docs: freeze Regime Trend v1 robustness plan

Author: dharmanan

research/regime-trend-v1/robustness-plan-v1.md

@@ -0,0 +1,95 @@
+# Regime Trend v1 — Frozen Robustness Plan
+
+Status: **Frozen before robustness results are viewed**
+
+This plan evaluates the already frozen `regime-trend-v1` baseline. It is not an optimization plan. No result from these checks may be used to replace the baseline parameters or select a better-looking neighbor.
+
+## 1. Data boundary
+
+Only these partitions may be read:
+
+- Development: 2019-01-01 through 2022-12-31 UTC
+- Validation: 2023-01-01 through 2024-12-31 UTC
+
+The final holdout beginning 2025-01-01 remains closed.
+
+## 2. Symbol-distribution gate
+
+Report BTCUSDT, ETHUSDT and BNBUSDT separately.
+
+For validation:
+
+- total net PnL must remain positive,
+- no single symbol may contribute more than 60% of the sum of positive symbol profits,
+- symbols with negative net PnL remain visible and are not excluded.
+
+This gate prevents one exceptional symbol from hiding weak cross-asset behavior.
+
+## 3. Chronological-block gate
+
+Validation is divided into eight fixed calendar quarters:
+
+- 2023-Q1 through 2024-Q4.
+
+Every quarter is counted, including quarters with zero trades.
+
+Pass condition:
+
+- at least 5 of 8 quarters must have positive net expectancy.
+
+Annual 2023 and 2024 results are also reported but are descriptive rather than a separate pass gate.
+
+## 4. Doubled-cost stress
+
+The frozen baseline uses:
+
+- commission: 0.10% per side,
+- slippage: 0.05% per side.
+
+The stress run uses:
+
+- commission: 0.20% per side,
+- slippage: 0.10% per side.
+
+All fills, fees, stops and trade outcomes are recalculated from the strategy engine. Existing trade ledgers are not merely repriced.
+
+Pass condition on validation:
+
+- total net PnL remains non-negative,
+- profit factor remains at least 1.00.
+
+## 5. One-factor parameter-neighborhood test
+
+The frozen baseline remains unchanged. Each neighbor changes exactly one parameter while all others stay at baseline.
+
+Neighbors:
+
+- EMA fast: 45, 55
+- EMA slow: 180, 220
+- Donchian lookback: 18, 22
+- normalized ATR floor: 0.0045, 0.0055
+- initial stop ATR multiple: 2.25, 2.75
+- trailing stop ATR multiple: 2.70, 3.30
+
+There are 12 neighbors. No combinatorial grid is allowed.
+
+Pass condition on validation:
+
+- at least 7 of 12 neighbors have positive total net PnL,
+- median neighbor profit factor is at least 1.00,
+- no more than 3 neighbors have profit factor below 0.90.
+
+A neighbor with no closed trades fails the positive-net count and is reported with an undefined profit factor.
+
+## 6. Overall robustness classification
+
+The robustness stage passes only if all three gates pass:
+
+1. symbol distribution,
+2. chronological blocks,
+3. doubled costs,
+4. parameter neighborhood.
+
+All four are required despite the numbering above. A failure does not permit retuning against validation. The baseline is either advanced unchanged to independent Pine parity work or stopped as `regime-trend-v1`.
+
+The final holdout remains unopened during this stage.