1j" dZddlmZddlmZddlZddlmZddlZ edGdd Z e d e hd Z e d e hd Z e de hdZe de hdZe de hdZe de hdZe de dhZe e e e e e eeeeeeeeeefZe e e eeeeefZdZeeezZeeeeeZddZd dZddZd dZ d!dZ d"dZ y)#a'Interpretable input routing for the grouped MLP experiment. Each first-layer unit receives one coherent family of features. Later layers remain dense, so the model can learn interactions *between* those family-level representations without mixing unrelated raw signals in the first transform. ) annotations) dataclassN)SequenceT)frozenc&eZdZUdZded<ded<y) FeatureFamilyzDA named set of raw features routed to one or more first-layer units.strnamezfrozenset[str]featuresN)__name__ __module__ __qualname____doc____annotations__H/Users/jameslopez/projects/TradingBot25/strategies/mlp_feature_groups.pyrrsN Irrlocal_technical> osc_norm stoch_norm m2_div_norm m3_div_norm m2_tiny_norm bb_pct_b_norm m2_nooff_norm vwap_div_normmacd_pred_normrsi_subtf_normstoch_div_normstoch_peak_normm3_momentum_normon_chain> hr_norm nupl_norm sopr_norm newaddr_norm active1y_norm sendaddr_norm totalvol_normmvrv_zscore_contmvrv_zscore_value candlesticks>shooting_star_scorebullish_hammer_scorebearish_engulfing_scorebullish_engulfing_scoremacro> gc_position dxy_roc_normbtc_spx_corr_30vix_pctrank_invbtc_dom_roc_signfed_net_liq_signgold_roc_pctrankspy_above_200emayield_curve_signus2y_roc_inv_signus10y_roc_inv_signrsid_structure> rsid_rt_bear_normrsid_rt_bull_normrsid_hid_bear_normrsid_hid_bull_normrsid_reg_bear_normrsid_reg_bull_normrsid_slow_bear_normrsid_slow_bull_normrsid_delayed_dip_normrsid_delayed_peak_norm market_flow>cvd_norm rvol_norm basis_norm oi_roc_norm usdt_d_norm cross_asset btc_gold_normc l|ttk7rtdttd|dtjdtD}t ||z }|rtdt |tjtDcgc]}|Dcgc]}||jvc} c}}t}|jdjs?t||jdDcgc] \}}|r | }}}td ||Scc}wcc}}wcc}}w) uReturn the allowed ``(hidden_size, n_features)`` first-layer edges. The experiment deliberately targets the established 55→16→8→1 shape. Failing on a different width prevents a silently arbitrary allocation. z!Grouped input structure requires  first-layer units; received .c34K|]}|jywNr .0familys r z+grouped_first_layer_mask..^sV>UF>Uz.Grouped input structure has unknown features: dtyperaxisz0Grouped input structure left features unrouted: )lenGROUPED_UNIT_ALLOCATION ValueError frozensetunionsetsortednparrayr boolanyallzip) feature_cols hidden_sizeknownunknownrYfeaturemaskpresentmissings rgrouped_first_layer_maskruSs? c122/4K0L/MN#}A '   IK  V>UV WE,%'GI&QX/IZ[\\ 88. 0.V5A AL'V__ $L A. 0 D 888    !36|TXXSTXEU3Vf3V/w^e73VfKG9UVV K B 0 gs$ D* D%1D*  D0D0%D* c t||ttjt}dtDcgc]}|j c}|Dcic]!j t fd|D#c}dScc}wcc}w)z>Serializable provenance saved beside an experimental artifact.grouped_first_layer_v1c3:K|]}|jvywrUrV)rXrqrYs rrZz-grouped_structure_metadata..usT|GW7|s)r unit_familiesfamily_feature_counts)rutupledictfromkeysrar sum)rmrnfamiliesrYs `rgrouped_structure_metadatarms\;7T]]#:;&Bc |tk7rtdtd|dtjdtD}t ||z }|rtdt |tjtDcgc]}|Dcgc]}||jvc} c}}t}tjttt|ft}tj||fScc}wcc}}w)a^Return the v1 specialist-plus-dense first-layer connectivity mask. Rows 0--7 are hard-routed specialists, in the documented family order. Rows 8--15 remain fully trainable across every input feature. Keeping the architecture check strict makes the allocation auditable rather than an accidental choice for another hidden width. z(Hybrid grouped input structure requires rRrSc34K|]}|jywrUrVrWs rrZz2hybrid_grouped_first_layer_mask..sc>bF>br[z5Hybrid grouped input structure has unknown features: r\)HYBRID_GROUPED_HIDDEN_SIZErbrcrd$HYBRID_GROUPED_SPECIALIST_ALLOCATIONrerfrgrhr rionesr` HYBRID_GROUPED_DENSE_ROW_INDICESvstack)rmrnrorprYrqspecialist_rows dense_rowss rhybrid_grouped_first_layer_maskr{s0067Q6RS++6-q :   IK  c>bc dE,%'GPQWX_Q`Pabcchh; =;V5A AL'V__ $L A; =O #>?\ARS[_`J 99oz2 33 B =s D C<D <D ct||}dj|}tt}ddt tDcic]\}}t ||j c}}tttj|jjt|jt|j|jz |jdDcgc] }t|c}|Dcgc]%}t!j"|j%'c}|ttd Scc}}wcc}wcc}w)z@Return complete, reproducible provenance for the hybrid v1 mask.hybrid_grouped_first_layer_v1r^) r versionspecialist_row_familiesdense_row_indicesfeature_cols_sha256 active_edges blocked_edgesrow_active_edgesrow_feature_indicesspecialist_row_countdense_row_count)rjoinr`r enumerater r listrhashlibsha256encode hexdigestintr~sizerg flatnonzerotolist)rmrnrr feature_textspecialist_countrowrYcounts r!hybrid_grouped_structure_metadatars *< ED;;|,L?@/ ))MN$ N V Hfkk !N$ ""BC&~~l.A.A.CDNNPDHHJTYY3459XX1X5EF5EESZ5EFHLMs 3 : : <M 0?@ $ GMs E 3E *Ec2t||}tjj|}tj|t }t |jdD]2\}}|jt|t|d}d|||f<4|S)awReturn a seeded random control mask with the grouped row capacities. This deliberately samples features independently per hidden unit. It is a capacity-matched control, not a second semantic feature routing scheme. Calling the grouped builder first preserves its architecture and feature validation while providing the exact per-row active-edge counts. r\rr^F)rreplaceT) rurgrandom default_rng zeros_likerirr~choicer`r) rmrn mask_seed grouped_maskrngrr row_indexrselecteds rrandom_sparse_first_layer_maskrs,L+FL ))   *C ==T 2D#,\-=-=1-=-E#F <::c,/c,6GQV:W$(Y !$G Krc t|||}dj|}d|dtj|j j t |jt |j|jz |jdDcgc] }t |c}|Dcgc]%}tj|j'c}dScc}wcc}w)zBReturn complete, serializable provenance for a random sparse mask.rrandom_sparse_first_layer_v1zDnumpy.default_rng(seed).choice(n_features, row_count, replace=False)rr^)r r generatorrrrrr) rrrrrrrr~rrgrr)rmrnrrrrrrs r random_sparse_structure_metadatars *, Y OD;;|,L.[&~~l.A.A.CDNNPDHHJTYY3459XX1X5EF5EESZ5EFHLMs 3 : : <M  GMs #C*;*C/)rm Sequence[str]rnrreturn np.ndarray)rmrrnrrdict[str, object])rmrrnrrrrr)rmrrnrrrrr)!r __future__r dataclassesrrtypingrnumpyrgrrcLOCAL_TECHNICALON_CHAIN CANDLESTICKSMACRORSID MARKET_FLOW CROSS_ASSETrarHYBRID_GROUPED_DENSE_UNIT_COUNTr`rr{rangerrurrrrrrrrrs #! $ 19>4 Y0& ^Y8.  gy*   %y2( M96, M9o5F+GH _o h 5%$_  ($#$ !EFIhh#( #2 35OP$ 4 44..1>A&.1>Ar