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

User Notes

• 用户当前关注 cyb50-quant/research/dragon/v2 下的回测文件分析与交易数据总结。
• 用户明确说明：回测表里的重复“卖点”应视为看空信号，不计入真实交易。

• 用户确认：C1=C4；QL凤凰线 的 B/S 是上穿/下穿；只做 399673；持仓中再次出现的 BUY 归类为辅助看多信号。

• User preference: treat repeated SELL without reopen as auxiliary bearish signals, not real trades.

• User preference: BUY while already holding is auxiliary bullish signal, not add-on.

• User constraint: only reconstruct and optimize for index 399673, no ETF extension.

• User preference: robustness first, reduce useless trades second, preserve large-trend profit third, annualized return last.

• User preference: wants detailed design explained before major quant refactors, then direct execution.

• User preference: residual pruning should preserve full real-trade alignment first; profitable alpha trades and bridge trades should not be removed without downstream replacement proof.

• User approved execution order: finish Stage 2 rule-layer reconstruction first, then start Stage 3 robustness validation.

• User preference: after plan approval, continue working autonomously unless an important adjustment or clarification is needed.

• Historical operating assumption: alpha_first_selective_veto was kept as the formal alpha-first branch before the later stronger alpha-first direction was adopted.

• Latest direction signal: user supports the stronger alpha-first viewpoint and wants the next research plan oriented toward formalizing the better-performing branch rather than continuing workbook-style micro-alignment.

• Latest execution instruction: continue autonomously without small confirmations until the agreed stage target is reached.

• Current branch usage preference: use alpha_first_glued_refined_hot_cap as the forward research default, keep alpha_first_selective_veto as the benchmark control, and keep workbook_preserving only as the reconstruction reference.

• User preference: before switching Codex CLI/session surfaces, record important completed changes explicitly in workspace memory so follow-up work can resume without losing context.

• User preference: memory files in this workspace must also be committed to git, otherwise important context may be lost across sessions.

• User instruction: useful code changes can be committed directly without asking again.

• User preference: when requesting external consultation/review, prepare a compact zip bundle with issue summary, core source, parameter snapshots, and key evidence reports.

• User preference: live signal checks should support intraday realtime evaluation, using the current market price as the provisional close for today's bar when the official daily bar is not yet available.

• User relies on the assistant for directional judgment when a live signal looks discretionary or when it is unclear whether a veto is filtering alpha or protecting quality; answers should be explicit rather than hedged.

• User preference: strategy-governance and optimization execution should be advanced in OpenSpec form (openspec/changes/...) for traceable proposal/design/tasks/spec workflow.

• User preference: when requesting bulk execution ("全部执行完，再找我"), do not send intermediate progress updates and report only after full completion.

• User priority clarification: the core optimization target is to make money by following trends, especially to recover quickly after false vetoes and rejoin valid trends in time.

• User preference: future strategy optimization should prioritize false-veto -> timely trend reentry -> trend profit capture ahead of static veto purity.

• User preference: avoid turning the work into a large research program; keep changes small-team sized, pragmatic, and focused on a minimal regime-driven profit loop.

• User instruction: set false-veto recovery as the current first priority and continue development autonomously without mid-task interruptions.

• Current research direction after lean Phase 2 review: keep ql_rebound_weak_followthrough and high_zone_weak_b1 in hard-block territory for now; if further optimization is attempted, focus narrowly on mid_zone_very_weak_b1 execution timing and entry-specific exit treatment.

• Current shadow-tracking direction after the next experiment: keep the new alpha_first_glued_followthrough_mid_exit_probe branch under observation as the only promising followthrough-exit variant; do not promote it yet because the evidence is still one repaired-trend path.

• Latest operational state as of 2026-06-04: default branch alpha_first_glued_refined_hot_cap remains active and in position from 2026-05-06 dual_gold_resonance_buy; no latest-bar event fired; rollout decision FORWARD_OK.