# Dragon Refined Edge Review

## Scope
- Target branch: `alpha_first_glued_refined_hot_cap`
- Control branch: `alpha_first_selective_veto`
- Evaluation window: `2016-01-01` to `2025-12-31`

## Headline
- control: trades `102`, win_rate `47.06%`, avg_return `2.86%`, profit_factor `4.04`
- refined: trades `91`, win_rate `52.75%`, avg_return `3.42%`, profit_factor `5.11`
- refined minus control: trades `-11`, avg_return `0.56%`, profit_factor `1.06`

## Main Edge Source
- Refined alpha is still primarily a `glued_buy` story, but now with stricter removal of weak short-holding glued trades.
- The branch is not winning by adding new complex trade paths; it is winning by deleting low-quality short trades while preserving the medium and long-holding winners.

## Entry Family Decomposition
- `glued_buy` [core_alpha_family]: trades `50`, share `54.95%`, avg_return `4.92%`, sum_return `246.25%`, profit_factor `7.15`, top_exit `knife_take_profit_2_glued`
- `early_crash_probe_buy` [core_alpha_family]: trades `6`, share `6.59%`, avg_return `4.62%`, sum_return `27.70%`, profit_factor `10.52`, top_exit `knife_take_profit_2_glued`
- `dual_gold_resonance_buy` [support_alpha_family]: trades `13`, share `14.29%`, avg_return `1.06%`, sum_return `13.72%`, profit_factor `2.08`, top_exit `knife_take_profit_2_glued`
- `oversold_recovery_buy` [core_alpha_family]: trades `4`, share `4.40%`, avg_return `3.35%`, sum_return `13.39%`, profit_factor `6.65`, top_exit `knife_take_profit_2_glued`
- `post_sell_rebound_buy` [secondary_research_family]: trades `4`, share `4.40%`, avg_return `1.84%`, sum_return `7.35%`, profit_factor `2.56`, top_exit `negative_a1_no_b1_recovery:kdj_sell`

## Weak Research Pockets
- `deep_oversold_rebound_buy:deep_capitulation`: trades `1`, avg_return `-2.05%`, sum_return `-2.05%`, profit_factor `0.00`
- `oversold_reversal_after_ql_buy`: trades `1`, avg_return `-0.77%`, sum_return `-0.77%`, profit_factor `0.00`
- `deep_oversold_rebound_buy:positive_b1_rebound`: trades `1`, avg_return `-0.18%`, sum_return `-0.18%`, profit_factor `0.00`
- `deep_oversold_rebound_buy:mixed_oversold`: trades `2`, avg_return `0.01%`, sum_return `0.02%`, profit_factor `1.01`
- `deep_oversold_rebound_buy:shallow_false_start`: trades `1`, avg_return `0.12%`, sum_return `0.12%`, profit_factor `inf`

## Kept Winner Structure
- `glued_buy / 41d+`: winners `9`, avg_return `25.57%`, sum_return `230.16%`
- `glued_buy / 11-20d`: winners `10`, avg_return `2.50%`, sum_return `25.01%`
- `glued_buy / 21-40d`: winners `2`, avg_return `12.42%`, sum_return `24.84%`
- `early_crash_probe_buy / 41d+`: winners `1`, avg_return `20.56%`, sum_return `20.56%`
- `dual_gold_resonance_buy / 21-40d`: winners `2`, avg_return `9.32%`, sum_return `18.65%`

## Entry / Exit Interaction Attribution
- positive `mid_regime / glued_buy -> crash_protection_exit`: trades `2`, sum_return `79.18%`, avg_return `39.59%`, PF `inf`
- positive `mid_regime / glued_buy -> prewarning_reduction_exit`: trades `2`, sum_return `49.55%`, avg_return `24.77%`, PF `inf`
- positive `mid_regime / glued_buy -> predictive_b1_break_exit`: trades `1`, sum_return `36.85%`, avg_return `36.85%`, PF `inf`
- positive `mid_regime / glued_buy -> high_regime_confirmed_exit:kdj_sell`: trades `2`, sum_return `32.29%`, avg_return `16.14%`, PF `inf`
- positive `high_regime / glued_buy -> high_regime_momentum_break`: trades `1`, sum_return `32.17%`, avg_return `32.17%`, PF `inf`
- positive `crash_probe_regime / early_crash_probe_buy -> ql_high_zone_take_profit`: trades `1`, sum_return `20.56%`, avg_return `20.56%`, PF `inf`
- positive `low_oversold_regime / oversold_recovery_buy -> high_regime_confirmed_exit:kdj_sell`: trades `1`, sum_return `14.72%`, avg_return `14.72%`, PF `inf`
- positive `low_oversold_regime / dual_gold_resonance_buy -> prewarning_reduction_exit`: trades `1`, sum_return `13.29%`, avg_return `13.29%`, PF `inf`

## Drag Interaction Pockets
- drag `low_oversold_regime / deep_oversold_rebound_buy:positive_b1_rebound -> knife_take_profit_2_glued`: trades `1`, sum_return `-0.18%`, avg_return `-0.18%`, PF `0.00`
- drag `low_oversold_regime / glued_buy -> knife_take_profit_2_glued`: trades `2`, sum_return `-0.31%`, avg_return `-0.16%`, PF `0.69`
- drag `rebound_after_sell_regime / post_sell_rebound_buy -> knife_take_profit_2_glued`: trades `1`, sum_return `-0.49%`, avg_return `-0.49%`, PF `0.00`
- drag `rebound_after_sell_regime / oversold_reversal_after_ql_buy -> knife_take_profit_2_glued`: trades `1`, sum_return `-0.77%`, avg_return `-0.77%`, PF `0.00`
- drag `crash_probe_regime / early_crash_probe_buy -> hard_exit:kdj_sell`: trades `1`, sum_return `-0.85%`, avg_return `-0.85%`, PF `0.00`
- drag `high_regime / glued_buy -> knife_take_profit_2_glued`: trades `1`, sum_return `-0.89%`, avg_return `-0.89%`, PF `0.00`
- drag `low_oversold_regime / deep_oversold_rebound_buy:classic_oversold -> negative_a1_no_b1_recovery:kdj_sell`: trades `1`, sum_return `-0.92%`, avg_return `-0.92%`, PF `0.00`
- drag `low_oversold_regime / dual_gold_resonance_buy -> ql_mid_zone_take_profit`: trades `1`, sum_return `-1.20%`, avg_return `-1.20%`, PF `0.00`

## Removed-Trade Recheck
- removed trades vs control: `11`
- removed-set avg_return `-1.81%`
- removed-set win_rate `0.00%`
- removed-set profit_factor `0.00`
- KEEP_REMOVAL `11` | OBSERVE_REMOVAL `0` | OVER_REMOVAL `0`
- `hot_positive_b1_cap75`: trades `7`, avg_return `-1.50%`, avg_holding `5.1`, avg_mfe `1.40%`
- `low_weak_range`: trades `4`, avg_return `-2.35%`, avg_holding `3.0`, avg_mfe `0.96%`

## Quant Judgment
- Core alpha remains concentrated in `glued_buy`, `early_crash_probe_buy`, and the preserved medium/long holding structure.
- `dual_gold_resonance_buy` and `deep_oversold_rebound_buy:classic_oversold` remain support families, not the main alpha engine.
- Weak pockets still exist in secondary rebound / weak deep-oversold variants, but they are not where the refined branch gets its headline improvement.
- The refined branch improves mainly by deleting low-quality short glued trades; this remains explainable and not dependent on deleting profitable samples.
- The next step should therefore move to execution-aware robustness, not back to workbook-style residual tuning.