3
    sqs_v2 = (
        profitability * w.profitability
        + risk * w.risk
        + consistency * w.consistency
        + robustness * w.robustness
        + capital_efficiency * w.capital_efficiency
    )
    if metrics.trade_count < w.low_trade_penalty_threshold:
        sqs_v2 *= w.low_trade_penalty_factor
    sqs_v2 = round(sqs_v2, 1)
    breakdown = {
        "profitability": round(profitability, 1),
        "risk": round(risk, 1),
        "consistency": round(consistency, 1),
        "robustness": round(robustness, 1),
        "capital_efficiency": round(capital_efficiency, 1),
    }
    return sqs_v2, breakdown
 # ---------------------------------------------------------------------------
 # Helper builders
 # ---------------------------------------------------------------------------
@ -118,9 +219,248 @@ def build_split_result(split_name: str, run_id: str, metrics: MetricsBundle) ->
        sharpe_ratio=metrics.sharpe_ratio,
        monthly_win_rate=metrics.monthly_win_rate,
        equity_curve_r_squared=metrics.equity_curve_r_squared,
        avg_gross_exposure_pct=metrics.avg_gross_exposure_pct,
        avg_net_exposure_pct=metrics.avg_net_exposure_pct,
        days_in_market_pct=metrics.days_in_market_pct,
    )
 def _metrics_from_split_result(result: SplitResult | None) -> MetricsBundle | None:
    if result is None:
        return None
    payload = result.model_dump(exclude={"run_id"})
    return MetricsBundle.model_validate(payload)
@functools.lru_cache(maxsize=512)
 def _load_run_metrics_summary(run_id: str) -> dict[str, Any] | None:
    metrics_path = Path("runs") / run_id / "metrics" / "metrics_summary.json"
    if not metrics_path.exists():
        return None
    summary = json.loads(metrics_path.read_text())
    needed_fields = (
        "avg_gross_exposure_pct",
        "avg_net_exposure_pct",
        "days_in_market_pct",
    )
    if all(summary.get(field) is not None for field in needed_fields):
        return summary
    equity_curve_path = Path("runs") / run_id / "artifacts" / "daily_equity_curve.parquet"
    if not equity_curve_path.exists():
        return summary
    try:
        import pyarrow.parquet as pq
        rows = pq.read_table(
            equity_curve_path,
            columns=["equity", "gross_exposure", "net_exposure"],
        ).to_pylist()
    except Exception:
        return summary
    exposure_rows = [row for row in rows if float(row["equity"]) > 0]
    if not exposure_rows:
        return summary
    if summary.get("avg_gross_exposure_pct") is None:
        summary["avg_gross_exposure_pct"] = sum(
            float(row["gross_exposure"]) / float(row["equity"]) * 100.0
            for row in exposure_rows
        ) / len(exposure_rows)
    if summary.get("avg_net_exposure_pct") is None:
        summary["avg_net_exposure_pct"] = sum(
            float(row["net_exposure"]) / float(row["equity"]) * 100.0
            for row in exposure_rows
        ) / len(exposure_rows)
    if summary.get("days_in_market_pct") is None:
        summary["days_in_market_pct"] = (
            sum(1 for row in rows if float(row["gross_exposure"]) > 0) / len(rows) * 100.0
        )
    return summary
 def _hydrate_split_result(result: SplitResult | None) -> SplitResult | None:
    if result is None:
        return None
    needed_fields = (
        "avg_gross_exposure_pct",
        "avg_net_exposure_pct",
        "days_in_market_pct",
    )
    if all(getattr(result, field) is not None for field in needed_fields):
        return result
    summary = _load_run_metrics_summary(result.run_id)
    if summary is None:
        return result
    updates = {
        field: summary.get(field)
        for field in needed_fields
        if getattr(result, field) is None and summary.get(field) is not None
    }
    if not updates:
        return result
    return result.model_copy(update=updates)
 def _compute_split_quality_score(
    result: SplitResult | None,
 ) -> tuple[float | None, dict[str, float], str | None]:
    metrics = _metrics_from_split_result(_hydrate_split_result(result))
    if metrics is None:
        return None, {}, None
    sqs_v2_score, sqs_v2_breakdown = compute_sqs_v2(metrics)
    if sqs_v2_score is not None:
        return sqs_v2_score, sqs_v2_breakdown, "sqs_v2"
    sqs_score, sqs_breakdown = compute_sqs(metrics)
    return sqs_score, sqs_breakdown, "sqs"
 def compute_unified_split_quality(
    metrics: MetricsBundle,
    weights: UnifiedScoreWeights | None = None,
 ) -> tuple[float | None, dict[str, float]]:
    """Compute a stricter split-level quality score."""
    if metrics.avg_gross_exposure_pct is None or metrics.days_in_market_pct is None:
        return None, {}
    w = weights or _DEFAULT_UNIFIED_WEIGHTS
    pf_score = _normalize(metrics.profit_factor, low=1.0, high=3.0)
    ret_score = _normalize(metrics.total_return_pct, low=0.0, high=3.0)
    profitability = pf_score * 0.5 + ret_score * 0.5
    dd_score = _normalize_inverse(metrics.max_drawdown_pct, low=5.0, high=0.5)
    sharpe_score = _normalize(metrics.sharpe_ratio, low=0.0, high=3.5)
    risk = dd_score * 0.5 + sharpe_score * 0.5
    wr_score = _normalize(metrics.win_rate, low=0.50, high=0.65)
    mwr_score = _normalize(metrics.monthly_win_rate, low=0.45, high=0.75)
    consistency = wr_score * 0.5 + mwr_score * 0.5
    r2_score = _normalize(metrics.equity_curve_r_squared, low=0.10, high=0.90)
    tc_score = _normalize(float(metrics.trade_count), low=20.0, high=80.0)
    robustness = r2_score * 0.5 + tc_score * 0.5
    return_on_gross_exposure = None
    if metrics.total_return_pct is not None and metrics.avg_gross_exposure_pct > 0:
        return_on_gross_exposure = metrics.total_return_pct / metrics.avg_gross_exposure_pct
    roe_score = _normalize(return_on_gross_exposure, low=0.05, high=0.40)
    dim_score = _normalize_band(
        metrics.days_in_market_pct,
        low_bad=20.0,
        low_good=45.0,
        high_good=75.0,
        high_bad=90.0,
    )
    capital_efficiency = roe_score * 0.7 + dim_score * 0.3
    split_quality = (
        profitability * w.split_profitability
        + risk * w.split_risk
        + consistency * w.split_consistency
        + robustness * w.split_robustness
        + capital_efficiency * w.split_capital_efficiency
    )
    if metrics.trade_count < 20:
        split_quality *= 0.75
    split_quality = round(split_quality, 1)
    breakdown = {
        "profitability": round(profitability, 1),
        "risk": round(risk, 1),
        "consistency": round(consistency, 1),
        "robustness": round(robustness, 1),
        "capital_efficiency": round(capital_efficiency, 1),
    }
    return split_quality, breakdown
 def compute_promotion_score(
    test_result: SplitResult | None,
    valid_result: SplitResult | None,
    weights: PromotionScoreWeights | None = None,
 ) -> tuple[float | None, dict[str, float]]:
    """Compute a promotion score using valid/test quality and a floor term."""
    test_score, _, _ = _compute_split_quality_score(test_result)
    valid_score, _, _ = _compute_split_quality_score(valid_result)
    if test_score is None or valid_score is None:
        return None, {}
    w = weights or _DEFAULT_PROMOTION_WEIGHTS
    floor_score = min(test_score, valid_score)
    promotion_score = (
        valid_score * w.valid_quality
        + test_score * w.test_quality
        + floor_score * w.floor_quality
    )
    breakdown = {
        "valid_quality": round(valid_score, 1),
        "test_quality": round(test_score, 1),
        "floor_quality": round(floor_score, 1),
    }
    return round(promotion_score, 1), breakdown
 def compute_unified_score(
    test_result: SplitResult | None,
    valid_result: SplitResult | None,
    weights: UnifiedScoreWeights | None = None,
 ) -> tuple[float | None, dict[str, float]]:
    """Compute one integrated score for real promotion decisions."""
    test_result = _hydrate_split_result(test_result)
    valid_result = _hydrate_split_result(valid_result)
    test_metrics = _metrics_from_split_result(test_result)
    valid_metrics = _metrics_from_split_result(valid_result)
    if test_metrics is None or valid_metrics is None:
        return None, {}
    w = weights or _DEFAULT_UNIFIED_WEIGHTS
    test_quality, _ = compute_unified_split_quality(test_metrics, w)
    valid_quality, _ = compute_unified_split_quality(valid_metrics, w)
    if test_quality is None or valid_quality is None:
        return None, {}
    floor_quality = min(test_quality, valid_quality)
    gap_quality = _normalize_inverse(abs(valid_quality - test_quality), low=35.0, high=5.0)
    unified_score = (
        valid_quality * w.valid_quality
        + test_quality * w.test_quality
        + floor_quality * w.floor_quality
        + gap_quality * w.gap_quality
    )
    breakdown = {
        "valid_quality": round(valid_quality, 1),
        "test_quality": round(test_quality, 1),
        "floor_quality": round(floor_quality, 1),
        "gap_quality": round(gap_quality, 1),
    }
    return _calibrate_sqs(unified_score), breakdown
 def compute_public_sqs(
    test_result: SplitResult | None,
    valid_result: SplitResult | None,
 ) -> tuple[float | None, dict[str, float], str | None]:
    """Return the public-facing SQS.
    Prefer the stricter integrated score when both valid/test are available.
    Fall back to the best available single-split quality score otherwise,
    using the same harsher calibration band.
    """
    integrated_score, integrated_breakdown = compute_unified_score(test_result, valid_result)
    if integrated_score is not None:
        return integrated_score, integrated_breakdown, "integrated"
    split_score, split_breakdown, split_source = _compute_split_quality_score(test_result)
    if split_score is not None:
        return _apply_single_split_penalty(_calibrate_sqs(split_score)), split_breakdown, split_source
    return None, {}, None
 def compute_config_delta(
    current: dict[str, Any],
    baseline: dict[str, Any],
@ -195,24 +535,58 @@ def rebuild_registry(
    registry_entries: list[RegistryEntry] = []
    for je in entries:
-        test_result = je.results.get("test")
+        test_result = _hydrate_split_result(je.results.get("test"))
        valid_result = _hydrate_split_result(je.results.get("valid"))
        computed_sqs_v2 = None
        test_metrics = _metrics_from_split_result(test_result)
        if test_metrics is not None:
            computed_sqs_v2, _ = compute_sqs_v2(test_metrics)
        computed_promotion_score, _ = compute_promotion_score(test_result, valid_result)
        computed_unified_score, _ = compute_unified_score(test_result, valid_result)
        computed_public_sqs, _, _ = compute_public_sqs(test_result, valid_result)
        canonical_sqs = computed_public_sqs if computed_public_sqs is not None else je.sqs_score or 0.0
        registry_entries.append(
            RegistryEntry(
                entry_id=je.entry_id,
                experiment_name=je.experiment_name,
-                sqs_score=je.sqs_score or 0.0,
+                sqs_score=canonical_sqs,
                sqs_v2_score=je.sqs_v2_score if je.sqs_v2_score is not None else computed_sqs_v2,
                promotion_score=(
                    je.promotion_score if je.promotion_score is not None else computed_promotion_score
                ),
                unified_score=(
                    je.unified_score if je.unified_score is not None else computed_unified_score
                ),
                profit_factor=test_result.profit_factor if test_result else None,
                total_return_pct=test_result.total_return_pct if test_result else None,
                win_rate=test_result.win_rate if test_result else None,
                sharpe_ratio=test_result.sharpe_ratio if test_result else None,
                max_drawdown_pct=test_result.max_drawdown_pct if test_result else None,
                trade_count=test_result.trade_count if test_result else 0,
                avg_gross_exposure_pct=test_result.avg_gross_exposure_pct if test_result else None,
                avg_net_exposure_pct=test_result.avg_net_exposure_pct if test_result else None,
                days_in_market_pct=test_result.days_in_market_pct if test_result else None,
                valid_profit_factor=valid_result.profit_factor if valid_result else None,
                valid_total_return_pct=valid_result.total_return_pct if valid_result else None,
                valid_win_rate=valid_result.win_rate if valid_result else None,
                valid_sharpe_ratio=valid_result.sharpe_ratio if valid_result else None,
                valid_max_drawdown_pct=valid_result.max_drawdown_pct if valid_result else None,
                valid_trade_count=valid_result.trade_count if valid_result else 0,
                valid_avg_gross_exposure_pct=valid_result.avg_gross_exposure_pct if valid_result else None,
                valid_avg_net_exposure_pct=valid_result.avg_net_exposure_pct if valid_result else None,
                valid_days_in_market_pct=valid_result.days_in_market_pct if valid_result else None,
                timestamp=je.timestamp,
            )
        )
-    # Sort by SQS descending
+    # Sort by canonical SQS descending, then promotion.
-    registry_entries.sort(key=lambda e: e.sqs_score, reverse=True)
+    registry_entries.sort(
        key=lambda e: (
            -(e.sqs_score or 0.0),
            e.promotion_score is None,
            -(e.promotion_score or 0.0),
        )
    )
    registry = ExperimentRegistry(
        entries=registry_entries,
@ -238,8 +612,8 @@ def _write_leaderboard_md(
    lines: list[str] = []
    lines.append("# Strategy Improvement Leaderboard")
    lines.append(f"_Updated: {registry.updated_at}_\n")
-    lines.append("| # | Experiment | SQS | PF | Ret% | WR | Sharpe | DD% | Trades | Date |")
+    lines.append("| # | Experiment | SQS | [T]PF | [T]Ret% | [T]WR | [T]Sharpe | [T]DD% | [T]N | [T]Gross% | [T]Net% | [T]DIM% | [V]PF | [V]Ret% | [V]WR | [V]Sharpe | [V]DD% | [V]N | [V]Gross% | [V]Net% | [V]DIM% | Date |")
-    lines.append("|---|-----------|-----|-----|------|-----|--------|-----|--------|------|")
+    lines.append("|---|-----------|-----|-------|---------|-------|-----------|--------|------|-----------|---------|---------|-------|---------|-------|-----------|--------|------|-----------|---------|---------|------|")
    for rank, e in enumerate(registry.entries, 1):
        pf = f"{e.profit_factor:.2f}" if e.profit_factor is not None else "-"
@ -247,19 +621,35 @@ def _write_leaderboard_md(
        wr = f"{e.win_rate:.0%}" if e.win_rate is not None else "-"
        sharpe = f"{e.sharpe_ratio:.1f}" if e.sharpe_ratio is not None else "-"
        dd = f"{e.max_drawdown_pct:.1f}" if e.max_drawdown_pct is not None else "-"
        tgross = f"{e.avg_gross_exposure_pct:.1f}" if e.avg_gross_exposure_pct is not None else "-"
        tnet = f"{e.avg_net_exposure_pct:+.1f}" if e.avg_net_exposure_pct is not None else "-"
        tdim = f"{e.days_in_market_pct:.1f}" if e.days_in_market_pct is not None else "-"
        vpf = f"{e.valid_profit_factor:.2f}" if e.valid_profit_factor is not None else "-"
        vret = f"{e.valid_total_return_pct:+.1f}" if e.valid_total_return_pct is not None else "-"
        vwr = f"{e.valid_win_rate:.0%}" if e.valid_win_rate is not None else "-"
        vsharpe = f"{e.valid_sharpe_ratio:.1f}" if e.valid_sharpe_ratio is not None else "-"
        vdd = f"{e.valid_max_drawdown_pct:.1f}" if e.valid_max_drawdown_pct is not None else "-"
        vgross = f"{e.valid_avg_gross_exposure_pct:.1f}" if e.valid_avg_gross_exposure_pct is not None else "-"
        vnet = f"{e.valid_avg_net_exposure_pct:+.1f}" if e.valid_avg_net_exposure_pct is not None else "-"
        vdim = f"{e.valid_days_in_market_pct:.1f}" if e.valid_days_in_market_pct is not None else "-"
        ts = e.timestamp[:10] if e.timestamp else "-"
        lines.append(
-            f"| {rank} | {e.experiment_name} | {e.sqs_score:.1f} | {pf} | {ret} | {wr} | {sharpe} | {dd} | {e.trade_count} | {ts} |"
+            f"| {rank} | {e.experiment_name} | {e.sqs_score:.1f}"
            f" | {pf} | {ret} | {wr} | {sharpe} | {dd} | {e.trade_count} | {tgross} | {tnet} | {tdim}"
            f" | {vpf} | {vret} | {vwr} | {vsharpe} | {vdd} | {e.valid_trade_count} | {vgross} | {vnet} | {vdim}"
            f" | {ts} |"
        )
    # Recent entries (last 5)
    recent = list(reversed(journal_entries))[:5]
    if recent:
        registry_by_id = {entry.entry_id: entry for entry in registry.entries}
        lines.append("\n## Recent Entries")
        for je in recent:
            canonical_sqs = registry_by_id.get(je.entry_id).sqs_score if je.entry_id in registry_by_id else je.sqs_score
            lines.append(f"### {je.entry_id} ({je.timestamp[:10]}) \u2014 {je.experiment_name}")
            lines.append(f"Hypothesis: {je.hypothesis}")
-            lines.append(f"Verdict: **{je.verdict.upper()}** (SQS {je.sqs_score})")
+            lines.append(f"Verdict: **{je.verdict.upper()}** (SQS {canonical_sqs})")
            if je.verdict_reasoning:
                lines.append(f"Reasoning: {je.verdict_reasoning}")
            if je.next_direction:
--- a/pyproject.toml
+++ b/pyproject.toml
@ -38,6 +38,9 @@ dev = [
    "types-beautifulsoup4",
 ]
 [project.scripts]
 fithia2 = "apps.tracker.cli:main"
 [tool.hatch.build.targets.wheel]
 packages = ["apps", "libs"]
--- a/tests/integration/backtest/test_backtest_run.py
+++ b/tests/integration/backtest/test_backtest_run.py
@ -323,6 +323,7 @@ class TestBacktestRunIntegration:
    def test_multi_engine_run_writes_per_engine_metrics(self, tmp_path):
        from apps.backtester.run import BacktestRunner
        from libs.backtest.domain import ExperimentManifest, StrategyEngineConfig
        import json
        import pyarrow.parquet as pq
        store = _build_multi_engine_store()
@ -382,3 +383,11 @@ class TestBacktestRunIntegration:
        trade_blotter = pq.read_table(run_dir / "artifacts" / "trade_blotter.parquet").to_pylist()
        engine_ids = {row["engine_id"] for row in trade_blotter}
        assert "earnings_after_close_short_v1" not in engine_ids
        equity_curve = pq.read_table(run_dir / "artifacts" / "daily_equity_curve.parquet").to_pylist()
        assert any(float(row["net_exposure"]) < 0 for row in equity_curve if float(row["gross_exposure"]) > 0)
        metrics_summary = json.loads((run_dir / "metrics" / "metrics_summary.json").read_text())
        assert "avg_gross_exposure_pct" in metrics_summary
        assert "avg_net_exposure_pct" in metrics_summary
        assert "days_in_market_pct" in metrics_summary
--- a/tests/unit/backtest/test_metrics.py
+++ b/tests/unit/backtest/test_metrics.py
@ -38,13 +38,19 @@ def _make_trade(
    )
-def _make_equity_state(date: dt.date, equity: float, n_positions: int = 0) -> DailyPortfolioState:
+def _make_equity_state(
    date: dt.date,
    equity: float,
    n_positions: int = 0,
    gross_exposure: float = 0.0,
    net_exposure: float = 0.0,
 ) -> DailyPortfolioState:
    return DailyPortfolioState(
        date=date,
        equity=equity,
        cash_available=equity,
-        gross_exposure=0.0,
+        gross_exposure=gross_exposure,
-        net_exposure=0.0,
+        net_exposure=net_exposure,
        reserved_risk_budget=0.0,
        unrealized_pnl=0.0,
        realized_pnl=0.0,
@ -168,6 +174,25 @@ class TestSharpeRatio:
        assert compute_sharpe_ratio(curve) is None
 class TestExposureMetrics:
    def test_avg_gross_and_net_exposure_pct(self):
        from libs.backtest.metrics import (
            compute_avg_gross_exposure_pct,
            compute_avg_net_exposure_pct,
            compute_days_in_market_pct,
        )
        curve = [
            _make_equity_state(dt.date(2026, 1, 5), 100_000, gross_exposure=0.0, net_exposure=0.0),
            _make_equity_state(dt.date(2026, 1, 6), 100_000, gross_exposure=20_000.0, net_exposure=-20_000.0),
            _make_equity_state(dt.date(2026, 1, 7), 100_000, gross_exposure=10_000.0, net_exposure=5_000.0),
        ]
        assert compute_avg_gross_exposure_pct(curve) == pytest.approx(10.0)
        assert compute_avg_net_exposure_pct(curve) == pytest.approx(-5.0)
        assert compute_days_in_market_pct(curve) == pytest.approx(66.6666666667)
 class TestStopExitRate:
    def test_basic(self):
        from libs.backtest.metrics import compute_stop_exit_rate
@ -218,6 +243,23 @@ class TestBuildMetricsBundle:
        assert m.trade_count == 2
        assert m.win_rate == pytest.approx(0.5)
        assert m.total_return_pct == pytest.approx(5.0)
        assert m.avg_gross_exposure_pct == pytest.approx(0.0)
        assert m.avg_net_exposure_pct == pytest.approx(0.0)
        assert m.days_in_market_pct == pytest.approx(0.0)
    def test_builds_exposure_metrics(self):
        from libs.backtest.metrics import build_metrics_bundle
        curve = [
            _make_equity_state(dt.date(2026, 1, 5), 100_000, gross_exposure=0.0, net_exposure=0.0),
            _make_equity_state(dt.date(2026, 1, 6), 100_000, gross_exposure=30_000.0, net_exposure=-10_000.0),
            _make_equity_state(dt.date(2026, 1, 7), 100_000, gross_exposure=10_000.0, net_exposure=10_000.0),
        ]
        m = build_metrics_bundle([], curve)
        assert m.avg_gross_exposure_pct == pytest.approx(13.3333333333)
        assert m.avg_net_exposure_pct == pytest.approx(0.0)
        assert m.days_in_market_pct == pytest.approx(66.6666666667)
    def test_empty_trades(self):
        from libs.backtest.metrics import build_metrics_bundle
--- a/tests/unit/backtest/test_tracker.py
+++ b/tests/unit/backtest/test_tracker.py
@ -14,11 +14,17 @@ from libs.backtest.domain import (
 )
 from libs.backtest.tracker import (
    _normalize,
    _normalize_band,
    _normalize_inverse,
    append_journal_entry,
    build_split_result,
    check_duplicate,
    compute_public_sqs,
    compute_promotion_score,
    compute_sqs,
    compute_sqs_v2,
    compute_unified_score,
    compute_unified_split_quality,
    get_next_entry_id,
    load_journal,
    rebuild_registry,
@ -66,6 +72,18 @@ class TestNormalizeInverse:
        assert _normalize_inverse(None, low=10.0, high=1.0) == 0.0
 class TestNormalizeBand:
    def test_band_plateau_scores_max(self):
        assert _normalize_band(60.0, low_bad=10.0, low_good=40.0, high_good=80.0, high_bad=100.0) == 100.0
    def test_below_band_ramps_up(self):
        assert _normalize_band(25.0, low_bad=10.0, low_good=40.0, high_good=80.0, high_bad=100.0) == pytest.approx(50.0)
    def test_outside_band_scores_zero(self):
        assert _normalize_band(5.0, low_bad=10.0, low_good=40.0, high_good=80.0, high_bad=100.0) == 0.0
        assert _normalize_band(100.0, low_bad=10.0, low_good=40.0, high_good=80.0, high_bad=100.0) == 0.0
 # ---------------------------------------------------------------------------
 # compute_sqs
 # ---------------------------------------------------------------------------
@ -162,6 +180,257 @@ class TestComputeSQS:
        assert sqs_profit > sqs_risk
 class TestComputeSQSv2:
    def test_missing_exposure_returns_none(self):
        m = MetricsBundle(
            trade_count=50,
            profit_factor=1.5,
            total_return_pct=2.0,
            max_drawdown_pct=2.0,
            sharpe_ratio=1.2,
            win_rate=0.55,
            monthly_win_rate=0.60,
            equity_curve_r_squared=0.50,
        )
        sqs_v2, breakdown = compute_sqs_v2(m)
        assert sqs_v2 is None
        assert breakdown == {}
    def test_capital_efficiency_can_break_tie(self):
        base = dict(
            trade_count=50,
            profit_factor=1.5,
            total_return_pct=2.0,
            max_drawdown_pct=2.0,
            sharpe_ratio=1.2,
            win_rate=0.55,
            monthly_win_rate=0.60,
            equity_curve_r_squared=0.50,
            days_in_market_pct=60.0,
        )
        efficient = MetricsBundle(
            **base,
            avg_gross_exposure_pct=4.0,
            avg_net_exposure_pct=-1.0,
        )
        inefficient = MetricsBundle(
            **base,
            avg_gross_exposure_pct=10.0,
            avg_net_exposure_pct=-1.0,
        )
        efficient_score, _ = compute_sqs_v2(efficient)
        inefficient_score, _ = compute_sqs_v2(inefficient)
        assert efficient_score is not None
        assert inefficient_score is not None
        assert efficient_score > inefficient_score
 class TestComputePromotionScore:
    def test_requires_valid_and_test(self):
        score, breakdown = compute_promotion_score(None, None)
        assert score is None
        assert breakdown == {}
    def test_penalizes_test_only_outperformance(self):
        overfit_test = SplitResult(
            run_id="bt_test",
            trade_count=56,
            profit_factor=2.0,
            total_return_pct=2.2,
            win_rate=0.61,
            max_drawdown_pct=0.4,
            sharpe_ratio=4.2,
            monthly_win_rate=1.0,
            equity_curve_r_squared=0.88,
            avg_gross_exposure_pct=6.2,
            avg_net_exposure_pct=1.1,
            days_in_market_pct=76.6,
        )
        overfit_valid = SplitResult(
            run_id="bt_valid",
            trade_count=53,
            profit_factor=1.24,
            total_return_pct=0.7,
            win_rate=0.51,
            max_drawdown_pct=0.9,
            sharpe_ratio=1.2,
            monthly_win_rate=0.75,
            equity_curve_r_squared=0.15,
            avg_gross_exposure_pct=7.5,
            avg_net_exposure_pct=1.6,
            days_in_market_pct=77.2,
        )
        robust_test = SplitResult(
            run_id="bt_test_robust",
            trade_count=22,
            profit_factor=4.19,
            total_return_pct=1.24,
            win_rate=0.73,
            max_drawdown_pct=0.22,
            sharpe_ratio=3.6,
            monthly_win_rate=0.67,
            equity_curve_r_squared=0.92,
            avg_gross_exposure_pct=2.3,
            avg_net_exposure_pct=-0.8,
            days_in_market_pct=42.6,
        )
        robust_valid = SplitResult(
            run_id="bt_valid_robust",
            trade_count=28,
            profit_factor=2.31,
            total_return_pct=1.3,
            win_rate=0.68,
            max_drawdown_pct=0.37,
            sharpe_ratio=3.05,
            monthly_win_rate=0.75,
            equity_curve_r_squared=0.49,
            avg_gross_exposure_pct=4.1,
            avg_net_exposure_pct=-1.9,
            days_in_market_pct=57.9,
        )
        overfit_score, overfit_breakdown = compute_promotion_score(overfit_test, overfit_valid)
        robust_score, robust_breakdown = compute_promotion_score(robust_test, robust_valid)
        assert overfit_score is not None
        assert robust_score is not None
        assert robust_score > overfit_score
        assert overfit_breakdown["floor_quality"] < overfit_breakdown["test_quality"]
 class TestComputeUnifiedScore:
    def test_requires_valid_and_test(self):
        score, breakdown = compute_unified_score(None, None)
        assert score is None
        assert breakdown == {}
    def test_split_quality_rewards_efficiency(self):
        efficient = MetricsBundle(
            trade_count=30,
            profit_factor=2.0,
            total_return_pct=1.2,
            max_drawdown_pct=0.4,
            sharpe_ratio=2.5,
            win_rate=0.60,
            monthly_win_rate=0.75,
            equity_curve_r_squared=0.70,
            avg_gross_exposure_pct=3.0,
            avg_net_exposure_pct=-1.0,
            days_in_market_pct=55.0,
        )
        inefficient = efficient.model_copy(
            update={"avg_gross_exposure_pct": 8.0, "days_in_market_pct": 85.0}
        )
        efficient_score, _ = compute_unified_split_quality(efficient)
        inefficient_score, _ = compute_unified_split_quality(inefficient)
        assert efficient_score is not None
        assert inefficient_score is not None
        assert efficient_score > inefficient_score
    def test_overfit_strategy_scores_below_robust_strategy(self):
        overfit_test = SplitResult(
            run_id="bt_step35_test",
            trade_count=56,
            profit_factor=2.009,
            total_return_pct=2.181,
            win_rate=0.607,
            max_drawdown_pct=0.442,
            sharpe_ratio=4.217,
            monthly_win_rate=1.0,
            equity_curve_r_squared=0.889,
            avg_gross_exposure_pct=6.228,
            avg_net_exposure_pct=1.126,
            days_in_market_pct=76.6,
        )
        overfit_valid = SplitResult(
            run_id="bt_step35_valid",
            trade_count=53,
            profit_factor=1.243,
            total_return_pct=0.708,
            win_rate=0.509,
            max_drawdown_pct=0.885,
            sharpe_ratio=1.180,
            monthly_win_rate=0.75,
            equity_curve_r_squared=0.146,
            avg_gross_exposure_pct=7.511,
            avg_net_exposure_pct=1.567,
            days_in_market_pct=77.2,
        )
        robust_test = SplitResult(
            run_id="bt_step52_test",
            trade_count=22,
            profit_factor=3.352,
            total_return_pct=1.011,
            win_rate=0.773,
            max_drawdown_pct=0.243,
            sharpe_ratio=3.032,
            monthly_win_rate=0.667,
            equity_curve_r_squared=0.847,
            avg_gross_exposure_pct=2.858,
            avg_net_exposure_pct=-2.223,
            days_in_market_pct=53.2,
        )
        robust_valid = SplitResult(
            run_id="bt_step52_valid",
            trade_count=25,
            profit_factor=5.660,
            total_return_pct=1.819,
            win_rate=0.72,
            max_drawdown_pct=0.217,
            sharpe_ratio=4.756,
            monthly_win_rate=1.0,
            equity_curve_r_squared=0.661,
            avg_gross_exposure_pct=3.584,
            avg_net_exposure_pct=-2.593,
            days_in_market_pct=56.1,
        )
        overfit_score, overfit_breakdown = compute_unified_score(overfit_test, overfit_valid)
        robust_score, robust_breakdown = compute_unified_score(robust_test, robust_valid)
        assert overfit_score is not None
        assert robust_score is not None
        assert robust_score > overfit_score
        assert overfit_breakdown["gap_quality"] < robust_breakdown["gap_quality"]
 class TestComputePublicSQS:
    def test_public_score_prefers_integrated_and_is_harsher(self):
        test_result = SplitResult(
            run_id="bt_test",
            trade_count=22,
            profit_factor=3.352,
            total_return_pct=1.011,
            win_rate=0.773,
            max_drawdown_pct=0.243,
            sharpe_ratio=3.032,
            monthly_win_rate=0.667,
            equity_curve_r_squared=0.847,
            avg_gross_exposure_pct=2.858,
            avg_net_exposure_pct=-2.223,
            days_in_market_pct=53.2,
        )
        valid_result = SplitResult(
            run_id="bt_valid",
            trade_count=25,
            profit_factor=5.660,
            total_return_pct=1.819,
            win_rate=0.72,
            max_drawdown_pct=0.217,
            sharpe_ratio=4.756,
            monthly_win_rate=1.0,
            equity_curve_r_squared=0.661,
            avg_gross_exposure_pct=3.584,
            avg_net_exposure_pct=-2.593,
            days_in_market_pct=56.1,
        )
        public_score, _, source = compute_public_sqs(test_result, valid_result)
        raw_integrated_score, _ = compute_unified_score(test_result, valid_result)
        assert source == "integrated"
        assert public_score == raw_integrated_score
        assert public_score is not None
        assert public_score < 60.0
 # ---------------------------------------------------------------------------
 # build_split_result
 # ---------------------------------------------------------------------------
@ -177,12 +446,18 @@ def test_build_split_result():
        sharpe_ratio=0.8,
        monthly_win_rate=0.60,
        equity_curve_r_squared=0.40,
        avg_gross_exposure_pct=18.5,
        avg_net_exposure_pct=-6.5,
        days_in_market_pct=27.0,
    )
    sr = build_split_result("test", "bt_run123", m)
    assert sr.run_id == "bt_run123"
    assert sr.trade_count == 50
    assert sr.profit_factor == 1.2
    assert sr.total_return_pct == 2.5
    assert sr.avg_gross_exposure_pct == 18.5
    assert sr.avg_net_exposure_pct == -6.5
    assert sr.days_in_market_pct == 27.0
 # ---------------------------------------------------------------------------
@ -298,6 +573,11 @@ class TestRebuildRegistry:
                profit_factor=1.0 + i * 0.2,
                total_return_pct=float(i),
                win_rate=0.5,
                monthly_win_rate=0.6,
                equity_curve_r_squared=0.5,
                avg_gross_exposure_pct=5.0,
                avg_net_exposure_pct=-1.0,
                days_in_market_pct=60.0,
            )
            entry = JournalEntry(
                entry_id=f"IMP-{i+1:04d}",
@ -305,6 +585,8 @@ class TestRebuildRegistry:
                experiment_name=name,
                hypothesis=f"h{i}",
                sqs_score=sqs,
                promotion_score=sqs - 5,
                unified_score=sqs - 10,
                results={"test": test_result},
                verdict="better" if sqs > 50 else "worse",
            )
@ -312,11 +594,13 @@ class TestRebuildRegistry:
        registry = rebuild_registry(journal_path, registry_path, leaderboard_path)
-        # Sorted by SQS descending
+        # Sorted by public SQS descending
        assert len(registry.entries) == 3
-        assert registry.entries[0].experiment_name == "exp_high"
+        assert registry.entries[0].experiment_name == "exp_mid"
        assert registry.entries[0].sqs_score == 70.0
        assert registry.entries[2].experiment_name == "exp_low"
        assert registry.entries[0].sqs_v2_score is not None
        assert registry.entries[0].promotion_score is not None
        assert registry.entries[0].sqs_score < 70.0
        # Files exist
        assert registry_path.exists()
@ -327,3 +611,5 @@ class TestRebuildRegistry:
        assert "exp_high" in lb_text
        assert "exp_low" in lb_text
        assert "| # |" in lb_text
        assert "| # | Experiment | SQS |" in lb_text
        assert "[T]Gross%" in lb_text