| Cell | Agent | EW | RP | Uplift | maxDD | turn |
|---|---|---|---|---|---|---|
| v6_main_w0_s0 | -0.868 | -0.820 | -1.731 | -0.048 | -16.99% | 0.0002 |
| v6_main_w0_s1 | -0.881 | -0.820 | -1.731 | -0.061 | -17.01% | 0.0005 |
| v6_main_w0_s2 | -0.864 | -0.820 | -1.731 | -0.044 | -17.11% | 0.0003 |
| v6_main_w0_s3 | -0.880 | -0.820 | -1.731 | -0.060 | -16.80% | 0.0002 |
| v6_main_w0_s4 | -0.873 | -0.820 | -1.731 | -0.053 | -16.96% | 0.0002 |
| v6_main_w1_s0 | +1.994 | +1.810 | +1.647 | +0.184 | -8.82% | 0.0004 |
| v6_main_w1_s1 | +1.985 | +1.810 | +1.647 | +0.174 | -8.85% | 0.0000 |
| v6_main_w1_s2 | +1.971 | +1.810 | +1.647 | +0.161 | -9.01% | 0.0002 |
| v6_main_w1_s3 | +1.997 | +1.810 | +1.647 | +0.186 | -8.75% | 0.0002 |
| v6_main_w1_s4 | +2.007 | +1.810 | +1.647 | +0.197 | -8.53% | 0.0002 |
| v6_main_w2_s0 | +4.729 | +4.974 | +3.290 | -0.244 | -3.14% | 0.0003 |
| v6_main_w2_s1 | +4.692 | +4.974 | +3.290 | -0.282 | -3.21% | 0.0002 |
| v6_main_w2_s2 | +4.709 | +4.974 | +3.290 | -0.265 | -3.13% | 0.0000 |
| v6_main_w2_s3 | +4.697 | +4.974 | +3.290 | -0.276 | -3.15% | 0.0002 |
| v6_main_w2_s4 | +4.716 | +4.974 | +3.290 | -0.258 | -3.00% | 0.0000 |
median uplift -0.053 | pooled p=0.495 | positive 5/15 | seeds+ 0/5 | worst DD -17.11% | mean turnover 0.0002
- uplift>0.10: NO
- p<0.05: NO
- net alpha: NO
- maxDD>-30%: OK
- turnover<=10%: OK
- seeds>=3: NO
verdict: REJECT
Interpretation (2026-09-09)
v6 tested v4 §13's two genuinely open items — factor observations and a long-horizon Sharpe reward — holding everything else at v4 config A. Neither rescues the allocator: median uplift −0.053, p=0.495.
The informative pattern is window-1 unanimity: all 5 seeds gain +0.16..+0.20 in w1 and lose in w0/w2. Seed-consistency means the agent LEARNED a systematic tilt (not noise); its failure is overfitting to regime, not failure to learn. Learning capacity was never the problem — stationarity was. This answers the "did we look clearly" challenge directly: across v3 (free/churny), v4 (constrained/EW), v5 (loose/worse), v6 (smart/regime-specific), each better-posed generation confirms the same null from a new angle, consistent with DeMiguel-Garlappi-Uppal 1/N: with diversified sleeves and ~300 lifetime allocation decisions, there is no robust learnable tilt. A v7 would need a different problem (execution is infeasible here), not a better setup of this one.