Pong RAM 策略实验

轮次: 49 · 局数: 1773 · 失败记录: 37 · 最新轮: 49

全部轮次

轮次	时间	策略	seed	局数	均分	最低	最高	步数	改动	建议
49	2026-05-10T00:05:23.369804+00:00	ram_intercept_v3_28_chain2_score00_initial_up	0	80	19.025	18.0	20.0	466333	Round 66 accepted marker-level chain2 score0-0 initial DOWN->UP guard after 80-seed counterfactual mean 19.025 min 18 max 20; expected seeds 2 and 15 improve 18->20 with no regressions.	和上一轮对比，每次只改一个 heuristic。
48	2026-05-09T21:16:24.260567+00:00	ram_intercept_v3_27_high_down_score160_bounce_noop	0	80	18.975	18.0	20.0	466965	Round 55: solidify score 16-0 high-down y165 bounce DOWN->NOOP after v3_26 score18 floor trace and full 80-seed counterfactual guard; v3_26 mean 18.70 min 18 max 20 to probe mean 18.98 min 18 max 20	和上一轮对比，每次只改一个 heuristic。
47	2026-05-09T20:56:51.415833+00:00	ram_intercept_v3_26_low_up_score10_final_noop	0	80	18.7	18.0	20.0	471695	Round 54: solidify score 1-0 low-up final DOWN->NOOP after v3_25 floor trace and full 80-seed counterfactual guard; v3_25 mean 18.2875 min 17 max 19 to probe mean 18.70 min 18 max 20	和上一轮对比，每次只改一个 heuristic。
46	2026-05-09T20:33:24.386269+00:00	ram_intercept_v3_25_y86_score140_serve_initial_noop	0	80	18.2875	17.0	19.0	713441	Round 53: solidify score 14-0 y86 serve-initial DOWN->NOOP after setup diagnostic and full 80-seed counterfactual guard; v3_24 mean 18.0125 min 17 max 19 to probe mean 18.29 min 17 max 19	和上一轮对比，每次只改一个 heuristic。
45	2026-05-09T20:11:08.785023+00:00	ram_intercept_v3_24_score00_mid_down_y154_noop	0	80	18.0125	17.0	19.0	712891	Round 51: solidify score 0-0 mid-down y154 final UP->NOOP after mixed seed2/15 trace and full 80-seed counterfactual guard; v3_23 mean 17.9625 min 16 max 19 to probe mean 18.01 min 17 max 19	和上一轮对比，每次只改一个 heuristic。
44	2026-05-09T19:53:06.444230+00:00	ram_intercept_v3_23_high_down_score120_bounce_noop	0	80	17.9625	16.0	19.0	713027	Round 50: solidify score 12-0 high-down bounce x144-148 y50-56 dy9 DOWN->NOOP after score16 floor trace and full 80-seed counterfactual guard; v3_22 mean 17.6875 min 16 max 19 to probe mean 17.96 min 16 max 19	和上一轮对比，每次只改一个 heuristic。
43	2026-05-09T19:33:24.229013+00:00	ram_intercept_v3_22_y86_score100_serve_initial_noop	0	80	17.6875	16.0	19.0	720397	Round 49: solidify score 10-0 y86 serve-initial DOWN->NOOP after setup diagnostic and full 80-seed counterfactual guard; v3_21 mean 17.4125 min 15 max 19 to probe mean 17.69 min 16 max 19	和上一轮对比，每次只改一个 heuristic。
42	2026-05-09T19:11:10.174793+00:00	ram_intercept_v3_21_high_down_score80_bounce_noop	0	80	17.4125	15.0	19.0	717207	Round 47: solidify score 8-0 high-down bounce x144-148 y50-56 dy9 DOWN->NOOP after full 80-seed counterfactual guard; v3_20 mean 17.1375 min 14 max 19 to probe mean 17.41 min 15 max 19	和上一轮对比，每次只改一个 heuristic。
41	2026-05-09T18:51:51.602778+00:00	ram_intercept_v3_20_y86_score60_serve_initial_noop	0	80	17.1375	14.0	19.0	727217	Round 45: solidify score 6-0 y86 serve-initial DOWN->NOOP after full 80-seed counterfactual guard; v3_19 mean 16.8625 min 13 max 19 to probe mean 17.14 min 14 max 19	和上一轮对比，每次只改一个 heuristic。
40	2026-05-09T17:46:03.441851+00:00	ram_intercept_v3_19_video_check_seed1	1	1	19.0	19.0	19.0	5262	video check for current retained v3_19 policy on seed 1; avoids known seed 0/9 inactive-tail loop	分数无波动；每次只改一个启发式，再多 seed 测。
39	2026-05-09T16:42:02.651397+00:00	ram_intercept_v3_19_video_check	0	1	17.0	17.0	17.0	27000	video check for current retained v3_19 policy; one recorded episode for report.html viewing	分数无波动；每次只改一个启发式，再多 seed 测。
38	2026-05-09T15:18:29.657105+00:00	ram_intercept_v3_19_y86_after_score22_bounce_up	0	80	16.8625	13.0	19.0	721387	Round 39: solidify stateful y86 x166 bounce DOWN->UP only after score 2-2 family marker; v3_18 mean 14.9125 min 13 max 19 to probe mean 16.86 min 13 max 19	和上一轮对比，每次只改一个 heuristic。
37	2026-05-09T14:20:12.603639+00:00	ram_intercept_v3_18_high_down_score40_bounce_noop	0	80	14.9125	13.0	19.0	763117	Round 37: solidify score 4-0 high-down bounce x144-148 y50-56 dy9 DOWN->NOOP after history diagnostic; v3_17 mean 14.6375 min 12 max 19 to probe mean 14.91 min 13 max 19	和上一轮对比，每次只改一个 heuristic。
36	2026-05-09T13:37:32.626697+00:00	ram_intercept_v3_17_y86_score82_early_setup_down	0	80	14.6375	12.0	19.0	769757	Round 34: solidify score 8-2 y86 offset8 UP->DOWN correction only during steps 2050..2110 after full guard; v3_16 mean 14.3125 min 12 max 19 to probe mean 14.64 min 12 max 19	和上一轮对比，每次只改一个 heuristic。
35	2026-05-09T12:05:35.206021+00:00	ram_intercept_v3_16_y86_score20_serve_initial_up	0	80	14.3125	12.0	19.0	778545	Round 27: solidify score 2-0 y86 serve-initial DOWN->UP correction after full guard; v3_15 mean 14.1625 min 11 max 19 to probe mean 14.31 min 12 max 19	和上一轮对比，每次只改一个 heuristic。
34	2026-05-09T11:19:26.190173+00:00	ram_intercept_v3_15_y86_score62_early_setup_down	0	80	14.1625	11.0	19.0	773925	Round 24: solidify step-gated y86 6-2 setup UP->DOWN correction after full guard; v3_14 mean 13.8375 min 11 max 19 to probe mean 14.16 min 11 max 19	和上一轮对比，每次只改一个 heuristic。
33	2026-05-09T10:51:55.791188+00:00	ram_intercept_v3_14_y86_score42_setup_down	0	80	13.8375	11.0	19.0	782713	Round 22: solidify one-shot y86 4-2 setup UP->DOWN correction after 80-seed counterfactual; v3_13 mean 13.5125 min 10 max 19 to probe mean 13.84 min 11 max 19	和上一轮对比，每次只改一个 heuristic。
32	2026-05-09T10:35:06.749627+00:00	ram_intercept_v3_13_y86_score22_setup_down	0	80	13.5125	10.0	19.0	791501	Round 21: solidify one-shot y86 2-2 setup UP->DOWN correction after 80-seed score-gated counterfactual; v3_12 mean 13.1875 min 9 max 19 to probe mean 13.51 min 10 max 19	和上一轮对比，每次只改一个 heuristic。
31	2026-05-09T09:48:19.656719+00:00	ram_intercept_v3_12_low_up_y111_down	0	80	13.1875	9.0	19.0	800289	Round 19: solidify low-up y111 precontact UP->DOWN correction after 80-seed counterfactual guard; v3_11 mean 12.75 min 9 max 19 to probe mean 13.19 min 9 max 19	和上一轮对比，每次只改一个 heuristic。
30	2026-05-09T09:24:48.512264+00:00	ram_intercept_v3_11_low_mid_score00_noop	0	80	12.75	9.0	19.0	792374	Round 18: solidify seed44/66 low-mid score00 DOWN->NOOP correction after 80-seed counterfactual guard	和上一轮对比，每次只改一个 heuristic。
29	2026-05-09T09:02:56.491867+00:00	ram_intercept_v3_10_low_up_score01_noop_seed60_79	60	20	13.8	-1.0	19.0	241136	expanded validation for retained v3_10 policy over seeds 60..79 after observed seeds 0..59 mean about 11.8 min -1 max 19; look for new collapse before further tuning	和上一轮对比，每次只改一个 heuristic。
28	2026-05-09T09:00:07.834203+00:00	ram_intercept_v3_10_low_up_score01_noop_seed40_59	40	20	12.1	-1.0	18.0	247890	expanded validation for retained v3_10 policy over seeds 40..59 after 0..39 mean 11.65 min -1 max 19; look for new collapse before further tuning	和上一轮对比，每次只改一个 heuristic。
27	2026-05-09T08:56:35.179654+00:00	ram_intercept_v3_10_low_up_score01_noop	0	40	11.65	-1.0	19.0	372732	expanded-set low-up score01 correction: keep v3_9 and add one-shot 0-1 score dy=-8 x=176..178 y=172..176 paddle=132..138 target=135..150 baseline DOWN to NOOP; 40-seed counterfactual improved min -21 to -1 with only seed33 changed	和上一轮对比，每次只改一个 heuristic。
26	2026-05-09T08:44:11.896065+00:00	ram_intercept_v3_9_late_low_loop_noop_limited_seed20_39	20	20	9.4	-21.0	19.0	171431	expanded validation for retained v3_9 policy over seeds 20..39; check for overfit or new collapse after 0..19 mean 12.9 min 9 max 19	和上一轮对比，每次只改一个 heuristic。
25	2026-05-09T08:42:11.067290+00:00	ram_intercept_v3_9_late_low_loop_noop_limited	0	20	12.9	9.0	19.0	176564	late low-loop correction capped to match validated probe: keep v3_8 and apply dy=3..4 x=169..177 y=49..57 paddle=80..105 target=60..75 UP-to-NOOP at most 64 times per episode; revalidate after uncapped formal run underperformed probe	和上一轮对比，每次只改一个 heuristic。
24	2026-05-09T08:40:31.349485+00:00	ram_intercept_v3_9_late_low_loop_noop	0	20	11.2	0.0	19.0	176564	late low-loop correction: keep v3_8 and change dy=3..4 x=169..177 y=49..57 paddle=80..105 target=60..75 baseline UP to NOOP; counterfactual 20-seed probe improved mean 9.4 to 12.9 and min -1 to 9, with max 20 to 19	和上一轮对比，每次只改一个 heuristic。
23	2026-05-09T08:32:14.021723+00:00	ram_intercept_v3_8_high_down_setup_noop	0	20	9.4	-1.0	20.0	219266	high-down setup correction: keep v3_7 and change dy=8 x=121..153 y=103..167 paddle=155..195 target=165..185 baseline NOOP to DOWN; counterfactual 20-seed probe improved mean 4.6 to 9.4 with min -1 unchanged	和上一轮对比，每次只改一个 heuristic。
22	2026-05-09T08:21:07.033577+00:00	ram_intercept_v3_7_mid_up_precontact_noop	0	20	4.6	-1.0	20.0	202825	mid-up y164 precontact phase correction: keep v3_6 and change only dy=-8 x=187..191 y=188..196 paddle=190..205 baseline UP to NOOP; counterfactual 20-seed probe improved mean 1.5 to 4.6 with min -1 unchanged	和上一轮对比，每次只改一个 heuristic。
21	2026-05-09T08:08:29.390957+00:00	ram_intercept_v3_6_mid_down_precontact_noop	0	20	1.5	-1.0	9.0	206374	mid-down precontact phase correction: keep v3_5 and change only dy=4 x=186..194 y=68..82 paddle=65..90 baseline UP to NOOP; counterfactual 20-seed probe improved mean 1.2 to 1.5 with min -1 unchanged	和上一轮对比，每次只改一个 heuristic。
20	2026-05-09T07:55:51.258328+00:00	ram_intercept_v3_5_low_serve_late_noop	0	20	1.2	-1.0	9.0	235425	low-serve late noop correction: keep v3_4 opening low-dy8 fix and change only late low-serve target=50 noop at x=189..201 y=55..67 pad=35..65 to UP; counterfactual 20-seed probe improved mean 0.2 to 1.2 and min -21 to -1	和上一轮对比，每次只改一个 heuristic。
19	2026-05-09T07:41:30.825519+00:00	ram_intercept_v3_4_opening_low_dy8	0	10	0.9	-1.0	9.0	138373	opening low dy=8 one-step correction: at 0-0 score, x=184..188 y=130..145 with paddle recovering upward, execute UP instead of baseline DOWN once; counterfactual probe improved 10-seed mean 0.5 to 0.9	和上一轮对比，每次只改一个 heuristic。
18	2026-05-09T06:46:18.399391+00:00	ram_intercept_v3_4_late_down_keep_down_probe	0	10	0.5	-1.0	9.0	137933	temporary probe: if late downward ball x=188..204 y=70..140 dy>0 and previous action was down, keep down instead of noop/up; validate action-rhythm hypothesis against v3_3 baseline mean 0.5 min -1 max 9	和上一轮对比，每次只改一个 heuristic。
17	2026-05-09T06:40:34.163055+00:00	ram_intercept_v3_4_late_low_down_y86_probe	0	10	0.5	-1.0	9.0	138299	temporary probe: for repeated x~202 y~86 dy=4 losses, add +8 target bias only at x=196..204 y=78..96 dy=3..5; validate against v3_3 baseline mean 0.5 min -1 max 9	和上一轮对比，每次只改一个 heuristic。
16	2026-05-09T06:21:02.806717+00:00	ram_intercept_v3_3_score_state_regression	0	10	0.5	-1.0	9.0	138299	state extraction update only: PongState now exposes ram[14] self_score and ram[13] opponent_score discovered from reward-event probes; verify no policy behavior change over seeds 0-9	和上一轮对比，每次只改一个 heuristic。
15	2026-05-09T06:19:02.799542+00:00	ram_intercept_v3_4_late_lock_5seed	0	5	-5.8	-21.0	-1.0	70301	hypothesis test: lock the last predicted intercept after ball passes SELF_PADDLE_X_RAM instead of chasing current y at the right edge; quick seeds 0-4 no-video max_steps=30000	先用 RGB/OBS 标注失败模式，再改 heuristic。
14	2026-05-09T06:16:49.146780+00:00	ram_intercept_v3_3_repro_baseline	0	10	0.5	-1.0	9.0	138299	repro baseline before next change: current ram_intercept_v3_3 over seeds 0-9, no-video max_steps=30000	先用 RGB/OBS 标注失败模式，再改 heuristic。
13	2026-05-09T05:57:01.222578+00:00	ram_intercept_v3_3_10seed_30k_check	0	10	0.5	-1.0	9.0	138299	no-video long robustness check for ram_intercept_v3_3 over seeds 0-9 with max_steps=30000; distinguishes true near-tie games from 10000-step truncation	先用 RGB/OBS 标注失败模式，再改 heuristic。
12	2026-05-09T05:55:36.672979+00:00	ram_intercept_v3_3_10seed_check	0	10	0.5	-2.0	9.0	84144	no-video robustness check for ram_intercept_v3_3 over seeds 0-9; scores expected around mean +0.5 with no -20 collapse	先用 RGB/OBS 标注失败模式，再改 heuristic。
11	2026-05-09T05:51:13.476925+00:00	ram_intercept_v3_3	0	5	1.4	-1.0	9.0	37646	low-serve pattern rule: after detecting reset serve near x=126..138 y=116..128 moving up-right, aim at y=50 for 24 policy steps; no-video sweep improved seeds 0-4 mean from -1.4 to +1.4	先用 RGB/OBS 标注失败模式，再改 heuristic。
10	2026-05-09T05:39:55.253489+00:00	ram_intercept_v3_2	0	5	-1.4	-20.0	9.0	34843	inactive serve-ready center changed from 121 to 100; no-video sweep over seeds 0-4 improved mean from -1.8 to -1.4 and kept target/intercept constants from v3_1	先用 RGB/OBS 标注失败模式，再改 heuristic。
9	2026-05-09T05:26:56.672212+00:00	ram_intercept_v3_1	0	5	-1.8	-20.0	9.0	23217	target clamp 46..196 added on top of ram_intercept_v3; no-video sweep over seeds 0-4 improved mean from -2.0 to -1.8	先用 RGB/OBS 标注失败模式，再改 heuristic。
8	2026-05-09T05:22:05.593598+00:00	ram_intercept_v3	0	5	-2.0	-20.0	8.0	23252	tuned intercept policy: runtime reflection bounds fixed; MAX_PLAY_Y_RAM=210 and PADDLE_DEADBAND=3 selected from no-video sweeps over seeds 0-4	先用 RGB/OBS 标注失败模式，再改 heuristic。
7	2026-05-09T05:12:41.196821+00:00	ram_intercept_v2	0	5	-14.4	-19.0	0.0	20904	stateful velocity policy: estimate dx/dy from consecutive RAM states, predict right-side intercept with wall reflection; RGB overlay calibrated from visual probe	和上一轮对比，每次只改一个 heuristic。
6	2026-05-09T05:11:03.604459+00:00	minimal_ram_xgate_v1_calibrated_overlay_diag	0	1	-21.0	-21.0	-21.0	2145	诊断样本：RGB 叠加层已按 visual_x=ram[49]-48.5、visual_y=ram[54]-12.5 校准，用于分析下一轮失败模式	仍全输；看 RGB 失败片段，并复核 ball_y / paddle_y 映射。
5	2026-05-09T05:08:25.248804+00:00	minimal_ram_xgate_v1_overlay_diag	0	1	-21.0	-21.0	-21.0	2145	诊断样本：RGB 视频叠加 RAM 推断的球、挡板、动作、reward、score，用于检查状态检测和策略失败模式	仍全输；看 RGB 失败片段，并复核 ball_y / paddle_y 映射。
4	2026-05-09T03:00:51.076953+00:00	minimal_ram_xgate_v1	0	5	-20.4	-21.0	-20.0	7486	x-gated target: center paddle unless ball is active and in right half	仍全输；看 RGB 失败片段，并复核 ball_y / paddle_y 映射。
3	2026-05-09T02:46:35.736446+00:00	minimal_ram_v0	0	1	-20.0	-20.0	-20.0	1343	dual rgb and obs video recording	仍全输；看 RGB 失败片段，并复核 ball_y / paddle_y 映射。
2	2026-05-09T02:38:15.472083+00:00	minimal_ram_v0	0	3	-20.0	-20.0	-20.0	3606	minimal ram policy with per-episode gif recording	仍全输；看 RGB 失败片段，并复核 ball_y / paddle_y 映射。
1	2026-05-09T02:33:35.856289+00:00	minimal_ram_v0	0	3	-20.0	-20.0	-20.0	3606	minimal ram policy using ram[54] ball_y and ram[51] paddle_y	仍全输；看 RGB 失败片段，并复核 ball_y / paddle_y 映射。

轮次详情

Round 49: ram_intercept_v3_28_chain2_score00_initial_up

2026-05-10T00:05:23.369804+00:00

均分 19.025

改动: Round 66 accepted marker-level chain2 score0-0 initial DOWN->UP guard after 80-seed counterfactual mean 19.025 min 18 max 20; expected seeds 2 and 15 improve 18->20 with no regressions.
样本: seed 0 / 80 局 / 466333 步
分数: 18.0 / 19.025 / 20.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	0	20.0	5023	True
2	1	19.0	5262	True
3	2	20.0	5024	True
4	3	18.0	5458	True
5	4	18.0	5454	True
6	5	19.0	7573	True
7	6	18.0	5458	True
8	7	19.0	6904	True
9	8	19.0	7573	True
10	9	20.0	5019	True
11	10	19.0	5261	True
12	11	19.0	5257	True
13	12	19.0	6905	True
14	13	19.0	5544	True
15	14	19.0	6905	True
16	15	20.0	5024	True
17	16	19.0	5261	True
18	17	19.0	6904	True
19	18	18.0	5458	True
20	19	19.0	5256	True
21	20	19.0	7573	True
22	21	19.0	5261	True
23	22	19.0	5258	True
24	23	19.0	7569	True
25	24	18.0	5460	True
26	25	20.0	5025	True
27	26	19.0	5258	True
28	27	19.0	7569	True
29	28	19.0	5263	True
30	29	19.0	5258	True
31	30	19.0	5262	True
32	31	19.0	5544	True
33	32	19.0	7568	True
34	33	18.0	5455	True
35	34	19.0	5263	True
36	35	19.0	5261	True
37	36	19.0	5262	True
38	37	19.0	6904	True
39	38	19.0	5263	True
40	39	19.0	5261	True
41	40	18.0	5459	True
42	41	20.0	5026	True
43	42	18.0	5459	True
44	43	20.0	5025	True
45	44	18.0	5455	True
46	45	19.0	6905	True
47	46	19.0	5256	True
48	47	19.0	7572	True
49	48	20.0	5026	True
50	49	18.0	5458	True
51	50	20.0	5022	True
52	51	19.0	5257	True
53	52	19.0	5258	True
54	53	19.0	5257	True
55	54	19.0	6904	True
56	55	19.0	5260	True
57	56	19.0	5258	True
58	57	19.0	6904	True
59	58	19.0	7573	True
60	59	19.0	5259	True
61	60	19.0	5260	True
62	61	19.0	7573	True
63	62	19.0	5544	True
64	63	20.0	5023	True
65	64	20.0	5025	True
66	65	19.0	5545	True
67	66	18.0	5455	True
68	67	19.0	7573	True
69	68	19.0	5544	True
70	69	19.0	5543	True
71	70	19.0	5545	True
72	71	19.0	6905	True
73	72	20.0	5020	True
74	73	19.0	5543	True
75	74	19.0	6901	True
76	75	20.0	5020	True
77	76	19.0	5263	True
78	77	19.0	7568	True
79	78	19.0	5262	True
80	79	19.0	7568	True

Round 48: ram_intercept_v3_27_high_down_score160_bounce_noop

2026-05-09T21:16:24.260567+00:00

均分 18.975

改动: Round 55: solidify score 16-0 high-down y165 bounce DOWN->NOOP after v3_26 score18 floor trace and full 80-seed counterfactual guard; v3_26 mean 18.70 min 18 max 20 to probe mean 18.98 min 18 max 20
样本: seed 0 / 80 局 / 466965 步
分数: 18.0 / 18.975 / 20.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	0	20.0	5023	True
2	1	19.0	5262	True
3	2	18.0	5340	True
4	3	18.0	5458	True
5	4	18.0	5454	True
6	5	19.0	7573	True
7	6	18.0	5458	True
8	7	19.0	6904	True
9	8	19.0	7573	True
10	9	20.0	5019	True
11	10	19.0	5261	True
12	11	19.0	5257	True
13	12	19.0	6905	True
14	13	19.0	5544	True
15	14	19.0	6905	True
16	15	18.0	5340	True
17	16	19.0	5261	True
18	17	19.0	6904	True
19	18	18.0	5458	True
20	19	19.0	5256	True
21	20	19.0	7573	True
22	21	19.0	5261	True
23	22	19.0	5258	True
24	23	19.0	7569	True
25	24	18.0	5460	True
26	25	20.0	5025	True
27	26	19.0	5258	True
28	27	19.0	7569	True
29	28	19.0	5263	True
30	29	19.0	5258	True
31	30	19.0	5262	True
32	31	19.0	5544	True
33	32	19.0	7568	True
34	33	18.0	5455	True
35	34	19.0	5263	True
36	35	19.0	5261	True
37	36	19.0	5262	True
38	37	19.0	6904	True
39	38	19.0	5263	True
40	39	19.0	5261	True
41	40	18.0	5459	True
42	41	20.0	5026	True
43	42	18.0	5459	True
44	43	20.0	5025	True
45	44	18.0	5455	True
46	45	19.0	6905	True
47	46	19.0	5256	True
48	47	19.0	7572	True
49	48	20.0	5026	True
50	49	18.0	5458	True
51	50	20.0	5022	True
52	51	19.0	5257	True
53	52	19.0	5258	True
54	53	19.0	5257	True
55	54	19.0	6904	True
56	55	19.0	5260	True
57	56	19.0	5258	True
58	57	19.0	6904	True
59	58	19.0	7573	True
60	59	19.0	5259	True
61	60	19.0	5260	True
62	61	19.0	7573	True
63	62	19.0	5544	True
64	63	20.0	5023	True
65	64	20.0	5025	True
66	65	19.0	5545	True
67	66	18.0	5455	True
68	67	19.0	7573	True
69	68	19.0	5544	True
70	69	19.0	5543	True
71	70	19.0	5545	True
72	71	19.0	6905	True
73	72	20.0	5020	True
74	73	19.0	5543	True
75	74	19.0	6901	True
76	75	20.0	5020	True
77	76	19.0	5263	True
78	77	19.0	7568	True
79	78	19.0	5262	True
80	79	19.0	7568	True

Round 47: ram_intercept_v3_26_low_up_score10_final_noop

2026-05-09T20:56:51.415833+00:00

均分 18.7

改动: Round 54: solidify score 1-0 low-up final DOWN->NOOP after v3_25 floor trace and full 80-seed counterfactual guard; v3_25 mean 18.2875 min 17 max 19 to probe mean 18.70 min 18 max 20
样本: seed 0 / 80 局 / 471695 步
分数: 18.0 / 18.7 / 20.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	0	20.0	5023	True
2	1	19.0	5262	True
3	2	18.0	5340	True
4	3	18.0	5458	True
5	4	18.0	5454	True
6	5	18.0	7788	True
7	6	18.0	5458	True
8	7	18.0	7119	True
9	8	18.0	7788	True
10	9	20.0	5019	True
11	10	19.0	5261	True
12	11	19.0	5257	True
13	12	18.0	7120	True
14	13	19.0	5544	True
15	14	18.0	7120	True
16	15	18.0	5340	True
17	16	19.0	5261	True
18	17	18.0	7119	True
19	18	18.0	5458	True
20	19	19.0	5256	True
21	20	18.0	7788	True
22	21	19.0	5261	True
23	22	19.0	5258	True
24	23	18.0	7784	True
25	24	18.0	5460	True
26	25	20.0	5025	True
27	26	19.0	5258	True
28	27	18.0	7784	True
29	28	19.0	5263	True
30	29	19.0	5258	True
31	30	19.0	5262	True
32	31	19.0	5544	True
33	32	18.0	7783	True
34	33	18.0	5455	True
35	34	19.0	5263	True
36	35	19.0	5261	True
37	36	19.0	5262	True
38	37	18.0	7119	True
39	38	19.0	5263	True
40	39	19.0	5261	True
41	40	18.0	5459	True
42	41	20.0	5026	True
43	42	18.0	5459	True
44	43	20.0	5025	True
45	44	18.0	5455	True
46	45	18.0	7120	True
47	46	19.0	5256	True
48	47	18.0	7787	True
49	48	20.0	5026	True
50	49	18.0	5458	True
51	50	20.0	5022	True
52	51	19.0	5257	True
53	52	19.0	5258	True
54	53	19.0	5257	True
55	54	18.0	7119	True
56	55	19.0	5260	True
57	56	19.0	5258	True
58	57	18.0	7119	True
59	58	18.0	7788	True
60	59	19.0	5259	True
61	60	19.0	5260	True
62	61	18.0	7788	True
63	62	19.0	5544	True
64	63	20.0	5023	True
65	64	20.0	5025	True
66	65	19.0	5545	True
67	66	18.0	5455	True
68	67	18.0	7788	True
69	68	19.0	5544	True
70	69	19.0	5543	True
71	70	19.0	5545	True
72	71	18.0	7120	True
73	72	20.0	5020	True
74	73	19.0	5543	True
75	74	18.0	7116	True
76	75	20.0	5020	True
77	76	19.0	5263	True
78	77	18.0	7783	True
79	78	19.0	5262	True
80	79	18.0	7783	True

Round 46: ram_intercept_v3_25_y86_score140_serve_initial_noop

2026-05-09T20:33:24.386269+00:00

均分 18.2875

改动: Round 53: solidify score 14-0 y86 serve-initial DOWN->NOOP after setup diagnostic and full 80-seed counterfactual guard; v3_24 mean 18.0125 min 17 max 19 to probe mean 18.29 min 17 max 19
样本: seed 0 / 80 局 / 713441 步
分数: 17.0 / 18.2875 / 19.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	0	17.0	27000	True
2	1	19.0	5262	True
3	2	18.0	5340	True
4	3	18.0	5458	True
5	4	18.0	5454	True
6	5	18.0	7788	True
7	6	18.0	5458	True
8	7	18.0	7119	True
9	8	18.0	7788	True
10	9	17.0	27000	True
11	10	19.0	5261	True
12	11	19.0	5257	True
13	12	18.0	7120	True
14	13	19.0	5544	True
15	14	18.0	7120	True
16	15	18.0	5340	True
17	16	19.0	5261	True
18	17	18.0	7119	True
19	18	18.0	5458	True
20	19	19.0	5256	True
21	20	18.0	7788	True
22	21	19.0	5261	True
23	22	19.0	5258	True
24	23	18.0	7784	True
25	24	18.0	5460	True
26	25	17.0	27000	True
27	26	19.0	5258	True
28	27	18.0	7784	True
29	28	19.0	5263	True
30	29	19.0	5258	True
31	30	19.0	5262	True
32	31	19.0	5544	True
33	32	18.0	7783	True
34	33	18.0	5455	True
35	34	19.0	5263	True
36	35	19.0	5261	True
37	36	19.0	5262	True
38	37	18.0	7119	True
39	38	19.0	5263	True
40	39	19.0	5261	True
41	40	18.0	5459	True
42	41	17.0	27000	True
43	42	18.0	5459	True
44	43	17.0	27000	True
45	44	18.0	5455	True
46	45	18.0	7120	True
47	46	19.0	5256	True
48	47	18.0	7787	True
49	48	17.0	27000	True
50	49	18.0	5458	True
51	50	17.0	27000	True
52	51	19.0	5257	True
53	52	19.0	5258	True
54	53	19.0	5257	True
55	54	18.0	7119	True
56	55	19.0	5260	True
57	56	19.0	5258	True
58	57	18.0	7119	True
59	58	18.0	7788	True
60	59	19.0	5259	True
61	60	19.0	5260	True
62	61	18.0	7788	True
63	62	19.0	5544	True
64	63	17.0	27000	True
65	64	17.0	27000	True
66	65	19.0	5545	True
67	66	18.0	5455	True
68	67	18.0	7788	True
69	68	19.0	5544	True
70	69	19.0	5543	True
71	70	19.0	5545	True
72	71	18.0	7120	True
73	72	17.0	27000	True
74	73	19.0	5543	True
75	74	18.0	7116	True
76	75	17.0	27000	True
77	76	19.0	5263	True
78	77	18.0	7783	True
79	78	19.0	5262	True
80	79	18.0	7783	True

Round 45: ram_intercept_v3_24_score00_mid_down_y154_noop

2026-05-09T20:11:08.785023+00:00

均分 18.0125

改动: Round 51: solidify score 0-0 mid-down y154 final UP->NOOP after mixed seed2/15 trace and full 80-seed counterfactual guard; v3_23 mean 17.9625 min 16 max 19 to probe mean 18.01 min 17 max 19
样本: seed 0 / 80 局 / 712891 步
分数: 17.0 / 18.0125 / 19.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	0	17.0	27000	True
2	1	19.0	5262	True
3	2	18.0	5340	True
4	3	18.0	5458	True
5	4	18.0	5454	True
6	5	17.0	7763	True
7	6	18.0	5458	True
8	7	17.0	7094	True
9	8	17.0	7763	True
10	9	17.0	27000	True
11	10	19.0	5261	True
12	11	19.0	5257	True
13	12	17.0	7095	True
14	13	19.0	5544	True
15	14	17.0	7095	True
16	15	18.0	5340	True
17	16	19.0	5261	True
18	17	17.0	7094	True
19	18	18.0	5458	True
20	19	19.0	5256	True
21	20	17.0	7763	True
22	21	19.0	5261	True
23	22	19.0	5258	True
24	23	17.0	7759	True
25	24	18.0	5460	True
26	25	17.0	27000	True
27	26	19.0	5258	True
28	27	17.0	7759	True
29	28	19.0	5263	True
30	29	19.0	5258	True
31	30	19.0	5262	True
32	31	19.0	5544	True
33	32	17.0	7758	True
34	33	18.0	5455	True
35	34	19.0	5263	True
36	35	19.0	5261	True
37	36	19.0	5262	True
38	37	17.0	7094	True
39	38	19.0	5263	True
40	39	19.0	5261	True
41	40	18.0	5459	True
42	41	17.0	27000	True
43	42	18.0	5459	True
44	43	17.0	27000	True
45	44	18.0	5455	True
46	45	17.0	7095	True
47	46	19.0	5256	True
48	47	17.0	7762	True
49	48	17.0	27000	True
50	49	18.0	5458	True
51	50	17.0	27000	True
52	51	19.0	5257	True
53	52	19.0	5258	True
54	53	19.0	5257	True
55	54	17.0	7094	True
56	55	19.0	5260	True
57	56	19.0	5258	True
58	57	17.0	7094	True
59	58	17.0	7763	True
60	59	19.0	5259	True
61	60	19.0	5260	True
62	61	17.0	7763	True
63	62	19.0	5544	True
64	63	17.0	27000	True
65	64	17.0	27000	True
66	65	19.0	5545	True
67	66	18.0	5455	True
68	67	17.0	7763	True
69	68	19.0	5544	True
70	69	19.0	5543	True
71	70	19.0	5545	True
72	71	17.0	7095	True
73	72	17.0	27000	True
74	73	19.0	5543	True
75	74	17.0	7091	True
76	75	17.0	27000	True
77	76	19.0	5263	True
78	77	17.0	7758	True
79	78	19.0	5262	True
80	79	17.0	7758	True

Round 44: ram_intercept_v3_23_high_down_score120_bounce_noop

2026-05-09T19:53:06.444230+00:00

均分 17.9625

改动: Round 50: solidify score 12-0 high-down bounce x144-148 y50-56 dy9 DOWN->NOOP after score16 floor trace and full 80-seed counterfactual guard; v3_22 mean 17.6875 min 16 max 19 to probe mean 17.96 min 16 max 19
样本: seed 0 / 80 局 / 713027 步
分数: 16.0 / 17.9625 / 19.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	0	17.0	27000	True
2	1	19.0	5262	True
3	2	16.0	5408	True
4	3	18.0	5458	True
5	4	18.0	5454	True
6	5	17.0	7763	True
7	6	18.0	5458	True
8	7	17.0	7094	True
9	8	17.0	7763	True
10	9	17.0	27000	True
11	10	19.0	5261	True
12	11	19.0	5257	True
13	12	17.0	7095	True
14	13	19.0	5544	True
15	14	17.0	7095	True
16	15	16.0	5408	True
17	16	19.0	5261	True
18	17	17.0	7094	True
19	18	18.0	5458	True
20	19	19.0	5256	True
21	20	17.0	7763	True
22	21	19.0	5261	True
23	22	19.0	5258	True
24	23	17.0	7759	True
25	24	18.0	5460	True
26	25	17.0	27000	True
27	26	19.0	5258	True
28	27	17.0	7759	True
29	28	19.0	5263	True
30	29	19.0	5258	True
31	30	19.0	5262	True
32	31	19.0	5544	True
33	32	17.0	7758	True
34	33	18.0	5455	True
35	34	19.0	5263	True
36	35	19.0	5261	True
37	36	19.0	5262	True
38	37	17.0	7094	True
39	38	19.0	5263	True
40	39	19.0	5261	True
41	40	18.0	5459	True
42	41	17.0	27000	True
43	42	18.0	5459	True
44	43	17.0	27000	True
45	44	18.0	5455	True
46	45	17.0	7095	True
47	46	19.0	5256	True
48	47	17.0	7762	True
49	48	17.0	27000	True
50	49	18.0	5458	True
51	50	17.0	27000	True
52	51	19.0	5257	True
53	52	19.0	5258	True
54	53	19.0	5257	True
55	54	17.0	7094	True
56	55	19.0	5260	True
57	56	19.0	5258	True
58	57	17.0	7094	True
59	58	17.0	7763	True
60	59	19.0	5259	True
61	60	19.0	5260	True
62	61	17.0	7763	True
63	62	19.0	5544	True
64	63	17.0	27000	True
65	64	17.0	27000	True
66	65	19.0	5545	True
67	66	18.0	5455	True
68	67	17.0	7763	True
69	68	19.0	5544	True
70	69	19.0	5543	True
71	70	19.0	5545	True
72	71	17.0	7095	True
73	72	17.0	27000	True
74	73	19.0	5543	True
75	74	17.0	7091	True
76	75	17.0	27000	True
77	76	19.0	5263	True
78	77	17.0	7758	True
79	78	19.0	5262	True
80	79	17.0	7758	True

Round 43: ram_intercept_v3_22_y86_score100_serve_initial_noop

2026-05-09T19:33:24.229013+00:00

均分 17.6875

改动: Round 49: solidify score 10-0 y86 serve-initial DOWN->NOOP after setup diagnostic and full 80-seed counterfactual guard; v3_21 mean 17.4125 min 15 max 19 to probe mean 17.69 min 16 max 19
样本: seed 0 / 80 局 / 720397 步
分数: 16.0 / 17.6875 / 19.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	0	17.0	27000	True
2	1	19.0	5262	True
3	2	16.0	5408	True
4	3	18.0	5458	True
5	4	18.0	5454	True
6	5	16.0	8098	True
7	6	18.0	5458	True
8	7	16.0	7429	True
9	8	16.0	8098	True
10	9	17.0	27000	True
11	10	19.0	5261	True
12	11	19.0	5257	True
13	12	16.0	7430	True
14	13	19.0	5544	True
15	14	16.0	7430	True
16	15	16.0	5408	True
17	16	19.0	5261	True
18	17	16.0	7429	True
19	18	18.0	5458	True
20	19	19.0	5256	True
21	20	16.0	8098	True
22	21	19.0	5261	True
23	22	19.0	5258	True
24	23	16.0	8094	True
25	24	18.0	5460	True
26	25	17.0	27000	True
27	26	19.0	5258	True
28	27	16.0	8094	True
29	28	19.0	5263	True
30	29	19.0	5258	True
31	30	19.0	5262	True
32	31	19.0	5544	True
33	32	16.0	8093	True
34	33	18.0	5455	True
35	34	19.0	5263	True
36	35	19.0	5261	True
37	36	19.0	5262	True
38	37	16.0	7429	True
39	38	19.0	5263	True
40	39	19.0	5261	True
41	40	18.0	5459	True
42	41	17.0	27000	True
43	42	18.0	5459	True
44	43	17.0	27000	True
45	44	18.0	5455	True
46	45	16.0	7430	True
47	46	19.0	5256	True
48	47	16.0	8097	True
49	48	17.0	27000	True
50	49	18.0	5458	True
51	50	17.0	27000	True
52	51	19.0	5257	True
53	52	19.0	5258	True
54	53	19.0	5257	True
55	54	16.0	7429	True
56	55	19.0	5260	True
57	56	19.0	5258	True
58	57	16.0	7429	True
59	58	16.0	8098	True
60	59	19.0	5259	True
61	60	19.0	5260	True
62	61	16.0	8098	True
63	62	19.0	5544	True
64	63	17.0	27000	True
65	64	17.0	27000	True
66	65	19.0	5545	True
67	66	18.0	5455	True
68	67	16.0	8098	True
69	68	19.0	5544	True
70	69	19.0	5543	True
71	70	19.0	5545	True
72	71	16.0	7430	True
73	72	17.0	27000	True
74	73	19.0	5543	True
75	74	16.0	7426	True
76	75	17.0	27000	True
77	76	19.0	5263	True
78	77	16.0	8093	True
79	78	19.0	5262	True
80	79	16.0	8093	True

Round 42: ram_intercept_v3_21_high_down_score80_bounce_noop

2026-05-09T19:11:10.174793+00:00

均分 17.4125

改动: Round 47: solidify score 8-0 high-down bounce x144-148 y50-56 dy9 DOWN->NOOP after full 80-seed counterfactual guard; v3_20 mean 17.1375 min 14 max 19 to probe mean 17.41 min 15 max 19
样本: seed 0 / 80 局 / 717207 步
分数: 15.0 / 17.4125 / 19.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	0	17.0	27000	True
2	1	19.0	5262	True
3	2	16.0	5408	True
4	3	18.0	5458	True
5	4	18.0	5454	True
6	5	15.0	7953	True
7	6	18.0	5458	True
8	7	15.0	7284	True
9	8	15.0	7953	True
10	9	17.0	27000	True
11	10	19.0	5261	True
12	11	19.0	5257	True
13	12	15.0	7285	True
14	13	19.0	5544	True
15	14	15.0	7285	True
16	15	16.0	5408	True
17	16	19.0	5261	True
18	17	15.0	7284	True
19	18	18.0	5458	True
20	19	19.0	5256	True
21	20	15.0	7953	True
22	21	19.0	5261	True
23	22	19.0	5258	True
24	23	15.0	7949	True
25	24	18.0	5460	True
26	25	17.0	27000	True
27	26	19.0	5258	True
28	27	15.0	7949	True
29	28	19.0	5263	True
30	29	19.0	5258	True
31	30	19.0	5262	True
32	31	19.0	5544	True
33	32	15.0	7948	True
34	33	18.0	5455	True
35	34	19.0	5263	True
36	35	19.0	5261	True
37	36	19.0	5262	True
38	37	15.0	7284	True
39	38	19.0	5263	True
40	39	19.0	5261	True
41	40	18.0	5459	True
42	41	17.0	27000	True
43	42	18.0	5459	True
44	43	17.0	27000	True
45	44	18.0	5455	True
46	45	15.0	7285	True
47	46	19.0	5256	True
48	47	15.0	7952	True
49	48	17.0	27000	True
50	49	18.0	5458	True
51	50	17.0	27000	True
52	51	19.0	5257	True
53	52	19.0	5258	True
54	53	19.0	5257	True
55	54	15.0	7284	True
56	55	19.0	5260	True
57	56	19.0	5258	True
58	57	15.0	7284	True
59	58	15.0	7953	True
60	59	19.0	5259	True
61	60	19.0	5260	True
62	61	15.0	7953	True
63	62	19.0	5544	True
64	63	17.0	27000	True
65	64	17.0	27000	True
66	65	19.0	5545	True
67	66	18.0	5455	True
68	67	15.0	7953	True
69	68	19.0	5544	True
70	69	19.0	5543	True
71	70	19.0	5545	True
72	71	15.0	7285	True
73	72	17.0	27000	True
74	73	19.0	5543	True
75	74	15.0	7281	True
76	75	17.0	27000	True
77	76	19.0	5263	True
78	77	15.0	7948	True
79	78	19.0	5262	True
80	79	15.0	7948	True

Round 41: ram_intercept_v3_20_y86_score60_serve_initial_noop

2026-05-09T18:51:51.602778+00:00

均分 17.1375

改动: Round 45: solidify score 6-0 y86 serve-initial DOWN->NOOP after full 80-seed counterfactual guard; v3_19 mean 16.8625 min 13 max 19 to probe mean 17.14 min 14 max 19
样本: seed 0 / 80 局 / 727217 步
分数: 14.0 / 17.1375 / 19.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	0	17.0	27000	True
2	1	19.0	5262	True
3	2	16.0	5408	True
4	3	18.0	5458	True
5	4	18.0	5454	True
6	5	14.0	8408	True
7	6	18.0	5458	True
8	7	14.0	7739	True
9	8	14.0	8408	True
10	9	17.0	27000	True
11	10	19.0	5261	True
12	11	19.0	5257	True
13	12	14.0	7740	True
14	13	19.0	5544	True
15	14	14.0	7740	True
16	15	16.0	5408	True
17	16	19.0	5261	True
18	17	14.0	7739	True
19	18	18.0	5458	True
20	19	19.0	5256	True
21	20	14.0	8408	True
22	21	19.0	5261	True
23	22	19.0	5258	True
24	23	14.0	8404	True
25	24	18.0	5460	True
26	25	17.0	27000	True
27	26	19.0	5258	True
28	27	14.0	8404	True
29	28	19.0	5263	True
30	29	19.0	5258	True
31	30	19.0	5262	True
32	31	19.0	5544	True
33	32	14.0	8403	True
34	33	18.0	5455	True
35	34	19.0	5263	True
36	35	19.0	5261	True
37	36	19.0	5262	True
38	37	14.0	7739	True
39	38	19.0	5263	True
40	39	19.0	5261	True
41	40	18.0	5459	True
42	41	17.0	27000	True
43	42	18.0	5459	True
44	43	17.0	27000	True
45	44	18.0	5455	True
46	45	14.0	7740	True
47	46	19.0	5256	True
48	47	14.0	8407	True
49	48	17.0	27000	True
50	49	18.0	5458	True
51	50	17.0	27000	True
52	51	19.0	5257	True
53	52	19.0	5258	True
54	53	19.0	5257	True
55	54	14.0	7739	True
56	55	19.0	5260	True
57	56	19.0	5258	True
58	57	14.0	7739	True
59	58	14.0	8408	True
60	59	19.0	5259	True
61	60	19.0	5260	True
62	61	14.0	8408	True
63	62	19.0	5544	True
64	63	17.0	27000	True
65	64	17.0	27000	True
66	65	19.0	5545	True
67	66	18.0	5455	True
68	67	14.0	8408	True
69	68	19.0	5544	True
70	69	19.0	5543	True
71	70	19.0	5545	True
72	71	14.0	7740	True
73	72	17.0	27000	True
74	73	19.0	5543	True
75	74	14.0	7736	True
76	75	17.0	27000	True
77	76	19.0	5263	True
78	77	14.0	8403	True
79	78	19.0	5262	True
80	79	14.0	8403	True

Round 40: ram_intercept_v3_19_video_check_seed1

2026-05-09T17:46:03.441851+00:00

均分 19.0

改动: video check for current retained v3_19 policy on seed 1; avoids known seed 0/9 inactive-tail loop
样本: seed 1 / 1 局 / 5262 步
分数: 19.0 / 19.0 / 19.0 低/均/高
建议: 分数无波动；每次只改一个启发式，再多 seed 测。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束	RGB	RGB 速度	RGB 预览	OBS	OBS 速度	OBS 预览
1	1	19.0	5262	True	打开			打开

Round 39: ram_intercept_v3_19_video_check

2026-05-09T16:42:02.651397+00:00

均分 17.0

改动: video check for current retained v3_19 policy; one recorded episode for report.html viewing
样本: seed 0 / 1 局 / 27000 步
分数: 17.0 / 17.0 / 17.0 低/均/高
建议: 分数无波动；每次只改一个启发式，再多 seed 测。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束	RGB	RGB 速度	RGB 预览	OBS	OBS 速度	OBS 预览
1	0	17.0	27000	True	打开	2x		打开	1x 2x

Round 38: ram_intercept_v3_19_y86_after_score22_bounce_up

2026-05-09T15:18:29.657105+00:00

均分 16.8625

改动: Round 39: solidify stateful y86 x166 bounce DOWN->UP only after score 2-2 family marker; v3_18 mean 14.9125 min 13 max 19 to probe mean 16.86 min 13 max 19
样本: seed 0 / 80 局 / 721387 步
分数: 13.0 / 16.8625 / 19.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
2026-05-09T15:22:33+00:00	None	y86_after_score22_bounce_family	Round 39 diagnostics/counterfactual_y86_bounce_x166_80seed_round39.jsonl showed ungated x166 y86 bounce DOWN->UP has a split effect: 26 score-13 seeds improve to 19 but 22 score-13 seeds collapse to 6. Score-2-2-only in diagnostics/counterfactual_y86_bounce_score22_x166_80seed_round39.jsonl was inert. Excluding score 6-0 in diagnostics/counterfactual_y86_bounce_not_score60_x166_80seed_round39.jsonl still regressed the hidden family at score 8-1 to 7. A stateful gate that first observes the score-2-2 x166 bounce family, then allows later x166 DOWN->UP interventions, improved 26 seeds with no regressions in diagnostics/counterfactual_y86_bounce_after_score22_x166_80seed_round39.jsonl: mean 14.91 to 16.86, min 13, max 19. Formal v3_19 eval scored mean 16.8625 min 13 max 19.	The y86 x166 bounce action is causal but two visually similar score chains require opposite treatment. The safe discriminator is not a single score exclusion; it is an episode-level family marker set by seeing the score-2-2 x166 bounce before applying later x166 DOWN->UP corrections.	Retained v3_19; trace remaining score-13 seeds and determine whether the floor is now pure high-down/low-up, pure y86 score6-0/8-1, or a mixed residual family.

逐局记录

局	seed	分数	步数	结束
1	0	17.0	27000	True
2	1	19.0	5262	True
3	2	16.0	5408	True
4	3	18.0	5458	True
5	4	18.0	5454	True
6	5	13.0	8143	True
7	6	18.0	5458	True
8	7	13.0	7474	True
9	8	13.0	8143	True
10	9	17.0	27000	True
11	10	19.0	5261	True
12	11	19.0	5257	True
13	12	13.0	7475	True
14	13	19.0	5544	True
15	14	13.0	7475	True
16	15	16.0	5408	True
17	16	19.0	5261	True
18	17	13.0	7474	True
19	18	18.0	5458	True
20	19	19.0	5256	True
21	20	13.0	8143	True
22	21	19.0	5261	True
23	22	19.0	5258	True
24	23	13.0	8139	True
25	24	18.0	5460	True
26	25	17.0	27000	True
27	26	19.0	5258	True
28	27	13.0	8139	True
29	28	19.0	5263	True
30	29	19.0	5258	True
31	30	19.0	5262	True
32	31	19.0	5544	True
33	32	13.0	8138	True
34	33	18.0	5455	True
35	34	19.0	5263	True
36	35	19.0	5261	True
37	36	19.0	5262	True
38	37	13.0	7474	True
39	38	19.0	5263	True
40	39	19.0	5261	True
41	40	18.0	5459	True
42	41	17.0	27000	True
43	42	18.0	5459	True
44	43	17.0	27000	True
45	44	18.0	5455	True
46	45	13.0	7475	True
47	46	19.0	5256	True
48	47	13.0	8142	True
49	48	17.0	27000	True
50	49	18.0	5458	True
51	50	17.0	27000	True
52	51	19.0	5257	True
53	52	19.0	5258	True
54	53	19.0	5257	True
55	54	13.0	7474	True
56	55	19.0	5260	True
57	56	19.0	5258	True
58	57	13.0	7474	True
59	58	13.0	8143	True
60	59	19.0	5259	True
61	60	19.0	5260	True
62	61	13.0	8143	True
63	62	19.0	5544	True
64	63	17.0	27000	True
65	64	17.0	27000	True
66	65	19.0	5545	True
67	66	18.0	5455	True
68	67	13.0	8143	True
69	68	19.0	5544	True
70	69	19.0	5543	True
71	70	19.0	5545	True
72	71	13.0	7475	True
73	72	17.0	27000	True
74	73	19.0	5543	True
75	74	13.0	7471	True
76	75	17.0	27000	True
77	76	19.0	5263	True
78	77	13.0	8138	True
79	78	19.0	5262	True
80	79	13.0	8138	True

Round 37: ram_intercept_v3_18_high_down_score40_bounce_noop

2026-05-09T14:20:12.603639+00:00

均分 14.9125

改动: Round 37: solidify score 4-0 high-down bounce x144-148 y50-56 dy9 DOWN->NOOP after history diagnostic; v3_17 mean 14.6375 min 12 max 19 to probe mean 14.91 min 13 max 19
样本: seed 0 / 80 局 / 763117 步
分数: 13.0 / 14.9125 / 19.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
2026-05-09T14:20:12+00:00	None	high_down_score40_bounce_family	Round 37 diagnostics/high_down_history_x150_v3_17_round37.json showed the high-down floor diverges one frame before x150: losses use DOWN at the dy=9 bounce frame while wins enter x150 with a stable NOOP phase. Ungated DOWN->NOOP at diagnostics/counterfactual_high_down_y165_bounce_x146_80seed_round37.jsonl improved 22 seeds but regressed four score 0-0 hidden seeds to 11; score-4-0 gating in diagnostics/counterfactual_high_down_y165_bounce_score40_80seed_round37.jsonl improved 22 seeds with no regressions, raising mean 14.64 to 14.91 and min 12 to 13. Formal v3_18 eval scored mean 14.9125 min 13 max 19.	The high-down y165 floor has a safe causal correction at score 4-0 on the dy=9 bounce frame x=144..148,y=50..56,paddle=143..150,target=154..162: replacing baseline DOWN with NOOP preserves the next-frame paddle phase. The same geometry at score 0-0 is a hidden different family and must remain untouched.	Retained v3_18; trace remaining score-13 seeds and re-split the new floor before testing any further high-down or later-y86 correction.

逐局记录

局	seed	分数	步数	结束
1	0	17.0	27000	True
2	1	13.0	6867	True
3	2	16.0	5408	True
4	3	18.0	5458	True
5	4	18.0	5454	True
6	5	13.0	8143	True
7	6	18.0	5458	True
8	7	13.0	7474	True
9	8	13.0	8143	True
10	9	17.0	27000	True
11	10	13.0	6866	True
12	11	13.0	6862	True
13	12	13.0	7475	True
14	13	19.0	5544	True
15	14	13.0	7475	True
16	15	16.0	5408	True
17	16	13.0	6866	True
18	17	13.0	7474	True
19	18	18.0	5458	True
20	19	13.0	6861	True
21	20	13.0	8143	True
22	21	13.0	6866	True
23	22	13.0	6863	True
24	23	13.0	8139	True
25	24	18.0	5460	True
26	25	17.0	27000	True
27	26	13.0	6863	True
28	27	13.0	8139	True
29	28	13.0	6868	True
30	29	13.0	6863	True
31	30	13.0	6867	True
32	31	19.0	5544	True
33	32	13.0	8138	True
34	33	18.0	5455	True
35	34	13.0	6868	True
36	35	13.0	6866	True
37	36	13.0	6867	True
38	37	13.0	7474	True
39	38	13.0	6868	True
40	39	13.0	6866	True
41	40	18.0	5459	True
42	41	17.0	27000	True
43	42	18.0	5459	True
44	43	17.0	27000	True
45	44	18.0	5455	True
46	45	13.0	7475	True
47	46	13.0	6861	True
48	47	13.0	8142	True
49	48	17.0	27000	True
50	49	18.0	5458	True
51	50	17.0	27000	True
52	51	13.0	6862	True
53	52	13.0	6863	True
54	53	13.0	6862	True
55	54	13.0	7474	True
56	55	13.0	6865	True
57	56	13.0	6863	True
58	57	13.0	7474	True
59	58	13.0	8143	True
60	59	13.0	6864	True
61	60	13.0	6865	True
62	61	13.0	8143	True
63	62	19.0	5544	True
64	63	17.0	27000	True
65	64	17.0	27000	True
66	65	19.0	5545	True
67	66	18.0	5455	True
68	67	13.0	8143	True
69	68	19.0	5544	True
70	69	19.0	5543	True
71	70	19.0	5545	True
72	71	13.0	7475	True
73	72	17.0	27000	True
74	73	19.0	5543	True
75	74	13.0	7471	True
76	75	17.0	27000	True
77	76	13.0	6868	True
78	77	13.0	8138	True
79	78	13.0	6867	True
80	79	13.0	8138	True

Round 36: ram_intercept_v3_17_y86_score82_early_setup_down

2026-05-09T13:37:32.626697+00:00

均分 14.6375

改动: Round 34: solidify score 8-2 y86 offset8 UP->DOWN correction only during steps 2050..2110 after full guard; v3_16 mean 14.3125 min 12 max 19 to probe mean 14.64 min 12 max 19
样本: seed 0 / 80 局 / 769757 步
分数: 12.0 / 14.6375 / 19.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
2026-05-09T13:47:54+00:00	None	y86_score82_early_step_family	Round 34 diagnostics/counterfactual_y86_score82_early_step_80seed_round34.jsonl showed score 8-2 y86 offset8 UP->DOWN only during steps 2050..2110 improves 0..79 mean from 14.31 to 14.64 with min 12 max 19 and 26 interventions; formal v3_17 eval scored mean 14.6375 min 12 max 19.	The score 8-2 y86 setup correction is safe only in the early episode phase around steps 2050..2110; later score 8-2 triggers around steps 2332..2336 belong to a hidden phase that collapses under the same action replacement.	Retained v3_17; trace remaining score-12 seeds and identify whether floor is now high-down y165, later y86, or another family.
2026-05-09T13:55:47+00:00	None	high_down_y165_x150_up_to_noop_rejected	Round 36 diagnostics/counterfactual_high_down_y165_x150_80seed_round36.jsonl showed high-down x150/y61 UP->NOOP fired 12 times at score 4-0 around steps 1454..1459 but left every seed unchanged: mean 14.64, min 12, max 19, improved_count 0, worsened_count 0.	The x150 win/loss action difference in diagnostics/high_down_outcome_compare_v3_17_round36.json is a correlated phase signature rather than a causal one-frame action fix; replacing UP with NOOP at this anchor is inert for score outcomes.	Keep v3_17. Use the high-down outcome comparison to inspect earlier setup/controller history or build a phase-history diagnostic instead of adding another local high-down action replacement.

逐局记录

局	seed	分数	步数	结束
1	0	17.0	27000	True
2	1	13.0	6867	True
3	2	16.0	5408	True
4	3	18.0	5458	True
5	4	18.0	5454	True
6	5	12.0	8718	True
7	6	18.0	5458	True
8	7	12.0	7448	True
9	8	12.0	8718	True
10	9	17.0	27000	True
11	10	13.0	6866	True
12	11	13.0	6862	True
13	12	12.0	7449	True
14	13	19.0	5544	True
15	14	12.0	7449	True
16	15	16.0	5408	True
17	16	13.0	6866	True
18	17	12.0	7448	True
19	18	18.0	5458	True
20	19	13.0	6861	True
21	20	12.0	8718	True
22	21	13.0	6866	True
23	22	13.0	6863	True
24	23	12.0	8714	True
25	24	18.0	5460	True
26	25	17.0	27000	True
27	26	13.0	6863	True
28	27	12.0	8714	True
29	28	13.0	6868	True
30	29	13.0	6863	True
31	30	13.0	6867	True
32	31	19.0	5544	True
33	32	12.0	8713	True
34	33	18.0	5455	True
35	34	13.0	6868	True
36	35	13.0	6866	True
37	36	13.0	6867	True
38	37	12.0	7448	True
39	38	13.0	6868	True
40	39	13.0	6866	True
41	40	18.0	5459	True
42	41	17.0	27000	True
43	42	18.0	5459	True
44	43	17.0	27000	True
45	44	18.0	5455	True
46	45	12.0	7449	True
47	46	13.0	6861	True
48	47	12.0	8717	True
49	48	17.0	27000	True
50	49	18.0	5458	True
51	50	17.0	27000	True
52	51	13.0	6862	True
53	52	13.0	6863	True
54	53	13.0	6862	True
55	54	12.0	7448	True
56	55	13.0	6865	True
57	56	13.0	6863	True
58	57	12.0	7448	True
59	58	12.0	8718	True
60	59	13.0	6864	True
61	60	13.0	6865	True
62	61	12.0	8718	True
63	62	19.0	5544	True
64	63	17.0	27000	True
65	64	17.0	27000	True
66	65	19.0	5545	True
67	66	18.0	5455	True
68	67	12.0	8718	True
69	68	19.0	5544	True
70	69	19.0	5543	True
71	70	19.0	5545	True
72	71	12.0	7449	True
73	72	17.0	27000	True
74	73	19.0	5543	True
75	74	12.0	7445	True
76	75	17.0	27000	True
77	76	13.0	6868	True
78	77	12.0	8713	True
79	78	13.0	6867	True
80	79	12.0	8713	True

Round 35: ram_intercept_v3_16_y86_score20_serve_initial_up

2026-05-09T12:05:35.206021+00:00

均分 14.3125

改动: Round 27: solidify score 2-0 y86 serve-initial DOWN->UP correction after full guard; v3_15 mean 14.1625 min 11 max 19 to probe mean 14.31 min 12 max 19
样本: seed 0 / 80 局 / 778545 步
分数: 12.0 / 14.3125 / 19.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步
2026-05-09T12:09:34.765255+00:00	None	y86_score20_serve_initial_family	Round 27 diagnostics/counterfactual_y86_score20_serve_initial_80seed_round27.jsonl showed score 2-0 serve-initial DOWN->UP improves 0..79 mean 14.16 to 14.31 and min 11 to 12; formal v3_16 eval scored mean 14.3125 min 12 max 19.	The remaining score-2-0 y86 loop is best disrupted at the point-reset first active frame x=126 y=126: baseline DOWN sets the paddle into a bad loop phase, while UP avoids the later repeated y86 loss. DOWN->NOOP is unsafe.	Retained v3_16; trace score-12 seeds and identify whether y86 remains the floor or a new family dominates.
2026-05-09T12:23:53.525867+00:00	None	high_down_y165_offset4_rejected	Round 28 diagnostics/counterfactual_high_down_y165_offset4_80seed_round28.jsonl showed high-down y165 offset4 UP->NOOP and UP->DOWN both left 0..79 scores unchanged at mean 14.31 min 12 max 19 despite interventions.	The v3_16 high-down y165 floor is not controlled by the local x=190 y=141 paddle=155 UP action; the miss likely depends on earlier setup or a different collision phase.	Keep v3_16; continue diagnosing score-12 floor, comparing y86 and high-down families with earlier setup features.
2026-05-09T12:38:12+00:00	None	high_down_y165_offset4_h3_rejected	Round 29 diagnostics/counterfactual_high_down_y165_offset4_h3_80seed_round29.jsonl showed high-down y165 offset4 UP->NOOP and UP->DOWN with horizon 3 both left 0..79 scores unchanged at mean 14.31 min 12 max 19 despite 108 interventions each.	The high-down y165 floor is not controlled by the local offset4 action duration; extending the same replacement for three frames is still inert, so the relevant control point is earlier or the failure is determined by a different phase feature.	Keep v3_16; compare earlier high-down offsets or high-down success/failure setup before adding any policy code.
2026-05-09T12:55:02+00:00	None	high_down_y165_x166_rejected	Round 30 diagnostics/counterfactual_high_down_y165_x166_80seed_round30.jsonl showed x166 y93 DOWN->NOOP drops 0..79 mean from 14.31 to 8.90 with min -20, while DOWN->UP drops mean to 4.45 with min -20 over 97 interventions.	Although high-down wins often have NOOP at a nearby x166 anchor and losses have DOWN, directly replacing the loss-family DOWN action is unsafe; the observed win/loss difference is a correlated phase signature, not a safe local action edit.	Keep v3_16; inspect earlier setup/score phase or use a safer discriminator before any high-down policy correction.
2026-05-09T13:04:18+00:00	None	high_down_y165_initial_x166_rejected	Round 31 diagnostics/counterfactual_high_down_y165_initial_x166_80seed_round31.jsonl showed the narrower score 0-0 x166/y97/pad135 DOWN->UP trigger fires 15 times but lowers 0..79 mean from 14.31 to 14.24, with regressions on seeds 33, 44, and 66 and no improvements.	The x166 improvement seen inside the broad unsafe probe was not reproducible when isolated; the apparent +1 cases were downstream artifacts of the broader intervention pattern, not a safe one-shot first-point correction.	Keep v3_16; avoid x166 high-down action replacements and focus on state/phase diagnostics or the remaining y86 floor family.
2026-05-09T13:25:41+00:00	None	y86_final_down_rejected	Round 32 diagnostics/counterfactual_y86_final_down_80seed_round32.jsonl showed y86 final x202/y86 DOWN->NOOP leaves 0..79 scores unchanged at mean 14.31 min 12 max 19 despite 262 interventions, while DOWN->UP drops mean to 5.84 and min to -11 over 132 interventions.	The repeated y86 score-12 floor is not caused by the final contact-frame DOWN action; stopping it is inert and reversing it is harmful. The loop must be determined by earlier phase/setup or target model behavior.	Keep v3_16; use floor-family summary and y86 tail offsets to inspect earlier phase or state-model explanations instead of final-frame action replacement.
2026-05-09T13:37:16+00:00	None	y86_score82_setup_rejected	Round 33 diagnostics/counterfactual_y86_score82_80seed_round33.jsonl showed ungated score 8-2 y86 offset8 UP->DOWN improves 26 early y86 floor seeds from 12 to 13 but collapses 10 later phase seeds to -13, reducing 0..79 mean from 14.31 to 11.51 and min from 12 to -13.	The score 8-2 y86 setup correction has the same hidden phase split seen at score 6-2: geometry and score alone are unsafe because later step phases require the baseline action.	Keep v3_16 until an early step gate is validated; test score 8-2 only around the safe 2086..2093 intervention band.

逐局记录

局	seed	分数	步数	结束
1	0	17.0	27000	True
2	1	12.0	7205	True
3	2	16.0	5408	True
4	3	18.0	5458	True
5	4	18.0	5454	True
6	5	12.0	8718	True
7	6	18.0	5458	True
8	7	12.0	7448	True
9	8	12.0	8718	True
10	9	17.0	27000	True
11	10	12.0	7204	True
12	11	12.0	7200	True
13	12	12.0	7449	True
14	13	19.0	5544	True
15	14	12.0	7449	True
16	15	16.0	5408	True
17	16	12.0	7204	True
18	17	12.0	7448	True
19	18	18.0	5458	True
20	19	12.0	7199	True
21	20	12.0	8718	True
22	21	12.0	7204	True
23	22	12.0	7201	True
24	23	12.0	8714	True
25	24	18.0	5460	True
26	25	17.0	27000	True
27	26	12.0	7201	True
28	27	12.0	8714	True
29	28	12.0	7206	True
30	29	12.0	7201	True
31	30	12.0	7205	True
32	31	19.0	5544	True
33	32	12.0	8713	True
34	33	18.0	5455	True
35	34	12.0	7206	True
36	35	12.0	7204	True
37	36	12.0	7205	True
38	37	12.0	7448	True
39	38	12.0	7206	True
40	39	12.0	7204	True
41	40	18.0	5459	True
42	41	17.0	27000	True
43	42	18.0	5459	True
44	43	17.0	27000	True
45	44	18.0	5455	True
46	45	12.0	7449	True
47	46	12.0	7199	True
48	47	12.0	8717	True
49	48	17.0	27000	True
50	49	18.0	5458	True
51	50	17.0	27000	True
52	51	12.0	7200	True
53	52	12.0	7201	True
54	53	12.0	7200	True
55	54	12.0	7448	True
56	55	12.0	7203	True
57	56	12.0	7201	True
58	57	12.0	7448	True
59	58	12.0	8718	True
60	59	12.0	7202	True
61	60	12.0	7203	True
62	61	12.0	8718	True
63	62	19.0	5544	True
64	63	17.0	27000	True
65	64	17.0	27000	True
66	65	19.0	5545	True
67	66	18.0	5455	True
68	67	12.0	8718	True
69	68	19.0	5544	True
70	69	19.0	5543	True
71	70	19.0	5545	True
72	71	12.0	7449	True
73	72	17.0	27000	True
74	73	19.0	5543	True
75	74	12.0	7445	True
76	75	17.0	27000	True
77	76	12.0	7206	True
78	77	12.0	8713	True
79	78	12.0	7205	True
80	79	12.0	8713	True

Round 34: ram_intercept_v3_15_y86_score62_early_setup_down

2026-05-09T11:19:26.190173+00:00

均分 14.1625

改动: Round 24: solidify step-gated y86 6-2 setup UP->DOWN correction after full guard; v3_14 mean 13.8375 min 11 max 19 to probe mean 14.16 min 11 max 19
样本: seed 0 / 80 局 / 773925 步
分数: 11.0 / 14.1625 / 19.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步
2026-05-09T11:23:31.228809+00:00	None	y86_score62_early_step_family	Round 24 diagnostics/counterfactual_y86_score62_early_step_80seed_round23.jsonl showed step-gated score 6-2 y86 offset8 UP->DOWN improves 0..79 mean 13.84 to 14.16 with min 11 max 19; formal v3_15 eval scored mean 14.1625 min 11 max 19.	Score 6-2 y86 setup correction is only safe in the early episode phase around steps 1500..1700; later score 6-2 triggers around steps 2385..2390 belong to hidden families that collapse under the same action replacement.	Retained v3_15; trace remaining score-11 seeds and look for floor-improving families rather than extending y86 score gates blindly.
2026-05-09T11:37:47.292206+00:00	None	y86_score20_setup_rejected	Round 25 diagnostics/counterfactual_y86_score20_80seed_round25.jsonl showed score 2-0 y86 offset8 UP->NOOP had no score effect and UP->DOWN dropped mean 14.16 to 9.66 with min -19 over seeds 0..79.	Remaining score-11 seeds include early score 2-0 y86 losses, but the same offset8 action replacement is either inert or catastrophic; this floor needs a different discriminator or earlier model, not direct UP replacement.	Keep v3_15; inspect successful versus catastrophic y86 families for earlier trajectory or paddle phase differences beyond score and step.
2026-05-09T11:52:11.625142+00:00	None	y86_score20_approach_x82_rejected	Round 26 diagnostics/counterfactual_y86_score20_approach_x82_80seed_round26.jsonl showed score 2-0 x82 approach UP->NOOP had no effect and UP->DOWN dropped touched seeds 11 to 9, reducing 0..79 mean 14.16 to 13.86 and min 11 to 9.	The remaining score-2-0 y86 loop is not fixed by changing the near-target x82 approach action; direct earlier paddle-direction changes still worsen the loop.	Keep v3_15; investigate even earlier serve/return setup or non-action model features instead of direct y86 approach replacement.

逐局记录

局	seed	分数	步数	结束
1	0	17.0	27000	True
2	1	12.0	7205	True
3	2	16.0	5408	True
4	3	18.0	5458	True
5	4	18.0	5454	True
6	5	11.0	8333	True
7	6	18.0	5458	True
8	7	12.0	7448	True
9	8	11.0	8333	True
10	9	17.0	27000	True
11	10	12.0	7204	True
12	11	12.0	7200	True
13	12	12.0	7449	True
14	13	19.0	5544	True
15	14	12.0	7449	True
16	15	16.0	5408	True
17	16	12.0	7204	True
18	17	12.0	7448	True
19	18	18.0	5458	True
20	19	12.0	7199	True
21	20	11.0	8333	True
22	21	12.0	7204	True
23	22	12.0	7201	True
24	23	11.0	8329	True
25	24	18.0	5460	True
26	25	17.0	27000	True
27	26	12.0	7201	True
28	27	11.0	8329	True
29	28	12.0	7206	True
30	29	12.0	7201	True
31	30	12.0	7205	True
32	31	19.0	5544	True
33	32	11.0	8328	True
34	33	18.0	5455	True
35	34	12.0	7206	True
36	35	12.0	7204	True
37	36	12.0	7205	True
38	37	12.0	7448	True
39	38	12.0	7206	True
40	39	12.0	7204	True
41	40	18.0	5459	True
42	41	17.0	27000	True
43	42	18.0	5459	True
44	43	17.0	27000	True
45	44	18.0	5455	True
46	45	12.0	7449	True
47	46	12.0	7199	True
48	47	11.0	8332	True
49	48	17.0	27000	True
50	49	18.0	5458	True
51	50	17.0	27000	True
52	51	12.0	7200	True
53	52	12.0	7201	True
54	53	12.0	7200	True
55	54	12.0	7448	True
56	55	12.0	7203	True
57	56	12.0	7201	True
58	57	12.0	7448	True
59	58	11.0	8333	True
60	59	12.0	7202	True
61	60	12.0	7203	True
62	61	11.0	8333	True
63	62	19.0	5544	True
64	63	17.0	27000	True
65	64	17.0	27000	True
66	65	19.0	5545	True
67	66	18.0	5455	True
68	67	11.0	8333	True
69	68	19.0	5544	True
70	69	19.0	5543	True
71	70	19.0	5545	True
72	71	12.0	7449	True
73	72	17.0	27000	True
74	73	19.0	5543	True
75	74	12.0	7445	True
76	75	17.0	27000	True
77	76	12.0	7206	True
78	77	11.0	8328	True
79	78	12.0	7205	True
80	79	11.0	8328	True

Round 33: ram_intercept_v3_14_y86_score42_setup_down

2026-05-09T10:51:55.791188+00:00

均分 13.8375

改动: Round 22: solidify one-shot y86 4-2 setup UP->DOWN correction after 80-seed counterfactual; v3_13 mean 13.5125 min 10 max 19 to probe mean 13.84 min 11 max 19
样本: seed 0 / 80 局 / 782713 步
分数: 11.0 / 13.8375 / 19.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
2026-05-09T10:56:35.664758+00:00	None	y86_score42_setup_family	Round 22 diagnostics/counterfactual_y86_score42_80seed_round22.jsonl showed once_y86_setup_offset8_score42_up_to_down_h1 improves 0..79 mean 13.51 to 13.84 and min 10 to 11 with 26 interventions; formal v3_14 eval scored mean 13.8375 min 11 max 19.	After the v3_13 2-2 fix, the same y86 offset8 setup failure recurs at score 4-2; one-shot UP->DOWN at x=172..176 y=56..60 dy=4 paddle=84..90 target=74..78 breaks the next repeated loss without regressions.	Retained v3_14; trace score-11 seeds to see whether the y86 score-chain continues or another family becomes the floor.
2026-05-09T11:09:09.985099+00:00	None	y86_score62_setup_rejected	Round 23 diagnostics/counterfactual_y86_score62_80seed_round23.jsonl showed once_y86_setup_offset8_score62_up_to_down_h1 improved some y86 floor seeds 11 to 12 but regressed many hidden seeds to -15; full 0..79 mean dropped 13.84 to 10.26 and min 11 to -15.	The y86 score-chain cannot be safely extended by simply adding score 6-2; this gate overlaps hidden seed families where the same geometry is harmful.	Keep v3_14; do not add score 6-2 policy correction. Trace score-11 floor and look for safer discriminators or a different floor family.

逐局记录

局	seed	分数	步数	结束
1	0	17.0	27000	True
2	1	11.0	7543	True
3	2	16.0	5408	True
4	3	18.0	5458	True
5	4	18.0	5454	True
6	5	11.0	8333	True
7	6	18.0	5458	True
8	7	12.0	7448	True
9	8	11.0	8333	True
10	9	17.0	27000	True
11	10	11.0	7542	True
12	11	11.0	7538	True
13	12	12.0	7449	True
14	13	19.0	5544	True
15	14	12.0	7449	True
16	15	16.0	5408	True
17	16	11.0	7542	True
18	17	12.0	7448	True
19	18	18.0	5458	True
20	19	11.0	7537	True
21	20	11.0	8333	True
22	21	11.0	7542	True
23	22	11.0	7539	True
24	23	11.0	8329	True
25	24	18.0	5460	True
26	25	17.0	27000	True
27	26	11.0	7539	True
28	27	11.0	8329	True
29	28	11.0	7544	True
30	29	11.0	7539	True
31	30	11.0	7543	True
32	31	19.0	5544	True
33	32	11.0	8328	True
34	33	18.0	5455	True
35	34	11.0	7544	True
36	35	11.0	7542	True
37	36	11.0	7543	True
38	37	12.0	7448	True
39	38	11.0	7544	True
40	39	11.0	7542	True
41	40	18.0	5459	True
42	41	17.0	27000	True
43	42	18.0	5459	True
44	43	17.0	27000	True
45	44	18.0	5455	True
46	45	12.0	7449	True
47	46	11.0	7537	True
48	47	11.0	8332	True
49	48	17.0	27000	True
50	49	18.0	5458	True
51	50	17.0	27000	True
52	51	11.0	7538	True
53	52	11.0	7539	True
54	53	11.0	7538	True
55	54	12.0	7448	True
56	55	11.0	7541	True
57	56	11.0	7539	True
58	57	12.0	7448	True
59	58	11.0	8333	True
60	59	11.0	7540	True
61	60	11.0	7541	True
62	61	11.0	8333	True
63	62	19.0	5544	True
64	63	17.0	27000	True
65	64	17.0	27000	True
66	65	19.0	5545	True
67	66	18.0	5455	True
68	67	11.0	8333	True
69	68	19.0	5544	True
70	69	19.0	5543	True
71	70	19.0	5545	True
72	71	12.0	7449	True
73	72	17.0	27000	True
74	73	19.0	5543	True
75	74	12.0	7445	True
76	75	17.0	27000	True
77	76	11.0	7544	True
78	77	11.0	8328	True
79	78	11.0	7543	True
80	79	11.0	8328	True

Round 32: ram_intercept_v3_13_y86_score22_setup_down

2026-05-09T10:35:06.749627+00:00

均分 13.5125

改动: Round 21: solidify one-shot y86 2-2 setup UP->DOWN correction after 80-seed score-gated counterfactual; v3_12 mean 13.1875 min 9 max 19 to probe mean 13.51 min 10 max 19
样本: seed 0 / 80 局 / 791501 步
分数: 10.0 / 13.5125 / 19.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
2026-05-09T10:39:17.244369+00:00	None	y86_score22_setup_family	Round 21 diagnostics/counterfactual_y86_offset8_scoregates_80seed_round21.jsonl showed once_y86_setup_offset8_tied22_up_to_down_h1 improves 0..79 mean 13.19 to 13.51 and min 9 to 10 with 26 interventions; formal v3_13 eval scored mean 13.5125 min 10 max 19.	The repeated y86 floor family has a safe one-shot phase correction only at score 2-2: x=172..176 y=56..60 dy=4 paddle=84..90 target=74..78 baseline UP overcorrects; DOWN breaks the repeated loss cycle. Broader offset8 gates are unsafe.	Retained v3_13; next trace remaining score-10 seeds and continue floor-focused diagnostics.

逐局记录

局	seed	分数	步数	结束
1	0	17.0	27000	True
2	1	10.0	7881	True
3	2	16.0	5408	True
4	3	18.0	5458	True
5	4	18.0	5454	True
6	5	11.0	8333	True
7	6	18.0	5458	True
8	7	12.0	7448	True
9	8	11.0	8333	True
10	9	17.0	27000	True
11	10	10.0	7880	True
12	11	10.0	7876	True
13	12	12.0	7449	True
14	13	19.0	5544	True
15	14	12.0	7449	True
16	15	16.0	5408	True
17	16	10.0	7880	True
18	17	12.0	7448	True
19	18	18.0	5458	True
20	19	10.0	7875	True
21	20	11.0	8333	True
22	21	10.0	7880	True
23	22	10.0	7877	True
24	23	11.0	8329	True
25	24	18.0	5460	True
26	25	17.0	27000	True
27	26	10.0	7877	True
28	27	11.0	8329	True
29	28	10.0	7882	True
30	29	10.0	7877	True
31	30	10.0	7881	True
32	31	19.0	5544	True
33	32	11.0	8328	True
34	33	18.0	5455	True
35	34	10.0	7882	True
36	35	10.0	7880	True
37	36	10.0	7881	True
38	37	12.0	7448	True
39	38	10.0	7882	True
40	39	10.0	7880	True
41	40	18.0	5459	True
42	41	17.0	27000	True
43	42	18.0	5459	True
44	43	17.0	27000	True
45	44	18.0	5455	True
46	45	12.0	7449	True
47	46	10.0	7875	True
48	47	11.0	8332	True
49	48	17.0	27000	True
50	49	18.0	5458	True
51	50	17.0	27000	True
52	51	10.0	7876	True
53	52	10.0	7877	True
54	53	10.0	7876	True
55	54	12.0	7448	True
56	55	10.0	7879	True
57	56	10.0	7877	True
58	57	12.0	7448	True
59	58	11.0	8333	True
60	59	10.0	7878	True
61	60	10.0	7879	True
62	61	11.0	8333	True
63	62	19.0	5544	True
64	63	17.0	27000	True
65	64	17.0	27000	True
66	65	19.0	5545	True
67	66	18.0	5455	True
68	67	11.0	8333	True
69	68	19.0	5544	True
70	69	19.0	5543	True
71	70	19.0	5545	True
72	71	12.0	7449	True
73	72	17.0	27000	True
74	73	19.0	5543	True
75	74	12.0	7445	True
76	75	17.0	27000	True
77	76	10.0	7882	True
78	77	11.0	8328	True
79	78	10.0	7881	True
80	79	11.0	8328	True

Round 31: ram_intercept_v3_12_low_up_y111_down

2026-05-09T09:48:19.656719+00:00

均分 13.1875

改动: Round 19: solidify low-up y111 precontact UP->DOWN correction after 80-seed counterfactual guard; v3_11 mean 12.75 min 9 max 19 to probe mean 13.19 min 9 max 19
样本: seed 0 / 80 局 / 800289 步
分数: 9.0 / 13.1875 / 19.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
2026-05-09T09:52:48.464977+00:00	None	score9_low_up_y111_family	Round 19 trace diagnostics/point_traces_score9_subset_round19.jsonl found repeated low_up/paddle_above_target losses for seeds 2 and 15 at x=203 y=111 dy=-8 target=111 paddle=101; counterfactual diagnostics/counterfactual_low_up_y111_80seed_round19.jsonl improved seeds 2,15,33,44,66 with no 0..79 regressions; formal v3_12 scored mean 13.1875 min 9 max 19.	The x=187..195 y=127..143 dy=-8 precontact UP action overcorrects upward before the y111 low-up return; replacing it with DOWN preserves paddle phase.	Retained v3_12; next investigate remaining score-9 y86 near-target family and avoid the rejected late-up replacement that collapsed to -17.

逐局记录

局	seed	分数	步数	结束
1	0	17.0	27000	True
2	1	9.0	8219	True
3	2	16.0	5408	True
4	3	18.0	5458	True
5	4	18.0	5454	True
6	5	11.0	8333	True
7	6	18.0	5458	True
8	7	12.0	7448	True
9	8	11.0	8333	True
10	9	17.0	27000	True
11	10	9.0	8218	True
12	11	9.0	8214	True
13	12	12.0	7449	True
14	13	19.0	5544	True
15	14	12.0	7449	True
16	15	16.0	5408	True
17	16	9.0	8218	True
18	17	12.0	7448	True
19	18	18.0	5458	True
20	19	9.0	8213	True
21	20	11.0	8333	True
22	21	9.0	8218	True
23	22	9.0	8215	True
24	23	11.0	8329	True
25	24	18.0	5460	True
26	25	17.0	27000	True
27	26	9.0	8215	True
28	27	11.0	8329	True
29	28	9.0	8220	True
30	29	9.0	8215	True
31	30	9.0	8219	True
32	31	19.0	5544	True
33	32	11.0	8328	True
34	33	18.0	5455	True
35	34	9.0	8220	True
36	35	9.0	8218	True
37	36	9.0	8219	True
38	37	12.0	7448	True
39	38	9.0	8220	True
40	39	9.0	8218	True
41	40	18.0	5459	True
42	41	17.0	27000	True
43	42	18.0	5459	True
44	43	17.0	27000	True
45	44	18.0	5455	True
46	45	12.0	7449	True
47	46	9.0	8213	True
48	47	11.0	8332	True
49	48	17.0	27000	True
50	49	18.0	5458	True
51	50	17.0	27000	True
52	51	9.0	8214	True
53	52	9.0	8215	True
54	53	9.0	8214	True
55	54	12.0	7448	True
56	55	9.0	8217	True
57	56	9.0	8215	True
58	57	12.0	7448	True
59	58	11.0	8333	True
60	59	9.0	8216	True
61	60	9.0	8217	True
62	61	11.0	8333	True
63	62	19.0	5544	True
64	63	17.0	27000	True
65	64	17.0	27000	True
66	65	19.0	5545	True
67	66	18.0	5455	True
68	67	11.0	8333	True
69	68	19.0	5544	True
70	69	19.0	5543	True
71	70	19.0	5545	True
72	71	12.0	7449	True
73	72	17.0	27000	True
74	73	19.0	5543	True
75	74	12.0	7445	True
76	75	17.0	27000	True
77	76	9.0	8220	True
78	77	11.0	8328	True
79	78	9.0	8219	True
80	79	11.0	8328	True

Round 30: ram_intercept_v3_11_low_mid_score00_noop

2026-05-09T09:24:48.512264+00:00

均分 12.75

改动: Round 18: solidify seed44/66 low-mid score00 DOWN->NOOP correction after 80-seed counterfactual guard
样本: seed 0 / 80 局 / 792374 步
分数: 9.0 / 12.75 / 19.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
2026-05-09T09:29:32.301536+00:00	None	low_mid_score00_shared_weak_family	Round 18 counterfactual diagnostics/counterfactual_seed44_first_fix_80seed_round18.jsonl changed only seeds 33,44,66 from -1 to 11; formal v3_11 80-seed eval scored mean 12.75 min 9 max 19	At score 0-0, dy=4 x=177..181 y=53..57 paddle=45..55 target=65..70, baseline DOWN keeps the paddle too low for a shared seed33/44/66 first-loss family; NOOP preserves phase.	Retained v3_11 correction; next trace remaining score-9 seeds to look for a new repeated min-limiting family.

逐局记录

局	seed	分数	步数	结束
1	0	17.0	27000	True
2	1	9.0	8219	True
3	2	9.0	3825	True
4	3	18.0	5458	True
5	4	18.0	5454	True
6	5	11.0	8333	True
7	6	18.0	5458	True
8	7	12.0	7448	True
9	8	11.0	8333	True
10	9	17.0	27000	True
11	10	9.0	8218	True
12	11	9.0	8214	True
13	12	12.0	7449	True
14	13	19.0	5544	True
15	14	12.0	7449	True
16	15	9.0	3825	True
17	16	9.0	8218	True
18	17	12.0	7448	True
19	18	18.0	5458	True
20	19	9.0	8213	True
21	20	11.0	8333	True
22	21	9.0	8218	True
23	22	9.0	8215	True
24	23	11.0	8329	True
25	24	18.0	5460	True
26	25	17.0	27000	True
27	26	9.0	8215	True
28	27	11.0	8329	True
29	28	9.0	8220	True
30	29	9.0	8215	True
31	30	9.0	8219	True
32	31	19.0	5544	True
33	32	11.0	8328	True
34	33	11.0	3872	True
35	34	9.0	8220	True
36	35	9.0	8218	True
37	36	9.0	8219	True
38	37	12.0	7448	True
39	38	9.0	8220	True
40	39	9.0	8218	True
41	40	18.0	5459	True
42	41	17.0	27000	True
43	42	18.0	5459	True
44	43	17.0	27000	True
45	44	11.0	3872	True
46	45	12.0	7449	True
47	46	9.0	8213	True
48	47	11.0	8332	True
49	48	17.0	27000	True
50	49	18.0	5458	True
51	50	17.0	27000	True
52	51	9.0	8214	True
53	52	9.0	8215	True
54	53	9.0	8214	True
55	54	12.0	7448	True
56	55	9.0	8217	True
57	56	9.0	8215	True
58	57	12.0	7448	True
59	58	11.0	8333	True
60	59	9.0	8216	True
61	60	9.0	8217	True
62	61	11.0	8333	True
63	62	19.0	5544	True
64	63	17.0	27000	True
65	64	17.0	27000	True
66	65	19.0	5545	True
67	66	11.0	3872	True
68	67	11.0	8333	True
69	68	19.0	5544	True
70	69	19.0	5543	True
71	70	19.0	5545	True
72	71	12.0	7449	True
73	72	17.0	27000	True
74	73	19.0	5543	True
75	74	12.0	7445	True
76	75	17.0	27000	True
77	76	9.0	8220	True
78	77	11.0	8328	True
79	78	9.0	8219	True
80	79	11.0	8328	True

Round 29: ram_intercept_v3_10_low_up_score01_noop_seed60_79

2026-05-09T09:02:56.491867+00:00

均分 13.8

改动: expanded validation for retained v3_10 policy over seeds 60..79 after observed seeds 0..59 mean about 11.8 min -1 max 19; look for new collapse before further tuning
样本: seed 60 / 20 局 / 241136 步
分数: -1.0 / 13.8 / 19.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	60	9.0	8217	True
2	61	11.0	8333	True
3	62	19.0	5544	True
4	63	17.0	27000	True
5	64	17.0	27000	True
6	65	19.0	5545	True
7	66	-1.0	27000	True
8	67	11.0	8333	True
9	68	19.0	5544	True
10	69	19.0	5543	True
11	70	19.0	5545	True
12	71	12.0	7449	True
13	72	17.0	27000	True
14	73	19.0	5543	True
15	74	12.0	7445	True
16	75	17.0	27000	True
17	76	9.0	8220	True
18	77	11.0	8328	True
19	78	9.0	8219	True
20	79	11.0	8328	True

Round 28: ram_intercept_v3_10_low_up_score01_noop_seed40_59

2026-05-09T09:00:07.834203+00:00

均分 12.1

改动: expanded validation for retained v3_10 policy over seeds 40..59 after 0..39 mean 11.65 min -1 max 19; look for new collapse before further tuning
样本: seed 40 / 20 局 / 247890 步
分数: -1.0 / 12.1 / 18.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	40	18.0	5459	True
2	41	17.0	27000	True
3	42	18.0	5459	True
4	43	17.0	27000	True
5	44	-1.0	27000	True
6	45	12.0	7449	True
7	46	9.0	8213	True
8	47	11.0	8332	True
9	48	17.0	27000	True
10	49	18.0	5458	True
11	50	17.0	27000	True
12	51	9.0	8214	True
13	52	9.0	8215	True
14	53	9.0	8214	True
15	54	12.0	7448	True
16	55	9.0	8217	True
17	56	9.0	8215	True
18	57	12.0	7448	True
19	58	11.0	8333	True
20	59	9.0	8216	True

Round 27: ram_intercept_v3_10_low_up_score01_noop

2026-05-09T08:56:35.179654+00:00

均分 11.65

改动: expanded-set low-up score01 correction: keep v3_9 and add one-shot 0-1 score dy=-8 x=176..178 y=172..176 paddle=132..138 target=135..150 baseline DOWN to NOOP; 40-seed counterfactual improved min -21 to -1 with only seed33 changed
样本: seed 0 / 40 局 / 372732 步
分数: -1.0 / 11.65 / 19.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	0	17.0	27000	True
2	1	9.0	8219	True
3	2	9.0	3825	True
4	3	18.0	5458	True
5	4	18.0	5454	True
6	5	11.0	8333	True
7	6	18.0	5458	True
8	7	12.0	7448	True
9	8	11.0	8333	True
10	9	17.0	27000	True
11	10	9.0	8218	True
12	11	9.0	8214	True
13	12	12.0	7449	True
14	13	19.0	5544	True
15	14	12.0	7449	True
16	15	9.0	3825	True
17	16	9.0	8218	True
18	17	12.0	7448	True
19	18	18.0	5458	True
20	19	9.0	8213	True
21	20	11.0	8333	True
22	21	9.0	8218	True
23	22	9.0	8215	True
24	23	11.0	8329	True
25	24	18.0	5460	True
26	25	17.0	27000	True
27	26	9.0	8215	True
28	27	11.0	8329	True
29	28	9.0	8220	True
30	29	9.0	8215	True
31	30	9.0	8219	True
32	31	19.0	5544	True
33	32	11.0	8328	True
34	33	-1.0	27000	True
35	34	9.0	8220	True
36	35	9.0	8218	True
37	36	9.0	8219	True
38	37	12.0	7448	True
39	38	9.0	8220	True
40	39	9.0	8218	True

Round 26: ram_intercept_v3_9_late_low_loop_noop_limited_seed20_39

2026-05-09T08:44:11.896065+00:00

均分 9.4

改动: expanded validation for retained v3_9 policy over seeds 20..39; check for overfit or new collapse after 0..19 mean 12.9 min 9 max 19
样本: seed 20 / 20 局 / 171431 步
分数: -21.0 / 9.4 / 19.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	20	11.0	8333	True
2	21	9.0	8218	True
3	22	9.0	8215	True
4	23	11.0	8329	True
5	24	18.0	5460	True
6	25	17.0	27000	True
7	26	9.0	8215	True
8	27	11.0	8329	True
9	28	9.0	8220	True
10	29	9.0	8215	True
11	30	9.0	8219	True
12	31	19.0	5544	True
13	32	11.0	8328	True
14	33	-21.0	2263	True
15	34	9.0	8220	True
16	35	9.0	8218	True
17	36	9.0	8219	True
18	37	12.0	7448	True
19	38	9.0	8220	True
20	39	9.0	8218	True

Round 25: ram_intercept_v3_9_late_low_loop_noop_limited

2026-05-09T08:42:11.067290+00:00

均分 12.9

改动: late low-loop correction capped to match validated probe: keep v3_8 and apply dy=3..4 x=169..177 y=49..57 paddle=80..105 target=60..75 UP-to-NOOP at most 64 times per episode; revalidate after uncapped formal run underperformed probe
样本: seed 0 / 20 局 / 176564 步
分数: 9.0 / 12.9 / 19.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	0	17.0	27000	True
2	1	9.0	8219	True
3	2	9.0	3825	True
4	3	18.0	5458	True
5	4	18.0	5454	True
6	5	11.0	8333	True
7	6	18.0	5458	True
8	7	12.0	7448	True
9	8	11.0	8333	True
10	9	17.0	27000	True
11	10	9.0	8218	True
12	11	9.0	8214	True
13	12	12.0	7449	True
14	13	19.0	5544	True
15	14	12.0	7449	True
16	15	9.0	3825	True
17	16	9.0	8218	True
18	17	12.0	7448	True
19	18	18.0	5458	True
20	19	9.0	8213	True

Round 24: ram_intercept_v3_9_late_low_loop_noop

2026-05-09T08:40:31.349485+00:00

均分 11.2

改动: late low-loop correction: keep v3_8 and change dy=3..4 x=169..177 y=49..57 paddle=80..105 target=60..75 baseline UP to NOOP; counterfactual 20-seed probe improved mean 9.4 to 12.9 and min -1 to 9, with max 20 to 19
样本: seed 0 / 20 局 / 176564 步
分数: 0.0 / 11.2 / 19.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	0	0.0	27000	True
2	1	9.0	8219	True
3	2	9.0	3825	True
4	3	18.0	5458	True
5	4	18.0	5454	True
6	5	11.0	8333	True
7	6	18.0	5458	True
8	7	12.0	7448	True
9	8	11.0	8333	True
10	9	0.0	27000	True
11	10	9.0	8218	True
12	11	9.0	8214	True
13	12	12.0	7449	True
14	13	19.0	5544	True
15	14	12.0	7449	True
16	15	9.0	3825	True
17	16	9.0	8218	True
18	17	12.0	7448	True
19	18	18.0	5458	True
20	19	9.0	8213	True

Round 23: ram_intercept_v3_8_high_down_setup_noop

2026-05-09T08:32:14.021723+00:00

均分 9.4

改动: high-down setup correction: keep v3_7 and change dy=8 x=121..153 y=103..167 paddle=155..195 target=165..185 baseline NOOP to DOWN; counterfactual 20-seed probe improved mean 4.6 to 9.4 with min -1 unchanged
样本: seed 0 / 20 局 / 219266 步
分数: -1.0 / 9.4 / 20.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	0	20.0	5267	True
2	1	9.0	8219	True
3	2	9.0	3825	True
4	3	-1.0	27000	True
5	4	-1.0	27000	True
6	5	11.0	8333	True
7	6	-1.0	27000	True
8	7	12.0	7448	True
9	8	11.0	8333	True
10	9	20.0	5263	True
11	10	9.0	8218	True
12	11	9.0	8214	True
13	12	12.0	7449	True
14	13	19.0	5544	True
15	14	12.0	7449	True
16	15	9.0	3825	True
17	16	9.0	8218	True
18	17	12.0	7448	True
19	18	-1.0	27000	True
20	19	9.0	8213	True

Round 22: ram_intercept_v3_7_mid_up_precontact_noop

2026-05-09T08:21:07.033577+00:00

均分 4.6

改动: mid-up y164 precontact phase correction: keep v3_6 and change only dy=-8 x=187..191 y=188..196 paddle=190..205 baseline UP to NOOP; counterfactual 20-seed probe improved mean 1.5 to 4.6 with min -1 unchanged
样本: seed 0 / 20 局 / 202825 步
分数: -1.0 / 4.6 / 20.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	0	20.0	5267	True
2	1	-1.0	6770	True
3	2	9.0	3825	True
4	3	-1.0	27000	True
5	4	-1.0	27000	True
6	5	2.0	6523	True
7	6	-1.0	27000	True
8	7	5.0	6054	True
9	8	2.0	6523	True
10	9	20.0	5263	True
11	10	-1.0	6769	True
12	11	-1.0	6765	True
13	12	5.0	6055	True
14	13	19.0	5544	True
15	14	5.0	6055	True
16	15	9.0	3825	True
17	16	-1.0	6769	True
18	17	5.0	6054	True
19	18	-1.0	27000	True
20	19	-1.0	6764	True

Round 21: ram_intercept_v3_6_mid_down_precontact_noop

2026-05-09T08:08:29.390957+00:00

均分 1.5

改动: mid-down precontact phase correction: keep v3_5 and change only dy=4 x=186..194 y=68..82 paddle=65..90 baseline UP to NOOP; counterfactual 20-seed probe improved mean 1.2 to 1.5 with min -1 unchanged
样本: seed 0 / 20 局 / 206374 步
分数: -1.0 / 1.5 / 9.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	0	-1.0	6502	True
2	1	-1.0	6770	True
3	2	9.0	3825	True
4	3	-1.0	27000	True
5	4	-1.0	27000	True
6	5	2.0	6523	True
7	6	-1.0	27000	True
8	7	5.0	6054	True
9	8	2.0	6523	True
10	9	-1.0	6498	True
11	10	-1.0	6769	True
12	11	-1.0	6765	True
13	12	5.0	6055	True
14	13	-1.0	6623	True
15	14	5.0	6055	True
16	15	9.0	3825	True
17	16	-1.0	6769	True
18	17	5.0	6054	True
19	18	-1.0	27000	True
20	19	-1.0	6764	True

Round 20: ram_intercept_v3_5_low_serve_late_noop

2026-05-09T07:55:51.258328+00:00

均分 1.2

改动: low-serve late noop correction: keep v3_4 opening low-dy8 fix and change only late low-serve target=50 noop at x=189..201 y=55..67 pad=35..65 to UP; counterfactual 20-seed probe improved mean 0.2 to 1.2 and min -21 to -1
样本: seed 0 / 20 局 / 235425 步
分数: -1.0 / 1.2 / 9.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	0	-1.0	6502	True
2	1	-1.0	7319	True
3	2	9.0	3825	True
4	3	-1.0	27000	True
5	4	-1.0	27000	True
6	5	1.0	11558	True
7	6	-1.0	27000	True
8	7	4.0	10113	True
9	8	1.0	11558	True
10	9	-1.0	6498	True
11	10	-1.0	7318	True
12	11	-1.0	7314	True
13	12	4.0	10114	True
14	13	-1.0	6623	True
15	14	4.0	10114	True
16	15	9.0	3825	True
17	16	-1.0	7318	True
18	17	4.0	10113	True
19	18	-1.0	27000	True
20	19	-1.0	7313	True

Round 19: ram_intercept_v3_4_opening_low_dy8

2026-05-09T07:41:30.825519+00:00

均分 0.9

改动: opening low dy=8 one-step correction: at 0-0 score, x=184..188 y=130..145 with paddle recovering upward, execute UP instead of baseline DOWN once; counterfactual probe improved 10-seed mean 0.5 to 0.9
样本: seed 0 / 10 局 / 138373 步
分数: -1.0 / 0.9 / 9.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	0	-1.0	6502	True
2	1	-1.0	7319	True
3	2	9.0	3825	True
4	3	-1.0	27000	True
5	4	-1.0	27000	True
6	5	1.0	11558	True
7	6	-1.0	27000	True
8	7	4.0	10113	True
9	8	1.0	11558	True
10	9	-1.0	6498	True

Round 18: ram_intercept_v3_4_late_down_keep_down_probe

2026-05-09T06:46:18.399391+00:00

均分 0.5

改动: temporary probe: if late downward ball x=188..204 y=70..140 dy>0 and previous action was down, keep down instead of noop/up; validate action-rhythm hypothesis against v3_3 baseline mean 0.5 min -1 max 9
样本: seed 0 / 10 局 / 137933 步
分数: -1.0 / 0.5 / 9.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
2026-05-09T06:47:21.736760+00:00	None	late_down_action_rhythm_rules_no_gain	Round 2 action probe over seeds 0-9: low_down_force_down scored mean 0.4 min -2 max 9, improving seed7 4->5 but hurting seed5/8 -1->-2; late_down_no_up scored mean -7.4 with seed0/5/8/9 collapses; late_down_keep_down hit 8 times but scored exactly baseline. Formal temporary keep-down policy scored mean 0.5 min -1 max 9, identical to v3_3.	Simple late downward action overrides do not solve the residual misses; the deterministic losses likely depend on earlier rally setup or paddle phase over a longer horizon than the last forced action.	Compare winning and losing point trajectories by serve/return family and first-contact setup, not only final 12-frame action rhythm.

逐局记录

局	seed	分数	步数	结束
1	0	-1.0	6502	True
2	1	-1.0	7319	True
3	2	9.0	3825	True
4	3	-1.0	27000	True
5	4	-1.0	27000	True
6	5	-1.0	11338	True
7	6	-1.0	27000	True
8	7	4.0	10113	True
9	8	-1.0	11338	True
10	9	-1.0	6498	True

Round 17: ram_intercept_v3_4_late_low_down_y86_probe

2026-05-09T06:40:34.163055+00:00

均分 0.5

改动: temporary probe: for repeated x~202 y~86 dy=4 losses, add +8 target bias only at x=196..204 y=78..96 dy=3..5; validate against v3_3 baseline mean 0.5 min -1 max 9
样本: seed 0 / 10 局 / 138299 步
分数: -1.0 / 0.5 / 9.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
2026-05-09T06:41:32.361820+00:00	None	late_low_down_target_bias_no_effect	Round probe over seeds 0-9: broad low_mid_down_plus8 improved seed5/8 from -1 to 1 but hurt seed7 from 4 to -2, mean 0.3 vs baseline 0.5. Narrow y86 +8/+6 hit 76 frames but scored exactly baseline. Formal temporary policy v3_4_late_low_down_y86_probe also scored mean 0.5 min -1 max 9, identical to v3_3.	The repeated x~202 y~86 dy=4 losses are not fixed by a late target-y bias; paddle/action phase or collision timing is likely the limiting factor.	Inspect action rhythm and mode/target jumps in the last 30-80 frames, especially noop/up/down phase before right-edge contact, before trying another policy rule.

逐局记录

局	seed	分数	步数	结束
1	0	-1.0	6502	True
2	1	-1.0	7319	True
3	2	9.0	3825	True
4	3	-1.0	27000	True
5	4	-1.0	27000	True
6	5	-1.0	11521	True
7	6	-1.0	27000	True
8	7	4.0	10113	True
9	8	-1.0	11521	True
10	9	-1.0	6498	True

Round 16: ram_intercept_v3_3_score_state_regression

2026-05-09T06:21:02.806717+00:00

均分 0.5

改动: state extraction update only: PongState now exposes ram[14] self_score and ram[13] opponent_score discovered from reward-event probes; verify no policy behavior change over seeds 0-9
样本: seed 0 / 10 局 / 138299 步
分数: -1.0 / 0.5 / 9.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	0	-1.0	6502	True
2	1	-1.0	7319	True
3	2	9.0	3825	True
4	3	-1.0	27000	True
5	4	-1.0	27000	True
6	5	-1.0	11521	True
7	6	-1.0	27000	True
8	7	4.0	10113	True
9	8	-1.0	11521	True
10	9	-1.0	6498	True

Round 15: ram_intercept_v3_4_late_lock_5seed

2026-05-09T06:19:02.799542+00:00

均分 -5.8

改动: hypothesis test: lock the last predicted intercept after ball passes SELF_PADDLE_X_RAM instead of chasing current y at the right edge; quick seeds 0-4 no-video max_steps=30000
样本: seed 0 / 5 局 / 70301 步
分数: -21.0 / -5.8 / -1.0 低/均/高
建议: 先用 RGB/OBS 标注失败模式，再改 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
2026-05-09T06:19:35.564470+00:00	None	late_intercept_lock_regression	ram_intercept_v3_4_late_lock_5seed scored [-3,-21,-3,-1,-1], mean -5.8, regressing from v3_3 seeds 0-4 mean +1.4 and 10-seed mean +0.5; seed1 collapsed to -21.	Freezing the intercept after x>192 prevents some late chasing but breaks trajectories where the ball still needs the paddle to follow the post-bounce/current y near the right edge.	Revert late-lock rule; inspect narrower action-level fixes or tooling before changing policy again.

逐局记录

局	seed	分数	步数	结束
1	0	-3.0	5884	True
2	1	-21.0	4408	True
3	2	-3.0	6009	True
4	3	-1.0	27000	True
5	4	-1.0	27000	True

Round 14: ram_intercept_v3_3_repro_baseline

2026-05-09T06:16:49.146780+00:00

均分 0.5

改动: repro baseline before next change: current ram_intercept_v3_3 over seeds 0-9, no-video max_steps=30000
样本: seed 0 / 10 局 / 138299 步
分数: -1.0 / 0.5 / 9.0 低/均/高
建议: 先用 RGB/OBS 标注失败模式，再改 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	0	-1.0	6502	True
2	1	-1.0	7319	True
3	2	9.0	3825	True
4	3	-1.0	27000	True
5	4	-1.0	27000	True
6	5	-1.0	11521	True
7	6	-1.0	27000	True
8	7	4.0	10113	True
9	8	-1.0	11521	True
10	9	-1.0	6498	True

Round 13: ram_intercept_v3_3_10seed_30k_check

2026-05-09T05:57:01.222578+00:00

均分 0.5

改动: no-video long robustness check for ram_intercept_v3_3 over seeds 0-9 with max_steps=30000; distinguishes true near-tie games from 10000-step truncation
样本: seed 0 / 10 局 / 138299 步
分数: -1.0 / 0.5 / 9.0 低/均/高
建议: 先用 RGB/OBS 标注失败模式，再改 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
2026-05-09T06:01:44.246986+00:00	None	residual_high_intercept_misses	v3_3 long 10-seed check mean +0.5 min -1 max 9; trace seed0 residual losses repeat high/intercept misses around ball_y 164..173 and target 173/193. A fresh target-to-paddle offset sweep over seeds 0-9 at max_steps=30000 showed offset 0 remains best; negative/positive offsets regressed badly.	remaining errors are phase-specific intercept misses, not a global visual/RAM coordinate offset	if improving beyond usable baseline, add narrow pattern rules for the two residual high-intercept trajectories and validate on 10+ seeds

逐局记录

局	seed	分数	步数	结束
1	0	-1.0	6502	True
2	1	-1.0	7319	True
3	2	9.0	3825	True
4	3	-1.0	27000	True
5	4	-1.0	27000	True
6	5	-1.0	11521	True
7	6	-1.0	27000	True
8	7	4.0	10113	True
9	8	-1.0	11521	True
10	9	-1.0	6498	True

Round 12: ram_intercept_v3_3_10seed_check

2026-05-09T05:55:36.672979+00:00

均分 0.5

改动: no-video robustness check for ram_intercept_v3_3 over seeds 0-9; scores expected around mean +0.5 with no -20 collapse
样本: seed 0 / 10 局 / 84144 步
分数: -2.0 / 0.5 / 9.0 低/均/高
建议: 先用 RGB/OBS 标注失败模式，再改 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束
1	0	-1.0	6502	True
2	1	-1.0	7319	True
3	2	9.0	3825	True
4	3	0.0	10000	False
5	4	0.0	10000	False
6	5	-2.0	10000	False
7	6	0.0	10000	False
8	7	3.0	10000	False
9	8	-2.0	10000	False
10	9	-1.0	6498	True

Round 11: ram_intercept_v3_3

2026-05-09T05:51:13.476925+00:00

均分 1.4

改动: low-serve pattern rule: after detecting reset serve near x=126..138 y=116..128 moving up-right, aim at y=50 for 24 policy steps; no-video sweep improved seeds 0-4 mean from -1.4 to +1.4
样本: seed 0 / 5 局 / 37646 步
分数: -1.0 / 1.4 / 9.0 低/均/高
建议: 先用 RGB/OBS 标注失败模式，再改 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束	RGB	RGB 速度	OBS	OBS 速度
1	0	-1.0	6502	True	打开	1x 2x	打开	2x 3x
2	1	-1.0	7319	True	打开		打开
3	2	9.0	3825	True	打开		打开
4	3	0.0	10000	False	打开		打开
5	4	0.0	10000	False	打开		打开

Round 10: ram_intercept_v3_2

2026-05-09T05:39:55.253489+00:00

均分 -1.4

改动: inactive serve-ready center changed from 121 to 100; no-video sweep over seeds 0-4 improved mean from -1.8 to -1.4 and kept target/intercept constants from v3_1
样本: seed 0 / 5 局 / 34843 步
分数: -20.0 / -1.4 / 9.0 低/均/高
建议: 先用 RGB/OBS 标注失败模式，再改 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
2026-05-09T05:43:28.523217+00:00	None	seed0_unfixed_low_serve	ram_intercept_v3_2 scored [-20, 9, 4, 0, 0]; inactive center improved seeds 3/4 to max_steps but seed0 still collapses at -20	remaining weakness is a specific low serve pattern; global center/intercept constants are no longer the best lever	add a targeted serve-pattern detector or policy memory for repeated point-reset low serves, validated against seed0 without regressing seeds 1-4

逐局记录

局	seed	分数	步数	结束	RGB	OBS
1	0	-20.0	856	True	打开	打开
2	1	9.0	8782	True	打开	打开
3	2	4.0	5205	True	打开	打开
4	3	0.0	10000	False	打开	打开
5	4	0.0	10000	False	打开	打开

Round 9: ram_intercept_v3_1

2026-05-09T05:26:56.672212+00:00

均分 -1.8

改动: target clamp 46..196 added on top of ram_intercept_v3; no-video sweep over seeds 0-4 improved mean from -2.0 to -1.8
样本: seed 0 / 5 局 / 23217 步
分数: -20.0 / -1.8 / 9.0 低/均/高
建议: 先用 RGB/OBS 标注失败模式，再改 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
2026-05-09T05:29:26.601423+00:00	None	seed0_repeated_low_serve_failure	ram_intercept_v3_1 scored [-20, 9, 4, -1, -1]; seed0 repeatedly loses the same low serve pattern around ball_y 52 while other seeds are near break-even or positive	policy needs serve-pattern handling or paddle phase control after point reset; target clamp and reflection tuning are not enough	compare seed0 low-serve trace with seed1 successful serve trace and add reset/serve-specific positioning if evidence supports it

逐局记录

局	seed	分数	步数	结束	RGB	OBS
1	0	-20.0	856	True	打开	打开
2	1	9.0	3872	True	打开	打开
3	2	4.0	5265	True	打开	打开
4	3	-1.0	6614	True	打开	打开
5	4	-1.0	6610	True	打开	打开

Round 8: ram_intercept_v3

2026-05-09T05:22:05.593598+00:00

均分 -2.0

改动: tuned intercept policy: runtime reflection bounds fixed; MAX_PLAY_Y_RAM=210 and PADDLE_DEADBAND=3 selected from no-video sweeps over seeds 0-4
样本: seed 0 / 5 局 / 23252 步
分数: -20.0 / -2.0 / 8.0 低/均/高
建议: 先用 RGB/OBS 标注失败模式，再改 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
2026-05-09T05:24:35.196652+00:00	None	seed0_fast_collapse	ram_intercept_v3 over seeds 0-4 scored [-20, 8, 4, -1, -1]; mean improved to -2.0 but seed0 ended at -20 in 856 steps	current intercept constants work for common rallies but fail on one early serve pattern, likely target reflection or paddle-position phase during the first return	trace seed0 first loss sequence and compare with successful seed1 before changing another policy rule

逐局记录

局	seed	分数	步数	结束	RGB	OBS
1	0	-20.0	856	True	打开	打开
2	1	8.0	3907	True	打开	打开
3	2	4.0	5265	True	打开	打开
4	3	-1.0	6614	True	打开	打开
5	4	-1.0	6610	True	打开	打开

Round 7: ram_intercept_v2

2026-05-09T05:12:41.196821+00:00

均分 -14.4

改动: stateful velocity policy: estimate dx/dy from consecutive RAM states, predict right-side intercept with wall reflection; RGB overlay calibrated from visual probe
样本: seed 0 / 5 局 / 20904 步
分数: -19.0 / -14.4 / 0.0 低/均/高
建议: 和上一轮对比，每次只改一个 heuristic。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
2026-05-09T05:15:00.338451+00:00	None	late_or_overpredicted_intercept	ram_intercept_v2 improved mean score from -20.4 to -14.4 over 5 seeds, but seeds 0-3 still ended at -17 to -19; seed 4 reached max_steps with score 0	velocity prediction helps but needs better serve/return-state handling and maybe paddle target offset/speed limits	inspect calibrated RGB videos for the first missed return, then add one targeted adjustment or trace logging before changing more policy constants

逐局记录

局	seed	分数	步数	结束	RGB	OBS
1	0	-17.0	2418	True	打开	打开
2	1	-19.0	3715	True	打开	打开
3	2	-17.0	2419	True	打开	打开
4	3	-19.0	2352	True	打开	打开
5	4	0.0	10000	False	打开	打开

Round 6: minimal_ram_xgate_v1_calibrated_overlay_diag

2026-05-09T05:11:03.604459+00:00

均分 -21.0

改动: 诊断样本：RGB 叠加层已按 visual_x=ram[49]-48.5、visual_y=ram[54]-12.5 校准，用于分析下一轮失败模式
样本: seed 0 / 1 局 / 2145 步
分数: -21.0 / -21.0 / -21.0 低/均/高
建议: 仍全输；看 RGB 失败片段，并复核 ball_y / paddle_y 映射。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束	RGB	RGB 速度	RGB 预览	OBS	OBS 速度	OBS 预览
1	0	-21.0	2145	True	打开			打开

Round 5: minimal_ram_xgate_v1_overlay_diag

2026-05-09T05:08:25.248804+00:00

均分 -21.0

改动: 诊断样本：RGB 视频叠加 RAM 推断的球、挡板、动作、reward、score，用于检查状态检测和策略失败模式
样本: seed 0 / 1 局 / 2145 步
分数: -21.0 / -21.0 / -21.0 低/均/高
建议: 仍全输；看 RGB 失败片段，并复核 ball_y / paddle_y 映射。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束	RGB	RGB 速度	RGB 预览	OBS	OBS 速度	OBS 预览
1	0	-21.0	2145	True	打开			打开

Round 4: minimal_ram_xgate_v1

2026-05-09T03:00:51.076953+00:00

均分 -20.4

改动: x-gated target: center paddle unless ball is active and in right half
样本: seed 0 / 5 局 / 7486 步
分数: -21.0 / -20.4 / -20.0 低/均/高
建议: 仍全输；看 RGB 失败片段，并复核 ball_y / paddle_y 映射。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束	RGB	RGB 速度	OBS	OBS 速度
1	0	-21.0	2185	True	打开	0.25x 0.5x 1x 2x 3x	打开	0.25x 0.5x 1x 2x 3x
2	1	-20.0	1343	True	打开	0.25x 0.5x 1x	打开	0.25x 0.5x 1x
3	2	-21.0	2186	True	打开	0.25x 0.5x 1x	打开	0.25x 0.5x 1x
4	3	-20.0	858	True	打开	0.25x 0.5x 1x	打开	0.25x 0.5x 1x
5	4	-20.0	914	True	打开	0.25x 0.5x 1x	打开	0.25x 0.5x 1x

Round 3: minimal_ram_v0

2026-05-09T02:46:35.736446+00:00

均分 -20.0

改动: dual rgb and obs video recording
样本: seed 0 / 1 局 / 1343 步
分数: -20.0 / -20.0 / -20.0 低/均/高
建议: 仍全输；看 RGB 失败片段，并复核 ball_y / paddle_y 映射。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束	RGB	RGB 速度	RGB 预览	OBS	OBS 速度	OBS 预览
1	0	-20.0	1343	True	打开	0.25x 0.5x 1x		打开	0.25x 0.5x 1x

Round 2: minimal_ram_v0

2026-05-09T02:38:15.472083+00:00

均分 -20.0

改动: minimal ram policy with per-episode gif recording
样本: seed 0 / 3 局 / 3606 步
分数: -20.0 / -20.0 / -20.0 低/均/高
建议: 仍全输；看 RGB 失败片段，并复核 ball_y / paddle_y 映射。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束	RGB	RGB 速度
1	0	-20.0	1343	True	打开	0.25x 0.5x 1x
2	1	-20.0	919	True	打开	0.25x 0.5x 1x
3	2	-20.0	1344	True	打开	0.25x 0.5x 1x

Round 1: minimal_ram_v0

2026-05-09T02:33:35.856289+00:00

均分 -20.0

改动: minimal ram policy using ram[54] ball_y and ram[51] paddle_y
样本: seed 0 / 3 局 / 3606 步
分数: -20.0 / -20.0 / -20.0 低/均/高
建议: 仍全输；看 RGB 失败片段，并复核 ball_y / paddle_y 映射。

失败记录

时间	局	失败模式	证据	假设	下一步	视频
这一轮还没有失败模式记录。

逐局记录

局	seed	分数	步数	结束	RGB	RGB 速度	RGB 预览	OBS	OBS 速度	OBS 预览
这一轮没有逐局视频记录。