cyb50-quant/research/dragon/v2/dragon_formal_strategy_governance.md

Dragon Formal Strategy Governance

Objective

Turn alpha_first_glued_refined_hot_cap from a governed research candidate into a strategy asset with explicit:

• branch roles
• promotion rules
• downgrade rules
• change protocol
• monitoring protocol

Branch Roles

• workbook_preserving

  • role: reconstruction reference
  • use: workbook explanation, regression checks, reverse-engineering audit
  • change rule: do not optimize for alpha on this branch

• alpha_first_selective_veto

  • role: benchmark control branch
  • use: compare refined against the current conservative alpha baseline
  • change rule: keep frozen unless governance explicitly resets the control

• alpha_first_glued_refined_hot_cap

  • role: forward default branch
  • use: primary research, robustness validation, and formal promotion candidate
  • change rule: all new strategy research starts here unless the task is explicitly reconstruction-oriented

Governance State

• governance-neutral state: DUAL_TRACK_GOVERNANCE
• operational default for forward research: alpha_first_glued_refined_hot_cap
• benchmark control: alpha_first_selective_veto
• reconstruction reference: workbook_preserving

Promotion Rule

Promote alpha_first_glued_refined_hot_cap to the formal strategy branch if any one of the following governance choices is accepted:

• workbook overlap is treated as a cost metric rather than a hard gate
• overlap tolerance is widened to at least the current refined delta
• the official objective becomes "robust alpha first on 399673 under explainable rules"

Promotion also requires all of the following quantitative gates to remain true:

• headline quality gate: pass
• removed-trade quality gate: pass
• local stability gate: pass
• execution-aware robustness gate: pass
• drawdown quality gate: pass

Execution-Aware Robustness Gate

The branch remains promotion-eligible only if it continues to satisfy:

• next-open execution still beats control on avg return and profit factor
• next-open plus 20 bps/side still beats control on CAGR and profit factor
• max drawdown is not worse than control under next-open stress
• loss clustering is not worse than control on:
  • max loss streak
  • worst 5-trade sum
  • drawdown duration

Downgrade Rule

Downgrade refined from forward default back to control-default only if one of the following happens:

• a new rerun shows refined no longer beats control under next-open execution
• removed-trade review starts showing meaningful profitable over-removals
• local sensitivity turns knife-edge
• short-holding cleanup disappears while overlap cost remains large
• a newer branch dominates refined on quality and stability together

Change Protocol

Class A: Frozen Formal Changes

These changes may not go directly into the formal refined branch. They require a new named research branch:

• any change to glued selective veto thresholds
• any change to deep-oversold selective veto thresholds
• any change to predictive bridge exits that alters real trade flow
• any change that shifts the removed-trade set

Required process:

1. create a new branch name
2. rerun attribution pack
3. rerun execution-aware robustness pack
4. compare against refined and control
5. write an explicit promotion or rejection memo

Class B: Research Changes

These may be explored in research branches but not merged into refined without review:

• secondary rebound entry redesign
• weak deep-oversold subtype redesign
• exit interaction changes outside hygiene-only scope
• regime gating changes

Required process:

1. local rationale
2. full-sample rerun
3. stress rerun
4. family-level attribution check

Class C: Hygiene Changes

These may be accepted more easily if they do not alter real-trade structure materially:

• auxiliary signal compression
• reporting-only changes
• non-trading metadata tagging
• documentation and monitoring updates

Required process:

1. verify no harmful real-trade drift
2. verify no hidden metric regression
3. document the change

Default Comparison Rule

• If the question is "is the new branch a better strategy?", compare against alpha_first_glued_refined_hot_cap.
• If the question is "is workbook reconstruction preserved?", compare against workbook_preserving.
• Do not mix those two objectives in one headline decision.

Required Deliverables Before Future Promotion

Any candidate that wants to replace refined must provide:

• branch summary
• branch trade diff
• removed-trade review
• local sensitivity review
• execution-aware robustness review
• final governance judgment

Current Governance Judgment

• Refined is already the stronger quant branch.
• Non-promotion is currently a governance choice, not an evidence gap.
• Therefore:
  • refined should be treated as the default forward strategy branch
  • control should remain frozen for comparison
  • workbook branch should remain an audit baseline only