# ETH edge search v3 — Round B

**No validated edge.** The best pre-final rule is `8e3d24568c696980` (regime_switch, 60m, taker). Its development mean is 0.10607% equity/day on 57 trades; the strongest unrestricted development candidate remains `8c6176fab4c83b51` (0.12964%/day), with no calendar-valid historical ML final score.

Added **1,296** configurations; total recorded evaluation lower bound **6216** (7176 DSR arms). All earlier holdouts are burned. Earlier-inner-fold selection, complete-trade purging and 1-day embargo give -0.08580%/day on 68 next-fold trades across the expanded pool; Round A informed this research, so the development folds are adaptive evidence. White versus FLAT: p=0.929 on stored vectors; missing legacy neighbors force conservative global p=1.

Five finalists were frozen before the exclusive one-time final opening. **No final reselection.** The 2024 check now covers all 366 days, after 42 days of warm-up. BTC/SOL cover the same full untouched pre-development year, rather than their contaminated 2026 caches; these are not full listing histories. The short post-v2 window remains void where a full execution horizon cannot fit. The immutable seal retains its pre-opening flag; the lock and final results record the opening.

Best preselected rule, **net equity %/day [95% CI]**. Account margin is 1% per trade; these are not returns on an account fully exposed at the stated leverage.

| Check | Leverage | Net %/day [CI] | Trades (per day) | Max DD % | Liquidations | Fill rate |
|---|---:|---:|---:|---:|---:|---:|
| btc_full | 10x | -0.01173 [-0.02903, 0.00499] | 178 (0.486) | 8.566 | 0 | 100.0% |
| btc_full | 20x | -0.02316 [-0.05639, 0.01067] | 178 (0.486) | 16.345 | 1 | 100.0% |
| btc_full | 50x | -0.05173 [-0.13132, 0.03025] | 180 (0.492) | 32.264 | 48 | 100.0% |
| eth_2024 | 10x | -0.00511 [-0.02632, 0.01631] | 171 (0.467) | 7.987 | 0 | 100.0% |
| eth_2024 | 20x | -0.00996 [-0.05242, 0.03299] | 171 (0.467) | 15.366 | 1 | 100.0% |
| eth_2024 | 50x | 0.00382 [-0.09578, 0.10683] | 177 (0.484) | 29.358 | 55 | 100.0% |
| eth_new | — | not scored: insufficient full forward window | — | — | — | — |
| sol_full | 10x | 0.01685 [-0.01052, 0.04526] | 160 (0.437) | 3.405 | 0 | 100.0% |
| sol_full | 20x | 0.03470 [-0.01899, 0.09098] | 160 (0.437) | 5.751 | 10 | 100.0% |
| sol_full | 50x | 0.04991 [-0.07063, 0.18090] | 179 (0.489) | 15.660 | 92 | 100.0% |

The complete finalist distribution, daily vectors, costs, missed-fill returns, adverse markouts, parameter neighbors and all tried configurations are in `results_v3.json`; flat best-rule parameters are in `settings_v3.json`. At least 100 final trades, neighboring-parameter/fold/symbol stability and multiplicity are required for any edge claim. Positive point estimates alone do not meet that standard.

**What killed the ideas:** shorter holds pay away small moves; maker improvement depends on fill selection and unknown queues; long holds are sparse, concentrated and liquidation-exposed; gates often trade less; learned models fail chronology for 2024. Funding outside the Binance proxy history is a conservative debit, not measured venue funding. OHLC fills omit queues, spread and mark-price liquidation. The bounded experiment neither validates a deployable edge nor proves that ETH has no edge.

Offline: `python3 -m unittest discover -s tests -t .`; `bash scripts/check-composite.sh` (vendored toolchain); `python3 -m lab.report_b_v3 report` rebuilds outputs from saved results without reopening finals. `lab.final_v3` refuses a second opening.