{"key": "silent-step", "idea": "thrpol.ThrLattice (lattice filter, thr=0.9, qm_k1e-5 tables, all unchanged) extended so each fully-silent step (no own trade at all) is also conditioned on \"no own retail fill\", not only \"no own arb\". After the existing diffuse()+cut(lo,hi), each surviving lattice cell at price hypothesis p is reweighted by (1-P_fill(p))^alpha, where P_fill(p)=RHO*S(delta_sell(p))+RHO*S(delta_buy(p))-RHO^2*S(delta_sell)*S(delta_buy); delta_sell(p)=ln(p*G2)-ln(sp*(1-fb)) and delta_buy(p)=ln(sp/(1-fa))-ln(p/G2) are the router crossing gaps between our fixed quotes (sp,fb,fa in force since our last trade) and the rivals price-tracked quotes at p (via the same p->G2 band-clip real.py/latt.py use elsewhere for the rivals spot, so no new latent variable beyond p is introduced); S(x)=1 if x<=0 (we are already the better price, any arriving order reaches us) else 1-Phi((ln x - LD_FULL)/1.2), the survival function of the same lognormal order-size law real.pys fill-density branch already uses. RHO, LD_FULL, the 1.2 log-sd, and G2 are all existing constants, unchanged. The ordinary first-trade likelihoods (l_arb, l_ret, pi, thr=0.9) are untouched; only the silent-step belief evolution changes.", "settings_tried": [{"setting": "alpha=0 (no-op, equivalence check), seeds 62000..62015", "n": 16, "mean_diff": 0.0, "se": 0.0, "note": "bit-identical rows to thrpol.ThrLattice, confirmed cell by cell via pair.py"}, {"setting": "alpha=0.5 vs alpha=0 (thr09 base), seeds 62000..62063", "n": 64, "mean_diff": -0.0277, "se": 0.0193, "sum": -1.7743, "sumsq": 1.5543}, {"setting": "alpha=1.0 vs alpha=0 (thr09 base), seeds 62000..62063", "n": 64, "mean_diff": -0.031, "se": 0.0276, "sum": -1.9837, "sumsq": 3.1344}, {"setting": "alpha=2.0 vs alpha=0 (thr09 base), seeds 62000..62063", "n": 64, "mean_diff": -0.0387, "se": 0.0288, "sum": -2.479, "sumsq": 3.4308}], "tuning_conclusion": "All three non-zero strengths are null to mildly negative on the tuning range, none significant (all |mean_diff| < 1.5 se) and the trend worsens monotonically with alpha, so the derived likelihood direction does not show a gain on training data. Per the brief (validate direction before tuning strength), alpha=1.0, the physically-derived strength with no extra free parameter, was frozen for confirm and holdout rather than searching further.", "confirm_paired": {"n": 64, "mean_diff": 0.0203, "se": 0.0293, "sum": 1.2972, "sumsq": 3.489, "worst5": [[62128, -0.728], [62167, -0.656], [62150, -0.357], [62139, -0.275], [62156, -0.274]], "seeds": "62128..62191, alpha=1.0 vs thr09 base", "note": "sign flips vs the tuning range; consistent with a null effect, not a reproducible gain"}, "holdout_vs_base": {"n": 128, "mean_diff": 0.0066, "se": 0.0204, "sum": 0.8497, "sumsq": 6.7648, "worst5": [[10077, -0.776], [10095, -0.532], [10038, -0.464], [10066, -0.456], [10060, -0.424]], "seeds": "10000..10127, two 64-seed commands, alpha=1.0 vs thr09 base", "note": "partial holdout (128 of the 2048-seed holdout) reported honestly under the one-hour budget; not the full 10000..12047"}, "holdout_vs_lattice_qm": {"n": 128, "mean_diff": 0.3293, "se": 0.0414, "sum": 42.1448, "sumsq": 41.7548, "worst5": [[10006, -0.71], [10102, -0.649], [10081, -0.61], [10082, -0.546], [10072, -0.519]], "seeds": "10000..10127, alpha=1.0 vs ref/lattice_qm.json", "note": "for comparison, the unmodified thr09 base scored +0.3226 se 0.0434 vs ref/lattice_qm.json on the same 128 seeds (matches the base artifacts own holdout line of +0.3229 se 0.0107 on the full 2048), so the harness replicates correctly; the candidates +0.3293 is not distinguishable from the bases +0.3226"}, "misread_rate": {"base": 0.289, "mine": 0.2886, "seeds": "10000..10127", "note": "the belief update changed but the pi>thr call itself did not; the tiny misread-rate difference is an indirect side effect of the reshaped belief feeding l_arb=self.dens(lp_arb), not a change to the classifier"}, "price_error_bps": {"base": 13.24, "mine": 13.27, "seeds": "10000..10127", "metric": "sqrt(mean((1e4*(m-ln p))^2)) over every callback, as run.pys Blind wrapper reports it"}, "blindness_equivalence_checks": {"blind": "SilentFillLattice.on_trade takes exactly is_buy, amount_x, amount_y, timestamp, reserve_x, reserve_y (plus the diagnostic-only flag argument thrpol/latt already carry and this class never reads); run.pys Blind wrapper is the only thing that calls it and it strips the true price, rival spot and arb flag before calling, same as every other policy in solved/.", "zero_change_equivalence": "alpha=0 returns from _apply_nofill before touching self.P, so the silent-step loop is byte-for-byte thrpol.ThrLattices; verified with pair.py on seeds 62000..62015, mean_diff 0.0, se 0.0, all 16 seeds exactly 0.0", "direction_check": "done on the tuning range before any strength search, per settings_tried above: the sign is non-positive at every alpha tried, so no further strengths were searched"}, "gas_estimate": {"operations": "per silent step, per surviving lattice cell (up to 121, typically well under that once cut() has zeroed the out-of-band tails): 2 subtractions for delta_sell/delta_buy, 2 evaluations of the lognormal survival S(x) (1 ln plus 1 erf-based Phi each, i.e. the same fixed-point primitives the Gaussian branch of the port already needs), 3 multiplies to combine into P_fill, 1 subtraction and (for alpha!=1, not needed at alpha=1.0) 1 pow; roughly 4x the erf/ln evaluations of a single real.py-style first-trade call, repeated for every cell instead of once", "slots": 0, "measured": false, "note": "not ported or measured on the EVM; moot in any case since the holdout gain over the thr09 base is not distinguishable from zero, well short of the +0.3 bar the README sets for a Track B result to go into Track As next port, so this idea is not recommended for porting as implemented"}, "conclusion": "No measurable gain from conditioning silent steps on no-own-retail-fill as derived here: +0.0066 se 0.0204 over the unchanged thr09 base on a 128-seed holdout slice, after a tuning/confirm pattern (-0.031 se 0.0276 then +0.0203 se 0.0293) that also shows no reproducible sign. Likely cause, visible in the per-cell P_fill profile computed during tuning: within the no-arb band (lo,hi) defined by our own quotes, P_fill(p) is a shallow, roughly symmetric curve (about 0.34 at the band edges rising to about 0.45 at the centre for typical sp,fb,fa), so the reweighting mostly rescales the belief without reshaping it much after renormalising, leaving the quoted (w, R, sd) and the next pi call nearly unchanged. This is consistent with the Fleet rules warning that an earlier scalar router reweighting lost badly; this one at least does not lose, but it also does not win the +0.3 the brief is looking for.", "budget": {"wall_clock_note": "tuning+confirm+holdout runs were PROCS=4, ~4 minutes per 64-seed command; reported n is the actual n completed under the under-an-hour compute budget for this item, not imputed or extrapolated from a smaller sample", "engine_seeds_used": 0}, "code_zb64": "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", "code_sha256": "89919f8d010c16a2623590daf1b5f5929ddd2757b75f4560d9d009f7ce0355af"}