# ETH composite lab — measured step 2

252 unique configs (208 composites + 44 baselines), three adaptive rounds [96, 64, 48]. One unchanged step-1 scorer; each attempt has a config, lineage, ETH/BTC metrics and two parameter neighbours. Selection: highest mean entry net/day over four development test folds, minimum 20 trades. Exit winners are descriptive only.

Data: accepted OKX ETH/BTC cache, 160 days, 230,400 confirmed 1m bars each; 3m/5m/15m resampled. All file/canonical hashes checked. 130 development days; sealed 2026-10-13–2026-11-12 holdout, opened once after freezing. No full refetch was needed. BTC never selected parameters.

## Best development-selected composite per TF; 30-day holdout

All returns below are net **% equity/day**, 1% margin ×50 leverage, additive. CI is a seeded 1,000-replicate day-block interval.

| TF | Config | Dev entry | Holdout entry [95% CI] | PF / trades/day | DD% / liqs | Holdout exit |
|---|---|---:|---|---|---|---:|
| 1m | `4941d5d4a317` | 0.00143 | -0.00875 [-0.02163, 0.0] | 0.0 / 0.067 | 0.2624 / 0 | -0.00383 |
| 3m | `65c35983c839` | -0.01024 | -0.01586 [-0.04479, 0.00959] | 0.4048 / 0.267 | 0.6314 / 0 | 0.02121 |
| 5m | `115cad5fc8f8` | -0.01766 | -0.03429 [-0.08193, 0.00451] | 0.2586 / 0.5 | 1.2187 / 0 | 0.06806 |
| 15m | `fb01a045c2be` | 0.00358 | -0.00048 [-0.08698, 0.06885] | 0.9861 / 0.267 | 1.0269 / 0 | 0.11077 |

## Best baseline per TF (selected before opening)

| TF | Indicator | Holdout entry [95% CI] | Holdout exit | Entry DD% / liqs |
|---|---|---|---:|---|
| 1m | volume_spike  | -0.85298 [-1.02118, -0.69382] | -0.53823 | 25.6865 / 0 |
| 3m | sigma_pair diverge | -0.15619 [-0.27843, -0.02432] | -0.14317 | 4.8852 / 0 |
| 5m | layered_bb  | -0.14093 [-0.26475, -0.01474] | -0.35417 | 4.722 / 0 |
| 15m | sigma_pair diverge | -0.13734 [-0.33312, 0.04481] | 0.27940 | 4.9903 / 0 |

## Risk and adversarial verdict

Entry holdout point estimates beat costs on **none of the four timeframes**. No composite entry has a positive lower 95% bound.
Exit point estimates are positive on 3m/5m/15m, but every exit 95% interval includes zero (only 6/12/4 trades). No statistically supported cost-beating result emerged; every composite BTC entry is also negative. The low trade counts show that selection mainly reduced exposure to losing trades.
White reality check (entries), recorded at freeze against the best development baseline per TF: p=0.0010, 252 attempts, 1,000 paired circular five-day bootstraps. This p is degenerate: the 1m benchmark loses, so a strategy that never trades also rejects. Against max(best baseline, flat): p=1.0000 (observed max excess 0.00143%/day). Adaptive generation reuses development data; this is a search diagnostic, not independent significance.
White reality check (exits), recorded at freeze against the best development baseline per TF: p=0.2507, 252 attempts, 1,000 paired circular five-day bootstraps. Against max(best baseline, flat): p=1.0000 (observed max excess 0.02451%/day). Adaptive generation reuses development data; this is a search diagnostic, not independent significance.
Selection used the in-sample mean of dev test blocks 1-4, not walk-forward. A true expanding walk-forward of the same rule (pick on folds <k, score fold k) gives OOS entry net/day: 1m composite -0.01227 vs baseline -0.56019; 3m composite -0.03031 vs baseline -0.22204; 5m composite -0.02535 vs baseline -0.23342; 15m composite -0.04046 vs baseline -0.01425.
BTC and threshold ±5 neighbours were measured for every attempt; all numbers, entry MFE/MAE (5/15/30), ±1 ATR hit rates, exit capture/giveback/ZigZag timing, fold returns and intervals are in results.json/leaderboard.json. Caveat: the threshold-5 neighbour is identical to the main config for 208/208 composites (quantised scores), so only threshold+5 and BTC carry neighbour evidence. Weak fold or neighbour results and liquidations remain visible even for winners.
Limits inherited from step 1: overlapping entry trades have no aggregate margin cap; additive drawdowns can exceed 100%; exit scorer omits the final position and recognizes liquidation at its later closing event. Holdout initializes indicators cold: the 1m (mtf=align) and 15m finalists cannot signal for the first ~2.1-2.2 of 30 days. Fold-crossing trades are not purged. Only net/day and average R have bootstrap intervals. The 15m state uses closed higher bars only; forming candles never affect signals. Parity fixtures use development/synthetic data only.

Reproduce: `python3 -m lab.search --jobs 2` (development/checkpoints), `bash scripts/check-composite.sh` (offline tests). `python3 -m lab.search --final` is a one-time operation and refuses an existing opening log. `python3 -m lab.search --rebuild` regenerates the named outputs from checkpoints and the recorded opening, without selection. Frozen IDs/settings are in configs/frozen.json and artifacts/settings.json.