{"key": "fresh-2", "idea": "branch-blend", "tuning_paired": {"n": 256, "mean_diff": 0.2073, "se": 0.0273, "worst5": [[60210, -1.707], [60052, -0.947], [60046, -0.869], [60075, -0.744], [60206, -0.728]]}, "holdout_paired": {"n": 2048, "mean_diff": 0.1834, "se": 0.0107, "sum": 375.6647, "sumsq": 547.7818, "worst5": [[11487, -2.112], [10395, -1.435], [11003, -1.325], [10431, -1.213], [11663, -1.184]]}, "misread_rate": {"lattice": 0.1624, "mine": 0.1603}, "price_error_bps": {"lattice": 13.65, "mine": 13.62}, "gas_estimate": "Extra cost only on ambiguous first-of-step calls (0.0010.5. Instead I look up quotes under each branch separately - arb branch: price = arb reading, sd 0, phase 0, rival band-clipped to it; retail branch: cut posterior mean/sd, phase 1, rival updated by the router identity - and quote the pi-weighted average of the two log bid/ask prices, weight wA = pi^g/(pi^g+(1-pi)^g). Belief update unchanged; makes the mistake cheaper rather than rarer. Tuning 60000..60255 vs lattice: g=1 +0.168 se .032, g=0.5 -0.134, g=2 +0.207 se .027, g=4 +0.112; chose g=2. Holdout 10000..12047 once: +0.183 se 0.011 (vs HA +4.447). Failed first: a fee-bucket x move-bucket logistic correction to the retail likelihood (fit with true flag on 60256..60447): lattice LR already calibrated there, in-sample misread 16.16%->16.12%, dropped. Misread/price error measured by run.py Blind on 60000..60255 (g=2 vs lattice). code_zb64 = tar of bb.py (module bb:BB; run with bl=1 bw=2 qm=qm_k1e-5.json; needs latt.py/real.py/replica2.py from the solved folder; sha256 of bb.py da865b32c061e7b59c686e476f0ff365d593869c73039271e976cc2ed3dbc2c3). Next: blend inside the qm tables (solve against the two-point belief), and apply to the Gaussian port."}