Rank 107 · MTF CHOP charged-up count clean replication

硬结论：park / evidence pool

charged_count>=2 的 veto 确实让整体样本少做了一部分高噪声 long：baseline mean_net_ret 约 -15.42 bps，hard_veto 约 -15.57 bps；baseline fail_below_ema20_4bars 约 29.41%，hard_veto 约 28.76%。但它仍没有把总体样本拉到足够硬的 candidate 级别，而且 retention 只剩 96.39%；size_down 也主要是少亏，不是翻盘。

下一步：按顶板顺序把 Rank 107 收口为 park / evidence pool，并切 prebreak higher-low pressure ladder context gate 的 source intake。

这轮到底在测什么

• 只测它能不能当 long-side veto / size-down layer，不把它包装成独立 alpha。
• 底层 long proxy 只含两类：retest_hold_long 与 ema_continuation_long。
• 三臂固定为：baseline / charged_count>=2 hard veto / charged_count>=2 size-down(0.5x)。
• MTF CHOP 统一按 15m / 30m / 60m 三层、CHOP(14)、lookahead_off 对齐到当前 bar。

主表

variant	events	trade_count_retention	charged_ge2_share	mean_net_ret_6bps	median_net_ret_6bps	win_rate_6bps	fail_below_ema20_4bars	left_tail_p5_6bps	avg_position_size
baseline	1248	100.00%	4.57%	-15.42 bps	-15.32 bps	35.74%	29.41%	-126.89 bps	1.00
hard_veto	1203	96.39%	0.00%	-15.57 bps	-15.67 bps	35.58%	28.76%	-126.90 bps	1.00
size_down	1248	100.00%	4.57%	-15.41 bps	-15.26 bps	35.34%	29.41%	-122.77 bps	0.98

按 setup 拆开

variant	setup	events	trade_count_retention_vs_setup_baseline	charged_ge2_share	mean_net_ret_6bps	median_net_ret_6bps	win_rate_6bps	fail_below_ema20_4bars	left_tail_p5_6bps
baseline	ema_continuation_long	1248	100.00%	4.57%	-15.42 bps	-15.32 bps	35.74%	29.41%	-126.89 bps
hard_veto	ema_continuation_long	1203	96.39%	0.00%	-15.57 bps	-15.67 bps	35.58%	28.76%	-126.90 bps
size_down	ema_continuation_long	1248	100.00%	4.57%	-15.41 bps	-15.26 bps	35.34%	29.41%	-122.77 bps

按 charged bucket 拆开

variant	charged_bucket	events	mean_net_ret_6bps	median_net_ret_6bps	win_rate_6bps	fail_below_ema20_4bars	left_tail_p5_6bps
baseline	charged<2	1191	-15.55 bps	-15.63 bps	35.77%	28.80%	-126.89 bps
baseline	charged>=2	57	-12.69 bps	-13.05 bps	35.09%	42.11%	-123.42 bps
hard_veto	charged<2	1203	-15.57 bps	-15.67 bps	35.58%	28.76%	-126.90 bps
size_down	charged<2	1191	-15.55 bps	-15.63 bps	35.77%	28.80%	-126.89 bps
size_down	charged>=2	57	-12.34 bps	-12.52 bps	26.32%	42.11%	-67.71 bps

按币种拆开

variant	symbol	events	mean_net_ret_6bps	win_rate_6bps	fail_below_ema20_4bars	left_tail_p5_6bps
baseline	BTCUSDT	425	-17.25 bps	32.94%	32.00%	-105.57 bps
baseline	ETHUSDT	426	-10.59 bps	36.15%	28.40%	-118.16 bps
baseline	SOLUSDT	397	-18.66 bps	38.29%	27.71%	-155.52 bps
hard_veto	BTCUSDT	413	-17.73 bps	32.45%	32.20%	-105.25 bps
hard_veto	ETHUSDT	412	-10.63 bps	35.68%	27.43%	-119.39 bps
hard_veto	SOLUSDT	378	-18.59 bps	38.89%	26.46%	-159.62 bps
size_down	BTCUSDT	425	-17.37 bps	32.47%	32.00%	-105.57 bps
size_down	ETHUSDT	426	-10.80 bps	35.92%	28.40%	-118.16 bps
size_down	SOLUSDT	397	-18.24 bps	37.78%	27.71%	-155.52 bps

reader-facing 读法

• charged_count>=2 更像“少做高噪声 long”的 veto，而不是 shared admission gate。
• hard_veto 与 size_down 如果只是把亏损削浅、却没有把整体样本拉成更硬的 queue-facing candidate，就不该继续占住 Scout 主资源位。
• 因此这轮更诚实的 desk 读法是：保留“多周期一起变糊时别硬做 long”这条经验，但把 Rank 107 本体压回 park / evidence pool。

产物

• reports/artifacts/scout_rank107_mtf_chop_chargedup_15m/event_log.csv
• reports/artifacts/scout_rank107_mtf_chop_chargedup_15m/overall_summary.csv
• reports/artifacts/scout_rank107_mtf_chop_chargedup_15m/setup_summary.csv
• reports/artifacts/scout_rank107_mtf_chop_chargedup_15m/charged_bucket_summary.csv
• reports/artifacts/scout_rank107_mtf_chop_chargedup_15m/symbol_summary.csv
• reports/artifacts/scout_rank107_mtf_chop_chargedup_15m/verdict_summary.csv
• reports/artifacts/scout_rank107_mtf_chop_chargedup_15m/summary_snapshot.json