{"key": "idea-trained", "idea": "trained", "tuning_paired": {"n": 256, "mean_diff": 0.0084, "se": 0.0257, "worst5": [[60210, -3.25], [60075, -1.532], [60036, -0.805], [60191, -0.771], [60046, -0.763]]}, "holdout_paired": {"n": 2048, "mean_diff": 0.0227, "se": 0.008, "sum": 46.4915, "sumsq": 271.569, "worst5": [[10387, -1.754], [10671, -1.389], [11487, -1.089], [11821, -1.027], [11048, -1.023]]}, "misread_rate": {"lattice": 0.1624, "mine": 0.1619}, "price_error_bps": {"lattice": 13.65, "mine": 13.58}, "gas_estimate": "On top of the lattice/Gaussian filter: 20 features, 3 ln (d, size, dt), ~45 mul/add, 1 exp for the sigmoid; 20 weights as code constants, +1 storage slot (previous step's retail count). Roughly 4-6k gas per first-of-step afterSwap, ~0 otherwise.", "code_sha256": "35fc04876aaf6791c45efcc2b9087aec6eb6f5ebf5a48f1d09f68b6c7c065afd", "code_zb64": "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", "notes": "Idea: logistic correction to the lattice filter's arb/retail call, pi=sigmoid(W.f), f = lattice log-likelihood ratio llr, ln d, ln dt, prior sd, crossed/other quote offsets from prior mean, side, ln trade size (amountY), cut mass Z, our-vs-rival quote advantage, previous step's retail count, llr^2, and 7 interactions. W=(0,1,0..) reproduces latt.py exactly (checked: paired diff 0.0 on 64 seeds). Fit by pure-Python Newton L2 logistic on 200k ambiguous first trades from seeds 61000..61095 (true flag from replica, recorded only). Test error on ambiguous rows 17.32% -> 17.25%; calibration-only (bias+scale on llr) 17.28%. Binned residuals y - pi_lattice are <= 0.026 in every feature octile: the filter is already calibrated on all these features, so a trained correction has almost nothing to take. Threshold shifts on llr all lose on tuning (shift -1/-0.5/+0.5/+1/+2: -1.45/-0.31/-0.21/-0.67/-1.76). Tuning (60000..60255): cal -0.020, linear -0.027, interactions +0.008 (chosen). Holdout once: +0.023 se 0.008, misread 16.13% vs 16.22% on holdout. Not worth a port (bar 0.3). Next: info must come from outside these features, e.g. revising the call retroactively with the step's later own trades, or the router likelihood per hypothesis. code_sha256 is sha256 of the code_zb64 string. Files: trained.py (run.py module trained:Trained), trained_w.json, collect.py, fit.py."}