j \dZddlZddlZddlZddlZddlZddlZddl Z ddl Z ddl Z ddl m Z m Z ddlZddlZe j$j'ej$j)ej$j+ej$j-edddlZddlmZmZmZmZmZmZddl m!Z!ddl"m#Z#m$Z$m%Z%m&Z&m'Z'm(Z(ddl)m*Z*dd l+m,Z,m-Z-d Z.d Z/d Z0d Z1dZ2dZ3da4dZ5dZ6dZ7 d"dZ8d"dZ9d"dZ:dZ;dZdZ?dZ@dZA d#dZBd$dZCdZDd ZEeFd!k(reEyy)%u Two-phase trainer for the MLP activation-score strategy (strategy_mlp_scores). Phase 1 — supervised pretrain (PyTorch): Target: y_t = tanh( scale * (k-bar forward log return / k) / rolling_vol_t ) where rolling_vol is past-only (shifted) so there is no lookahead leakage. Temporal split: train on bars up to the last WFO fold's IS end, early-stop on the last fold's OOS span. (Full per-epoch expanding-window CV across all folds was deliberately simplified away — Phase 2 does the real fold-averaged model selection on the actual backtest metric.) Phase 2 — CMA-ES fine-tune (cma): Parameter vector = flattened MLP weights + [entry, exit, exit_conf, trail] thresholds. Fitness = -(mean per-fold OOS Calmar across config.WFO_FOLDS, invalid folds scoring 0, >= WFO_MIN_VALID_FOLDS valid folds required) + lambda * mean squared deviation from the pretrained weights (L2 anchor). `--fold-objective robust` keeps the same WFO-only training boundary but subtracts penalties for high fold drawdown, zero P&L/DD, low fold trade coverage, negative folds, and threshold fragility. Per-fold Calmar matches the existing WFO rescore convention in auto_optimize_loop.py (composite_score's 1000% P&L gate is meaningless on ~1-year folds, so Calmar is the per-fold metric). Usage: python3 tools/train_mlp.py --data data/COINBASE_BTCUSD-6H.csv [--hidden 16 8] [--target-k 10] [--phase1-epochs 300] [--es-generations 300] [--es-popsize 32] [--es-sigma 0.05] [--l2 1e-3] [--fold-objective mean|robust] [--es-workers 0] [--seed 42] [--smoke] [--skip-phase2] [--out PATH] Output: strategies/params/mlp/mlp_weights_{ASSET}_{TF}.json N)datetimetimezone..)ACTIVATION_NAME FEATURE_COLS_score_to_signalscalculate_metrics mlp_forwardsave_mlp_artifact)_prepare_features)grouped_first_layer_maskgrouped_structure_metadatahybrid_grouped_first_layer_mask!hybrid_grouped_structure_metadatarandom_sparse_first_layer_mask random_sparse_structure_metadata)inject_temporal_features)DEFAULT_WORKER_FRACTIONresolve_worker_count<g@Y@gI@皙??ctjdtjj |}|st d||j d|j dfS)Nz ([A-Z0-9_!]+)-([0-9]+[HDW])\.csvz&Cannot derive ASSET/TF from filename: )rematchospathbasename ValueErrorgroup) data_pathms :/Users/jameslopez/projects/TradingBot25/tools/train_mlp.pyderive_asset_tfr'RsS 4bgg6F6Fy6QRA A)MNN 771:qwwqz !!ctj|}|jjj jj |_tj |ddjjd|d<|dtjk\|dtjkz}|j|jdS)NtimeT)utcdrop)pdread_csvcolumnsstrlowerstrip to_datetimedt tz_localizeconfig TRAIN_START TRAIN_ENDloc reset_index)r$dfmasks r& load_datar>Ys Y B%%'++113BJ6 588DDTJBvJ vJ&,, ,Fv?O?O1O PD 66$< # # # ..r(ctj|djjtj}t j tj||d}|jtjjdj}tjt|tj}||d|d| z |z |d| tjt |z|dzz }tj"|}||dt$j&k\jz}||fS)zJtanh-squashed vol-normalised k-bar forward log return. Returns (y, valid).closer)prependrNư>r*)nplogvaluesastypefloat64r.Seriesdiffrolling VOL_WINDOWstdshiftfulllennantanh TARGET_SCALEisnanr7 SCORE_START)r<k log_closeret1volfwdyvalids r& build_targetr\asr'{))00<=I 99RWWY ! = >D ,,z " & & ( . .q 1 8 8C ''#b'266 "C!" #A2.!3C!H  s"cDj12A XXa[LE bjF... 6 66E e8Or(c  #$ddl$$jjrdn'$jjjrdnd#$j |t jj|jdgt|zdgz} g} g} tt| dz D]W}$jj| || |dz}| |$jjgz } | j!|Y$jj"| j%#}d}| W$j'| $j(#}$j+5| dj,j/|dddd|||k\j0z||kj0z}|||kj0z}| d#d | d |j3d |j3#$fd }||\}}||\}}$j4j7|j9d d}$jj;}t j<dddf\}}}}t|D]}|j?|jA||||} | jC|jE|9$j+5| dj,j/|dddd| jG}!|jI$j+5t|rtK||||n|!}"ddd"|dz kr|"d}}| Dcgc]}|j,jMjOjQjSjU|jVjMjOjQjSjUf}}n|dz }||k\r| d|d|ddn |dzdk(s| d|d|!dd|"d|#| dz|tY|j3tY|j3dfS#1swYxYw#1swYxYw#1swYZxYwcc}w)Nrcudampscpurdtypedevicezphase1: device=z arch=z train_bars=z val_bars=cj|jj|jjdfS)Nrar)tensorfloat32 unsqueeze)mask_arrXrctorchrZs r&tenszpretrain..tenssO Qx[ f M Qx[ f MWWXYZ\ \r(g{Gz?g-C6?)lr weight_decayrBzphase1: early stop at epoch z (best val MSE z.5f)2zphase1: epoch z train=z val=)rcarch epochs_run best_val_mse train_barsval_bars)-rkr^ is_availablebackendsr_ manual_seedrCrandomseedshapelistrangerOnnLinearTanhappend Sequentialtorfboolno_gradweight masked_fill_rEsumoptimAdam parametersMSELossinftrain zero_gradbackwardstepitemevalfloatdetachr`doublenumpycopybiasint)%rjrZr[timeshiddenepochsr{ val_startval_endrDfirst_layer_maskrrlayers_tmodulesilinmodelmask_tval_mask train_maskrlXtrytrXvayvaoptloss_fnbest_val best_statepatience since_bestepochloss train_lossval_lossrcrks%`` @@r&pretrainrss@ //1f!NN..;;=5  dIINN4 GGAJ<$v, &! ,DHG 3t9q= !hhood1gtAE{3C))" EHH   ) , ,V 4E F#.ejjP ]]_ QK   + +VGS 9*222ew6F5N5NNH%)+333J/&v\*..:J9KLLLN# %&\JHCH~HC ++  5++-$T  JChh G13r11D.Hj(Jv  uSz3'     ""//=!YY[  ]]_:=c(uWU3Z56 H ho %#+QjH&./%-c::,,.224;;=CCEJJL88??,00299;AACHHJL%- / !OJX%25'RUVWXY 2:? .wz#.>eHS>R S14 &$eai(0&)*..*:&;$' $79 99[_4!_/s+,!R#&!R0?&R==B+S #R-0R: =S c |-|Dcgc] \}}tj|t"}}}tjt ||Dcgc]V\\}}}tjtj ||tj |j gXc}}}Scc}}wcc}}}w)zEFlatten trainable weights, omitting structurally blocked connections.rb)rC ones_liker concatenatezipasarrayravel)layers weight_masksW_br=s r&flatten_layersrs@FG1 Qd3 G >> 55LFQD  1 d+RZZ]-@-@-BCD5 Hs %B3AB9 c Ngd}}tt|dz D]}||dz||}}|tj||ftntj ||t}|j ||fk7r td|d|j d||ftj||ftj} t|j} |||| z| |<|| z }||||z} ||z }|j| | f|t|k7r td|S)Nrrrz Weight mask z has shape z ; expected z=Parameter vector length does not match architecture and masks) r~rOrConesrrr|r"zerosrGrrr) thetarrrrrlin_outn_inr=rn_activers r&unflatten_layersrs)AAFCIM"26lDHt6B6J T2ZZ R 0= ::% &|B4{4::,kSXZ^R_Q`ab b HHeT]"** 5txxz?!h,'$h !AI U  q!f# CJXYY Mr(c:|dk(rttj|S|dk(rttj|S|dk(rCdttj|zdttj|zzSt d|)Nmeanminmean_mingffffff?333333?zUnknown fold objective: )rrCrrr")calmars objectives r&aggregate_fold_scorersFRWWW%&&ERVVG_%%JU2777+,,sU266'?5K/KKK / {; <>?a HD rvvdVm$ %%r(c Bt|d}t|dzttj|dz}tdt |}d}d} d} d} |D]} t| j dd} tt| j dd}t| j d d}t| j d d }|td|tz tz z }|dkr| d z } | td|| z |z z } | tdtd | z } ||z}| |z} | |z} | |z} t||t}t||t}d |ztd |dz z}|| | | ||||d}d|zd| zzd| zzd| zz|z}||z |fS)Nrg?rrd Total TradesrMax Drawdown % P&L/DD Ratio Calmar Ratio$rrg@@)drawdown_penaltypnl_dd_penalty trade_penaltynegative_fold_penaltyfragility_penalty fragile_0_05 fragile_1_0 trade_targetg?g?r)rmaxrrCceilrOgetrrROBUST_DRAWDOWN_TARGETrrROBUST_FRAGILITY_TIGHT_MARGINROBUST_FRAGILITY_WIDE_MARGIN)r fold_metricsrrmin_fold_tradesbasern_foldsrrrneg_fold_penaltymetricstradesdrawdownpnl_ddcalmar fragile_tight fragile_wider penaltiespenaltys r&robust_objective_scorers  4D*C#8M0N,OPL!S&'GNMW[[34uW[[)93?@Aw{{>378w{{>59:CX0F%FGJ``` S= c !NS,"78<GG CSvg%677 gNWM-fjB_`M,VZA]^L -S,:N1OO-(&!1.%#$ I        ! ! "     '>9 $$r(c |ddgj|dj|djjtj|djjtj|dddf|t |tj |tjt |||||| dS)Nr*r@high r)base_dfr close_arrhigh_arrstoch_peak_arrrjrrtheta0_wfoldsrmin_valid_foldsl2fold_objectiver)rrErFrCrGr}r) r<rjrrr r rr rrrs r&make_fitness_contextr svw'(--/F"[''..rzz:vJ%%,,RZZ8AqD' T JJxrzz:e**($ r(c Vtj|tj}|d}t|}t |d||d|d}tj ||dt zgdgd}t|d|}|dj}||d <|d |d <|d |d |dddtd|dd}t|||d|d|d }ggd } } } |d} |d} |dD]\}}}|j| |k\| |kz}t|dkr(| jd dddd| jdUt|| }| j||ddks|d| kr| jd| j|d| d z } d| i}|d }|d!k(r3t| | ||| \}}|j|t!| d"|d#<n| |d$krd}n t!| |}|d%t#tj$|d||z dzz}| |z| ||fS)&Nrr rrr) rrrd) @rrg>@rjractivation_scorer rrrrrFrdr)!i_long_entry_activation_threshold i_long_exit_activation_threshold-i_long_exit_activation_confirmation_thresholdi_use_long_exit_confirmationi_use_long_entry_confirmationi_trailing_stop_thresholdr rrrr  r)rrrr) min_tradesrr valid_foldsrrobustrbase_fold_scorer r)rCrrGrOrclipTHRESHOLD_SCALEr rrrr:rr rupdaterrr)rcontextr n_wrrrdparamsrrr[rrr oos_startoos_endslr diagnosticsrscoreranchors r&evaluate_thetar-4s JJuBJJ /Ez"H h-C eDSk76?GN "e +w~/F/X NN3  NN7>2 3 QJE"2!%(K-.N!1       y 9%)=gz)R %& *+ +$Wn= T]U277E$3K(,Bq+H#IJ JF 8f gz; >>r(c|ayN)_FITNESS_CONTEXT)r#s r&init_fitness_workerr1zsr(c"t|tSr/)r-r0)rs r&evaluate_theta_workerr3s %!1 22r(c Z t|||||||||| fd _i _ S)Nc>t|\}}}}|_|||fSr/)r-last_diagnostics)rfitrrr*r#fitnesss r&r8zmake_fitness..fitnesss-0>ug0N-Wj+#. GZ''r()rr#r6) r<rjrrr r rr rrrr#r8s @@r& make_fitnessr9s>" AtXuo G ( GO!G Nr(c 4ddl}|j||||ddtdttj |ddidfdfd} dk(rFdfd } t D])} | | | jsd | SStd d} | td d dtjd} tjj| t| f5t D]+} | | fdjs d | ndddS#1swYSxYw#t t"f$r)}d|dt%||||dcYd}~Sd}~wwxYw)Nri)popsizer{verboserr7rrthrr*workersc|\}}}}|j||dkr4|tj|j||t |dyy)Nr7r=)rrCrrdict) solutionresultfitsfrr>r*bestr?s r&observe_solutionz!run_cma..observe_solutions\'-$7C A tE{?H-224"#K0" D r(c jDcgc]'}tj|tj)}}||}g}t ||D]\}} ||| j |||dzdk(s| dz k(r8 d|d d dd d d xsgDcgc]}t |d c}yycc}wcc}w) Nrrrrz phase2: gen /z best_fitness=r7.4f fold_calmars=rr)askrCrrGrtellround)gen evaluators solutionsresultsrDrBrCcrFes generationsrDrGs r&run_generationzrun_cma..run_generations>@ffhGhRZZ4h GI& #Iw 7 Hf Xvt 4!8  4 8q=C;?2 ,se1[MU C?PQ 7;I7L"7L N7Lq!7L NOQ R3 H!Os ,C'C z1phase2: evaluating CMA-ES population sequentiallyc g}|D]7}|\}}}|j|||ttdif9|S)Nr6)rrAgetattr)rRoutrBrErr>r8s r&sequential_evaluatorz%run_cma..sequential_evaluatorsTC%")("37C Aw 2Db!IJLM&Jr(z phase2: CMA-ES converged at gen r#z*Parallel CMA-ES requires a fitness contextz*phase2: evaluating CMA-ES population with z worker processesspawn) max_workers mp_context initializerinitargscBtjt|Sr/)r}mapr3)rRpools r&zrun_cma..sd488DY[d;e6fr(z!phase2: worker pool unavailable (z(); falling back to sequential evaluation)r?)cmaCMAEvolutionStrategyrrrCrr~stoprYr"mp get_context concurrentfuturesProcessPoolExecutorr1OSErrorPermissionErrorrun_cma)r8theta0sigma0r;rVr{rDr?rerWr[rOr#r^excrFrUrGrcs` ` `` @@@@r&roros ! !&&.5tPR"S UB!S\"G66FtD' 3D  R R!| ?@ %C 3 4 5wwy6se<= &  gy$/GEFF 4WI=NOP(J\    3 3!+Z 4 [)s$fg779:3%@A *   K  K _ %\ /u4\ ]^wdCYZ[[\sB&(E+E:EEEEEF.F FFc ddlm}|jxs d|d|d}|d|tj|}|j dj }|d||jfi|} | dj} | jtj} tjt| tj } d} tt| D]$}| |dkDrt| d z d | |<| d z } #d} &| |d <| |d<t!| j#}t%| j'}|d|dt|d|dd|dzdd t(d dgzS)aCompute in_long_position + bars_held_norm from the current winner policy. One-cycle circularity (documented in roadmap): the position state used for training comes from the OLD model's decisions, not the new model. This is acceptable for the first retrain; a second retrain with the new policy's state would converge. Mutates df in-place, adding columns 'in_long_position' and 'bars_held_norm'. Returns the extended feature_cols list. r)generate_signalsz8results/winners/optimization_winner_strategy_mlp_scores_rz.csvz'approach-b: loading winner params from zNapproach-b: running generate_signals with old policy to get position state ...positionrrrrin_long_positionbars_held_normu+approach-b: position features injected — rIz& bars in-position, max bars_held_norm=z.3fz (=d.0fz bars))strategies.strategy_mlp_scoresrtapproach_b_paramsr.r/ilocto_dictrrErFrCrGrrOr~rrrrrr)r<assettfargsrD _gen_signals params_path params_dfr&signalsruin_long bars_heldbars_inrn_in_posmax_helds r&_inject_position_featuresrsP  W EeWAbTQU V 1+?@ K(I ^^A  & & (FXY2779//Gz"))Hoobjj)GXbjj9IG 3x= ! A;?w4IaL qLGG "%B$B7;;=!HY]]_%H 5hZqR J&s^3x|C.@ HI -/?@ @@r(cltjd}|jdd|jdtddd g |jd td |jdtd |jdtd |jdtd |jdtd |jdtd |jdgddd|jdtdd|jd tt d!|jd"td# |jd$d%d&'|jd(d%)|jd*d+,|jd-d.,|jd/d0,|jd1d%d2'|jd3d4,|jd5d%d6'|jd7gd8d9d:|jd;td<=|jd>gd?d@dA|j }|jdBk(r|j|jdC|jdDk(r|j|jdE|jr|j|jdFtj}dG}|jr%|jr|j|j}}nt|j \}}t#|j }|j$r|j'dHj)dI}|jd9k(r d dJg|_t-|j.dK|_t-|j0dL|_t-|j2d |_|dMj4tt7|dNz}dt9|j;t9|dMj4dOj;fg}dP\} } |} nKt<j>}t<j@} t<jB} tEjF|dOdQ} tI|jJ|j2|jL|_%|dMj4dO} |dR|dS|dTt7|dU|dMj4dj;dV|dMj4dOj;dW|j* |jNrtQ|||||} ntR} |jr4g| tU|} |dXt7| t7tRz dYd}|jdZk(r"tW| |j*d}|d[n|jdDk(r=tY| |j*d}|d\t|j[d]nb|jdBk(rSt]| |j*d|j}|d^|jd_t|j[t_|| }ta||jb\}}dQdOdd`}|jdd@k7r||jd}da|jfvr|dajh|k(}t|j[}||z}t|j[tkdQ|z dbz}|dc|jddd|det|j[df|dgdh n|di|jddjtm||||dM|j*|j.|jn| | || \}}|dk}i||jbtptrt9| t9| gdl|j tujvty|j dmj{j}t9|dMj4dj;t9|dMj4dOj;g|jn|jd|jdZk(rt| |j*dn^|jdDk(rt| |j*dn6|jdBk(r$t| |j*d|jndnd9itjtjjdopdq}t||}tjtj|drtj|dstj|dtdug}|dv|j-dwdx|jkdwdy|ddwdz|dQdwd{|d|dw |jsd}|N|gtdQt7|dQz Dcgc](}tj||dQz||ft}*c}z}t||}tj||tz g} t|||||| | |j|j| }!|!| \}"}#}$|d~|"dd|#D%cgc]}%t|%d|c}%t|!| |j|j2|j0|jn||jJ}&t7|}'t|&dd|'||}|&d}|j2|j|j0|j|&jd|jJ|j|&d|&dt tj|&dt tj,|&dd |d<|jdk(r|&jdi|dd<t |dt |dQt |d|t |dddd|d<|jxs)tjjdddd|d|d}(t|(||||| |d|(dtj|z dgd|d|dycc}wcc}%w)NzTwo-phase MLP trainer) descriptionz--dataT)requiredz--hidden+)typenargsdefaultz --target-kr)rrz--phase1-epochsi,z--es-generationsz --es-popsize z --es-sigmarz--l2gMbP?z--fold-objective)rrrrrzPhase-2 fold aggregation: mean matches legacy behavior; min emphasizes worst fold; mean_min blends 70%% mean / 30%% min; robust subtracts drawdown, P&L/DD, trade-count, negative-fold, and threshold-fragility penalties.)choicesrhelpz --es-workersrzWorker processes for parallel CMA-ES population evaluation. Use 0 for auto, currently 50%% of logical CPUs capped by popsize; use 1 for sequential evaluation.)rrrz--worker-fractionz.Logical CPU fraction used when --es-workers=0.z--seed*z--smoke store_truez/Tiny run on recent bars for plumbing validation)actionrz --skip-phase2)rz--outzArtifact path override)rz--assetzHAsset override (e.g. COINBASE_BTCUSD) when filename format doesn't matchz--tfz?Timeframe override (e.g. 4H) when filename format doesn't matchz --approach-bzProbe: inject in_long_position + bars_held_norm (cols 50-51) computed from the current winner artifact's position state (one-cycle circularity per roadmap). Output artifact will have 52 features.z--approach-b-paramszMWinner CSV to derive position state from (default: auto-detect from asset/tf)z--temporal-featureszLExperiment: append causal 3- and 12-bar EMA companions for ten fast signals.z--input-structure)densegroupedhybrid_grouped random_sparserzFirst-layer connectivity: dense is the production baseline; grouped routes each of the 16 first-layer units to one logical signal family; hybrid_grouped keeps eight family specialists plus eight dense rows; random_sparse is the matched seeded control.z --mask-seedzERequired with random_sparse; seed for the auditable first-layer mask.)rrz--regime)bullbearsidewaysallrzFilter Phase-1 supervised training to MVRV regime bars only (requires mvrv_regime column). Phase-2 CMA-ES always uses all IS bars.rz<--mask-seed is required with --input-structure random_sparserzQ--out is required with --input-structure hybrid_grouped to protect live artifactsz<--temporal-features requires --out to protect live artifactscPtdtjdd|dy)N[z%H:%M:%Sz] T)flush)printr*strftime)msgs r&rDzmain..logRs# $-- +,Bse4DAr(i r,rorr*g?)rrrzasset=z tf=z bars=z window=rz hidden=ztemporal-features: appended z causal EMA companionsrzTinput-structure=grouped: first-layer units are restricted to logical signal familiesz]input-structure=hybrid_grouped: eight family specialists plus eight dense rows (active_edges=rpz+input-structure=random_sparse: seeded mask=z active_edges=)rrr mvrv_regimerxzregime=z: Phase-1 training bars u → z (ryz%)zWARNING: --regime=z0 requested but mvrv_regime not in data; ignoringrr)target_k vol_window target_scaleval_spanrbnameseconds)timespec)phase1 data_file data_sha256 train_windowr{ regime_filterinput_structurecreatedrKrdzphase2: score range [z.1fz, z], seed thresholds entry=z exit=z conf=rrz"phase2: pretrain baseline fitness=rJrKrr>r?r7r) es_popsizees_sigmaes_generations l2_anchor es_workersr best_fitness fold_calmarsfold_mean_calmarfold_min_calmarphase2rr*robust_diagnosticsrrF)rrrrrrrecommended_thresholds strategiesr&mlp mlp_weights_rz.json)r~ timeframetraining feature_colszartifact written: z (zs total)zrecommended thresholds: )_argparseArgumentParser add_argumentrrr parse_argsr mask_seederrorrZtemporal_featuresr*r~rr'datar>smoketailr;rr phase1_epochsrrr|rOr1dater7 WFO_FOLDSWFO_MIN_OOS_TRADESWFO_MIN_VALID_FOLDSr. Timestamprrworker_fraction approach_brrrr rrrr r\rregimer0rErrr{rKrRhashlibsha256openread hexdigestrrrrnowrr+ isoformatr rCarray percentile skip_phase2r~rrrrr!r9rrrNrorrrrrr joinr ))prt0rDr~rr<t75r rr rrrrrjrZr[ _REGIME_INT regime_int regime_maskbeforepctrp1_metarr training_metascores0rrrr rpr8f0calmars0rrTrFr$rZs) r&mainrs ,CDANN8dN+NN:CsRGNDNNRN8NN NN>Q;<NN&U,OPNN(lnNN(fhNN&0gqxGH NN=s_aNN:'JTYab <<>D .4>>3I NO //DHH4D cd $(("2 NO BB zzdggJJr#DII. r 499 B zz WWS\ % %4 % 0   7 *a&DK !3!3R8!$"5"5q9dooq1jooc#b'D.12CHHJRZ__R-@-E-E-G)HIJ+/(    33 44LLr1. *4??DOOTMaMabDOjoob!G&tB4vc"gYh f:??1  " " $ %R6 (;(@(@(B'C8DKK= Z[ 0UBcJ #  EE(@(DE  *3|+vheCPUPYPYP[L\K]]_`cdg_hhjk l $T[[M1ab cq!UBvJ #11499iRU/1OFG 6?D?W?$--!+\ #IG =?YY~~d499d&;&@&@&BCMMORZ__Q/4467RZ__R=P=U=U=W9XY ##y0 '|T[[^ D##'773<QP##62, APTP^P^_'"<< -777K%M.!V$G gr" gr" gr" J   c2"W[[]34GH!!+As 36*Q-9LM1 c" $%     ',-EJ1cRViZ[mE\1E\a!ed1g.d;E\1L"&,7:+G HIr1dHe.#22LB"&/Ha 0C92:;(QU1a[(;< > ?w t**DIIsDOOM(m!$w-"5t\J%[ //t}}"11((9doo>"11 K O %bggd9o&> ?$RVVDO%<= # h   ( *<@HH]TVz!}9M%*:a=%9(+). /M*+ (( Ebggll<5%1%"U#CECc6",<I SETYY[2%5c$:(CD "=1I#J"KLMW1r\rrrrrrrr-r1r3r9rorr__name__r(r&r s<@  ' RWW__X-F MNO DFL    $""/ &#E9X$=&/%f>B(C?L 3 6:"I`-AhYNx zFr(