cyb50-quant/research/dragon/v2/memory/2026-04-05.md

2026-04-05

• Started the quasi-production phase after the formal research and governance pack was completed on 2026-04-04.
• The next practical objective shifted from further historical tuning to:
• freeze the current best branch as RC1
• generate daily signals on live-updated data

• generate a monitoring snapshot against formal thresholds

• Added dragon_rc1_release.py.

• Generated:

• dragon_rc1_release.md

• dragon_rc1_config_snapshot.json

• RC1 was frozen from branch alpha_first_glued_refined_hot_cap.

• Frozen headline:

• trades 91

• win_rate 52.75%

• avg_return 3.42%

• profit_factor 5.11

• compounded 1424.12%

• CAGR 31.31%

• max_drawdown -12.79%

• Added dragon_daily_signal_pipeline.py.

• Generated:

• dragon_daily_signal_snapshot.csv

• dragon_daily_branch_status.csv

• dragon_daily_monitor_snapshot.csv

• dragon_daily_rc1_manifest.json

• dragon_daily_signal_report.md

• plus dated archive copies under daily_reports/

• Live data fetch was validated through akshare -> data_fetcher_v2 -> dragon_indicators.

• Request date was 2026-04-05, but the actual latest available market bar was correctly identified as 2026-04-03.

• This avoids treating the weekend as a trading day.

• Daily branch status on latest available bar 2026-04-03:

• refined branch alpha_first_glued_refined_hot_cap: flat

• control branch alpha_first_selective_veto: flat

• latest real event for both branches: 2026-02-13 SELL knife_take_profit_2_glued

• latest bar values:

• close 3317.184

• a1 -0.0157

• b1 -0.1435

• c1 47.26

• latest bar had no new real or aux events for either branch

• Monitoring snapshot on 2026-04-03 after fixing the drawdown-threshold direction:

• warnings 0

• hard breaches 0

• next_open avg_return delta vs control +0.53%

• next_open profit_factor delta vs control +0.92

• next_open max_drawdown refined -13.19%

• next_open max loss streak refined 8

• next_open + 20 bps/side CAGR refined/control 25.17% / 22.07%

• A small implementation bug was found and fixed:

• dragon_strategy_monitoring_template.csv had the drawdown threshold direction wrong for negative drawdown values

• it was corrected from >= to <=

• after rerun, monitoring status became fully ok

• Practical workspace status after this step:

• the best branch is now frozen as RC1

• daily signal production is runnable

• daily monitoring is runnable

• the project is ready to move into forward observation / daily tracking rather than more in-sample tuning

• Continued immediately into the forward-observation layer.

• Patched dragon_daily_signal_pipeline.py so as_of_request_date is dynamic rather than hard-coded.

• Added dragon_forward_observation_pipeline.py.

• Generated:

• dragon_forward_observation_log.csv

• dragon_forward_observation_state.json

• dragon_signal_change_log.csv

• dragon_signal_change_review.md

• dragon_branch_divergence_log.csv

• dragon_branch_divergence_report.md

• dragon_monitor_history.csv

• dragon_monitor_health_report.md

• dragon_forward_weekly_summary.csv

• dragon_forward_weekly_review.md

• plus dated CSV archives under forward_reports/

• First forward-observation snapshot on latest bar 2026-04-03:

• both refined and control branches remained flat

• same latest real event: 2026-02-13|SELL|knife_take_profit_2_glued

• divergence level none

• signal change log only contains initial branch_initialized records for both branches

• monitor history now has the first dated row-set with all metrics in ok

• Practical implication:

• the project now has:

• frozen RC1

• daily signal generation

• daily monitor generation

• cumulative forward observation log

• branch divergence tracking

• monitor history with streak counters

• weekly forward summary scaffold

• next work should focus on accumulating real forward observations rather than changing the strategy core

• Added formal HTML reporting pages on top of the forward-observation layer.

• dragon_daily_signal_pipeline.py now emits three-strategy daily status including:

• workbook_preserving

• alpha_first_selective_veto

• alpha_first_glued_refined_hot_cap

• dragon_forward_weekly_summary.csv now also includes all three strategies.

• Rebuilt dragon_html_reports.py to generate:

• dragon_reports_index.html

• dragon_daily_signal_report.html

• dragon_forward_weekly_review.html

• dated archive copies under html_reports/

• html_reports/index.html

• Verified the HTML files are proper UTF-8 text; prior Chinese mojibake came from PowerShell display, not file corruption.

• Fixed HTML internal-link bug:

• root pages and archived pages now use different relative link maps

• links inside html_reports/ no longer incorrectly point to nested html_reports/html_reports/...

• Added next-step presentation upgrade:

• index page now includes headline RC1 summary cards

• index page now includes inline equity and yearly-return charts

• added one-click refresh script update_dragon_reports.ps1

• Added HTML versions of the Chinese usage docs:

• dragon_html_report_usage_cn.html

• dragon_html_report_quickstart_cn.html

• these are now linked from the main HTML index page and from the shared top navigation

• archived copies are also generated into html_reports/

• Added historical full-trade detail outputs for all three strategies:

• dragon_historical_trade_details.csv

• dragon_historical_trade_details.html

• dated archive CSV in daily_reports/

• dated archive HTML in html_reports/

• Detail fields now include:

• buy/sell date

• buy/sell price

• buy/sell reason

• holding days

• return_pct

• capital_before

• pnl_amount

• capital_after

• Daily report and index page now both link directly to the historical detail page for audit use.

• Added investor-facing indicator/strategy theory guide:

• source markdown: dragon_indicator_strategy_guide_cn.md

• html page: dragon_indicator_strategy_guide_cn.html

• archived html copy in html_reports/

• Upgraded the investor-facing guide page into a fuller illustrated explainer.

• Added sections:

• 指标速查卡

• 按市场状态来理解指标

• 详细指标值与触发举例

• 常见真实买卖规则速查表

• 三版本当前统计对照

• 普通投资者最容易误解的 4 件事

• 看完原理后，下一步看哪里

• Verified that both dragon_indicator_strategy_guide_cn.html and html_reports/dragon_indicator_strategy_guide_cn.html contain the new sections and still include the 3 SVG diagrams.

• The guide now also includes RC1-oriented rule tables with:

• representative real BUY / SELL rules

• typical trigger ranges / event conditions

• historical occurrence counts from dragon_historical_trade_details.csv

• Added cross-page drilldown:

• rule names in dragon_indicator_strategy_guide_cn.html now link directly to dragon_historical_trade_details*.html

• links carry URL params for branch=alpha_first_glued_refined_hot_cap plus rule keyword

• historical detail page now reads branch/year/keyword from query parameters and auto-applies filters on load

• historical detail page also now shows a visible source banner when opened from a filtered deep link

• added a 清空筛选，查看完整明细 link that resets back to the plain page path

• historical detail page was further upgraded with expandable per-trade detail rows

• each trade now has:

• 详情 / 收起 toggle

• verdict, point-difference, equivalent units, holding bucket

• quick links for same buy rule / same sell rule / same strategy-year / full detail

• Added adjacent-trade navigation inside expanded detail rows:

• 上一笔同策略交易

• 下一笔同策略交易

• clicking adjacent-trade links expands the target row and scrolls to it

• linked from index, daily report, and the shared top navigation as 指标原理

• Historical detail HTML was upgraded again for audit depth.

• Each expanded trade row now includes:

• 买入日前后 10 个交易日

• 卖出日前后 10 个交易日

• Snapshot fields:

• date

• marker (买入日 / 卖出日)

• close

• a1

• b1

• c1

• KDJ

• QL

• Snapshots are rendered on demand when a row is expanded, not pre-expanded for all trades.

• dragon_html_reports.py now auto-loads indicator coverage for this page:

• first tries local snapshot coverage

• if coverage is insufficient, fetches and recomputes full history through DragonIndicatorEngine

• writes dragon_indicator_snapshot_full.csv

• uses an in-process cache so root + archived historical detail pages only fetch once per rebuild

• Verified final embedded indicator coverage now reaches 2026-04-03.

• This fixes the earlier detail-page gap where embedded indicator rows stopped at 2026-01-30.

• Historical detail HTML was upgraded one more step from indicator-only snapshots to audit snapshots.

• Each snapshot card now also shows:

• rule summary in the card header

• same-window strategy events for the current branch

• event tags distinguish:

• 真实BUY

• 真实SELL

• 辅助BUY

• 辅助SELL

• dragon_html_reports.py now rebuilds branch event payloads for:

• workbook_preserving

• alpha_first_selective_veto

• alpha_first_glued_refined_hot_cap

• these event payloads are rendered inside the snapshot mini-table as a new 策略事件 column

• Historical detail snapshots were upgraded again with a window-level event summary strip.

• In each buy/sell snapshot card, dragon_html_reports.py now computes from embedded eventMap:

• total event count in the +/-10 trading-day window

• counts for 真实BUY / 真实SELL / 辅助BUY / 辅助SELL

• nearest event before the marker day

• same-day event summary

• nearest event after the marker day

• The summary is rendered as .snapshot-summary under the mini-table and above the indicator note.

• Rebuild verification completed successfully:

• py -3 -m py_compile dragon_html_reports.py

• py -3 dragon_html_reports.py

• verified root and archived historical detail HTML both contain the new summary script/output markers:

• snapshot-summary

• 总事件

• 前一条 / 当日 / 后一条

• Handoff note for future Codex CLI/session switch:

• latest completed UI/reporting change is the historical-detail snapshot event-summary strip in dragon_html_reports.py

• generated outputs already refreshed:

• dragon_historical_trade_details.html

• html_reports/dragon_historical_trade_details_2026-04-03.html

• status:

• implemented

• rebuilt

• verified

• scope remained reporting-only; no strategy-core or branch-logic changes were made in this step

• A formal code review was then run on the current productionized chain:

• dragon_strategy.py

• dragon_daily_signal_pipeline.py

• dragon_forward_observation_pipeline.py

• dragon_refined_execution_validation.py

• dragon_html_reports.py

• Main findings fixed immediately:

• daily monitor no longer hard-codes removed_trade_over_removal_count=0 and local_sensitivity_robust_case_count=52

• these two metrics are now loaded dynamically from:

• dragon_glued_refined_removed_trade_attribution.csv

• dragon_glued_refined_sensitivity.csv

• next-bar execution models no longer silently fall back to same-bar close when the next bar is missing

• missing next-bar prices are now set to NaN, then excluded inside _apply_execution_model(...)

• _summary(...) and _risk_cluster(...) were hardened for empty post-filter trade sets

• weekly summary no longer repeats global warning/divergence counts inside every branch row

• dragon_forward_weekly_summary.csv now keeps branch rows branch-local and adds a separate system_monitor row for global monitor/divergence counts

• Verification rerun completed:

• py -3 -m py_compile dragon_daily_signal_pipeline.py dragon_refined_execution_validation.py dragon_forward_observation_pipeline.py dragon_html_reports.py

• py -3 dragon_daily_signal_pipeline.py

• py -3 dragon_forward_observation_pipeline.py

• py -3 dragon_refined_execution_validation.py

• External review findings were then fixed:

• dragon_indicators.py no longer depends on hard-coded workstation path D:\\work\\project\\cyb50-quant\\dragon

• it now resolves repo root from __file__ and imports dependencies from <repo>/dragon

• dragon_alpha_first_baseline.py now trims trades by both:

• buy_date

• sell_date

• against workbook event window, so post-window exits are excluded from branch summary stats

• also aligned alpha-vs-workbook trade diff to the same filtered workbook trade set

• verified after rerun:

• py -3 -m py_compile dragon_indicators.py dragon_alpha_first_baseline.py

• py -3 dragon_alpha_first_baseline.py

• py -3 dragon_daily_signal_pipeline.py

• confirmed:

• dragon_alpha_first_branch_trade_diff.csv no longer contains 2026-01-05 -> 2026-01-20

• dragon_alpha_first_baseline.md now reports removed trades 4 instead of the stale 5

• A broader workbook-window consistency sweep was then completed across evaluation/research scripts.

• Rule used for this sweep:

• if a script is generating in-sample / workbook-window research statistics, trade filtering must constrain both:

• buy_date

• sell_date

• not just buy_date

• Files updated in this sweep:

• dragon_backtest.py

• dragon_cost_stress_test.py

• dragon_deep_oversold_confirmation_experiments.py

• dragon_deep_oversold_experiments.py

• dragon_deep_oversold_selective_veto_experiments.py

• dragon_equity_curve_review.py

• dragon_glued_alpha_candidate.py

• dragon_glued_refined_branch_review.py

• dragon_glued_refined_removed_trade_attribution.py

• dragon_glued_refined_sensitivity.py

• dragon_glued_refine_experiments.py

• dragon_predictive_break_experiments.py

• dragon_rc1_release.py

• dragon_refined_alpha_attribution.py

• dragon_refined_execution_validation.py

• dragon_rule_ablation.py

• dragon_short_holding_audit.py

• dragon_short_holding_experiments.py

• dragon_strategy_overview.py

• dragon_threshold_perturbation.py

• Explicit non-sweep exception:

• dragon_daily_signal_pipeline.py remained live-oriented and was not upper-bounded by workbook end date on purpose

• Regression verification completed:

• full py_compile passed for all updated scripts

• representative reruns passed for:

• dragon_backtest.py

• dragon_strategy_overview.py

• dragon_rc1_release.py

• dragon_refined_execution_validation.py

• dragon_cost_stress_test.py

• dragon_rule_ablation.py

• dragon_threshold_perturbation.py

• dragon_glued_refined_branch_review.py

• dragon_glued_refined_sensitivity.py

• dragon_glued_alpha_candidate.py

• dragon_short_holding_audit.py

• System review closure completed later on 2026-04-05.

• Root cause of the remaining review warnings was confirmed to be metric-window inconsistency, not strategy-core divergence.

• dragon_alpha_first_baseline.py was further aligned to the same bounded release window used by the main summary pack:

• 2016-01-01 to 2025-12-31

• This removed the stale real-BUY overlap drift caused by the workbook terminal 2026-01-05 BUY event being counted in only one report family.

• CAGR-year consistency was also unified:

• dragon_rc1_release.py no longer hard-codes 10.0 years

• dragon_refined_execution_validation.py no longer hard-codes 10.0 years

• both now derive evaluation years from the release window dates, matching dragon_strategy_overview.py

• Important operational lesson:

• the first rerun of dragon_system_review.py was launched in parallel with upstream generators, so it briefly read stale outputs

• rerunning dragon_system_review.py after upstream artifacts finished removed the false residual mismatches

• Final verified review state after the sequential rerun:

• dragon_review_branch_metric_consistency.md -> mismatches 0

• dragon_alpha_first_branch_summary.csv -> workbook 105/105, alpha 101/101

• dragon_strategy_overview.csv -> fully aligned with branch summaries

• dragon_rc1_config_snapshot.json CAGR -> aligned with dragon_strategy_overview.csv

• dragon_system_review_final.md now classifies dragon_strategy_overview.csv as directly trustable

• Remaining caution is now report interpretation / forward monitoring, not internal metric-definition drift.

• A second-pass governance/reporting review was completed after system-review closure.

• Confirmed and fixed two real governance/reporting issues:

• dragon_alpha_branch_governance.py

• fixed alignment_cost_gate so it only blocks on overlap loss worse than -8; positive overlap improvements no longer fail the gate by mistake

• fixed removal_quality_gate so an empty removed-trade set no longer fails promotion automatically

• fixed the final ## Recommendation section so it now matches final_decision instead of always biasing toward “keep current alpha”

• dragon_html_reports.py

• normalized system_monitor display labels from internal raw names to user-facing Chinese labels

• removed the ad-hoc post-dict assignment for BRANCH_ROLES["system_monitor"]

• verification rerun completed:

• py -3 -m py_compile dragon_alpha_branch_governance.py dragon_html_reports.py

• py -3 dragon_alpha_branch_governance.py

• py -3 dragon_html_reports.py

• current reviewed judgment:

• no new high-severity strategy-core issue found in this second pass

• remaining items are mainly governance-threshold choices and presentation polish, not signal correctness defects

• The highest-priority remaining engineering risk was then fixed:

• monitor-chain missing values no longer pass silently

• dragon_daily_signal_pipeline.py

• _compare_numeric(...) now returns missing_data for NaN

• monitor rows no longer get blank status when actual_value is missing

• daily monitor summary now reports:

• warning_count

• hard_breach_count where missing_data is treated as hard-breach-equivalent

• missing_data_count

• dragon_forward_observation_pipeline.py

• monitor summary now carries missing_data_count and includes missing_data inside hard_breach_count

• monitor change log now tracks monitor_missing_data_count_changed

• monitor history hard_breach_streak now also counts missing_data

• observation log and divergence log now persist monitor_missing_data_count / missing_data_count

• dragon_html_reports.py

• added missing_data badge color

• daily/weekly HTML now display 缺失指标数

• weekly HTML is backward-compatible with older logs that do not yet contain the new column

• verification completed:

• py -3 -m py_compile dragon_daily_signal_pipeline.py dragon_forward_observation_pipeline.py dragon_html_reports.py dragon_system_review.py

• py -3 dragon_daily_signal_pipeline.py

• py -3 dragon_forward_observation_pipeline.py

• py -3 dragon_html_reports.py

• py -3 dragon_system_review.py

• direct function-level verification:

• daily._compare_numeric(float('nan'), ...) -> 'missing_data'

• forward monitor summary treats missing_data as part of hard-breach-equivalent count

• P2 single-source refactor was then executed on the main reporting/evaluation chain.

• Added shared modules:

• dragon_shared.py

• centralizes START_DATE, END_DATE, evaluation_years(...), profit_factor(...), format_pct(...), format_num(...)

• dragon_execution_common.py

• centralizes execution-aware apply_execution_model(...), summary(...), risk_cluster(...)

• Main-chain scripts migrated to shared modules:

• dragon_strategy_overview.py

• dragon_rc1_release.py

• dragon_refined_execution_validation.py

• dragon_daily_signal_pipeline.py

• dragon_cost_stress_test.py

• dragon_equity_curve_review.py

• Also migrated core branch-summary family scripts:

• dragon_alpha_first_baseline.py

• dragon_glued_refined_branch_review.py

• dragon_glued_refined_sensitivity.py

• Important structural improvement:

• daily/RC1 scripts no longer import formatting helpers from dragon_strategy_overview.py

• daily signal pipeline no longer imports execution helpers from dragon_refined_execution_validation.py

• instead both use dedicated shared modules

• Verification reruns completed:

• py -3 -m py_compile passed for shared modules and migrated scripts

• reruns passed for:

• dragon_strategy_overview.py

• dragon_rc1_release.py

• dragon_refined_execution_validation.py

• dragon_daily_signal_pipeline.py

• dragon_cost_stress_test.py

• dragon_equity_curve_review.py

• dragon_alpha_first_baseline.py

• dragon_glued_refined_branch_review.py

• dragon_glued_refined_sensitivity.py

• dragon_system_review.py

• Post-refactor consistency check:

• dragon_review_branch_metric_consistency.md still shows Mismatches: 0

• dragon_strategy_overview.csv, dragon_alpha_first_branch_summary.csv, dragon_glued_refined_branch_summary.csv, and dragon_rc1_config_snapshot.json remain aligned

• Session persistence note:

• the useful dragon/v2 code changes were already committed as git commit 9f7ce16 (Add dragon v2 strategy research and reporting code)

• user then explicitly required that workspace memory files must also be committed, otherwise future sessions may lose context

• immediate next git action should be a separate memory-only commit covering MEMORY.md, USER.md, and the daily memory files