# Dragon Formal Strategy Memo

## Strategy Identity

- Instrument scope: `399673` only
- Current strongest branch: `alpha_first_glued_refined_hot_cap`
- Control branch: `alpha_first_selective_veto`
- Reconstruction branch: `workbook_preserving`

## Official Objective

Build an explainable, robust alpha strategy on `399673` where:

- robustness is more important than perfect workbook resemblance
- useless short trades are removed first
- large trend profits are preserved
- annualized return is a secondary consequence, not the first optimization target

## Current Strategy Judgment

### Quant Best Branch

- `alpha_first_glued_refined_hot_cap`

Why:

- better avg return
- better profit factor
- better win rate
- better drawdown profile
- better execution-aware stress behavior
- better short-loss clustering than control

### Conservative Governance Branch

- `alpha_first_selective_veto`

Why:

- smaller workbook divergence
- easier bridge from reconstruction to alpha-first research

## Edge Source

The current edge is not broad and diffuse. It is concentrated in:

- `glued_buy` as the main alpha engine
- preserved medium and long holding winners
- deletion of weak short-holding glued trades
- support from `early_crash_probe_buy`
- secondary support from `oversold_recovery_buy` and `dual_gold_resonance_buy`

The branch does **not** mainly improve by:

- adding many new trade paths
- aggressive curve-fit threshold searching
- deleting profitable samples

## Known Weak Points

- weak deep-oversold subtypes remain fragile
- secondary rebound families are not strong edge sources
- worst loss family is still `glued_buy`, even though refined has improved it
- governance still needs to decide whether workbook overlap is a hard gate or a cost metric

## What Is Already Proven

- removed trades vs control are low quality
- refined is locally stable under nearby threshold perturbations
- refined remains ahead under transaction cost pressure
- refined remains ahead under next-open and next-close execution assumptions
- refined has better loss clustering and drawdown duration than control

## What Must Not Be Done

- do not resume workbook-style micro-alignment as the main optimization path
- do not blind-tune predictive bridge exits
- do not change glued selective veto thresholds inside the formal branch without opening a new research branch
- do not treat auxiliary signal cleanup as the main alpha frontier

## Current Operating Rule

- forward default: `alpha_first_glued_refined_hot_cap`
- benchmark control: `alpha_first_selective_veto`
- reconstruction audit baseline: `workbook_preserving`

## Next Research Focus

- execution-aware robustness should remain the live validation layer
- future optimization should target weak secondary families, not the core glued alpha trunk
- any new candidate must beat refined on both quality and robustness, not just workbook resemblance

## Bottom Line

`alpha_first_glued_refined_hot_cap` is now the best strategy branch in the workspace from a quant point of view.

If it is not promoted, that should be treated as a governance choice.
It should no longer be treated as an unresolved research question.