{"key": "blend-tune", "idea": "branch-blend on the lattice (wA=pi^g/(pi^g+(1-pi)^g), g=2, log-price blend of arb/retail branch quotes) plus the same wA blend applied to the stored rival-spot estimate (geometric mean of the two branch rivals); belief update unchanged", "settings_tried": [{"setting": "plain blend g=2 (episode-1 BB, reference setting)", "n": 256, "mean_diff": 0.2075, "se": 0.0273}, {"setting": "g=1", "n": 256, "mean_diff": 0.1678, "se": 0.0323}, {"setting": "g=3", "n": 256, "mean_diff": 0.1749, "se": 0.0265}, {"setting": "g=2 on dt=1, g=3 on dt>=2", "n": 256, "mean_diff": 0.1699, "se": 0.0259}, {"setting": "g=2 on dt=1, g=1 on dt>=2", "n": 256, "mean_diff": 0.1935, "se": 0.0296}, {"setting": "g=2 on dt=1, g=1.5 on dt>=2", "n": 256, "mean_diff": 0.1899, "se": 0.0279}, {"setting": "g=2, wA capped 0.8 when d<8bps", "n": 256, "mean_diff": 0.2079, "se": 0.0285}, {"setting": "g=2, wA capped 0.5 when d<8bps", "n": 256, "mean_diff": 0.1686, "se": 0.0335}, {"setting": "g=2, wA capped 0.3 when d<8bps", "n": 256, "mean_diff": 0.0924, "se": 0.0351}, {"setting": "g=2, blend fees not log prices", "n": 256, "mean_diff": 0.2078, "se": 0.027}, {"setting": "g=2, retail branch forced s tables", "n": 256, "mean_diff": 0.1971, "se": 0.0256}, {"setting": "g=2, retail branch forced c tables", "n": 256, "mean_diff": 0.1669, "se": 0.0303}, {"setting": "g=1.5", "n": 256, "mean_diff": 0.2265, "se": 0.0335}, {"setting": "g=2.5", "n": 256, "mean_diff": 0.2001, "se": 0.0263}, {"setting": "g=2 + rival blend (CHOSEN)", "n": 256, "mean_diff": 0.2521, "se": 0.0275}, {"setting": "g=1.5 + rival blend", "n": 256, "mean_diff": 0.2536, "se": 0.0343}, {"setting": "g=1 + rival blend", "n": 256, "mean_diff": 0.2149, "se": 0.0338}, {"setting": "g=2.5 + rival blend", "n": 256, "mean_diff": 0.2154, "se": 0.0272}, {"setting": "g=2 + rival blend + fee blend", "n": 256, "mean_diff": 0.25, "se": 0.0277}, {"setting": "g=1.5 + rival blend + fee blend", "n": 256, "mean_diff": 0.254, "se": 0.0343}, {"setting": "g=2 + rival blend minus plain g=2 (paired, same seeds)", "n": 256, "mean_diff": 0.0446, "se": 0.0155}], "confirm_paired": {"n": 256, "mean_diff": 0.1833, "se": 0.0284, "vs_plain_g2": {"n": 256, "mean_diff": 0.0317, "se": 0.0141}, "plain_g2": {"n": 256, "mean_diff": 0.1516, "se": 0.0284}, "seeds": "60256..60511", "reference": "lattice qm_k1e-5 (fast_replica, bit-identical to latt.py)"}, "holdout_paired": {"n": 2048, "mean_diff": 0.2231, "se": 0.0112, "sum": 456.8143, "sumsq": 625.3713, "worst5": [[11487, -1.79], [10395, -1.548], [11651, -1.534], [10475, -1.468], [11433, -1.194]], "reference": "ref/lattice_qm.json", "seeds": "10000..12047", "setting": "bw=2 rb=1 qm=qm_k1e-5.json"}, "misread_rate": {"base": 0.1622, "mine": 0.1602}, "price_error_bps": {"base": 13.64, "mine": 13.59}, "gas_estimate": "Same as episode-1 branch-blend plus rival blend: on ambiguous first-of-step calls only (0.001=2 g=1,1.5,3 all below g=2), cap on wA for d<8bps (0.8 neutral, 0.5/0.3 worse: -0.04/-0.12), retail branch forced to s or c table (-0.01/-0.04), fee blend instead of log-price blend (neutral, +0.0003). g=1.5 ~ g=2. Win: blending the retail-branch rival update into the stored rival estimate, +0.045 se 0.016 over plain blend on tuning, +0.032 se 0.014 on confirm 60256..60511; holdout +0.223 se 0.011 vs ref/lattice_qm.json (episode-1 plain blend +0.183). Misread barely moves (the idea makes mistakes cheaper, not rarer). Next: carry wA into the next-step phase/cnt and s/c sync flag (soft phase), blend through the second trade of a step, and choose quotes for the mixture via table value arrays rather than averaging prices. Still well below the 0.3 bar for Track A."}