cyb50-quant/research/dragon/v2/openspec/changes/followthrough-lean-profit-loop/design.md

Context

The user clarified the real priority: make money by following trends, especially by recovering quickly when a hard veto blocks a move that later repairs into a valid trend.

GPT-Pro also already narrowed the next worthwhile work:

• execution timing
• entry-specific exit treatment
• optional high-zone shadow observation

The practical constraint is equally clear:

• keep it small-team sized
• do not build a broad research platform
• do not change RC1 production behavior before the evidence is stronger

Goals / Non-Goals

Goals

• Turn blocked glued weak rebounds into a reproducible observation dataset.
• Compare delayed reentry paths by subtype in a narrow shadow scope.
• Show whether the pain point is:
• no real followthrough
• execution timing slippage
• exit family clipping the reentry too early
• Keep outputs lightweight and easy to rerun.

Non-Goals

• Promote any followthrough branch to formal alpha in this phase.
• Relax default hard blocks in RC1.
• Introduce a new online monitoring subsystem.
• Rework all sell families.

Decisions

1. Research stays outside the RC1 path

This phase adds a standalone script and report only. No change to default branch config, governance, or forward pipeline behavior.

2. Candidate observation is rebuilt from actual strategy logic

Instead of inferring from static thresholds alone, the script replays the strategy state machine and captures every buy_block_glued_high_weak_rebound decision directly.

3. Reentry analysis stays subtype-specific

The script evaluates three isolated probe lanes:

• mid_zone_very_weak_b1
• high_zone_weak_b1
• ql_rebound_weak_followthrough

This keeps evidence clean and avoids mixed-lane conclusions.

4. Exit-treatment evidence is observational first

Before touching sell logic, the script measures:

• realized reentry return
• next-open vs same-close execution delta
• post-entry max-close opportunity
• post-exit followthrough left on the table

Risks / Trade-offs

• [Small sample sizes] accepted; the goal is to avoid broad, unjustified strategy edits.
• [Fetched raw history may change with new market data] accepted; this is a living research report, not a golden baseline.
• [Internal strategy hooks are reused] accepted; this keeps the review aligned with the real engine path.

Migration Plan

1. Add the lean review script and helper functions.
2. Add a focused regression test for known block/reentry mapping.
3. Generate outputs and publish the review report.
4. Record findings in OpenSpec acceptance and workspace memory.