"""PDF: Binzagr PoV model review - before and after figures, how each was reached, and what is more accurate."""
import sys, pickle
import pandas as pd
sys.path.insert(0, '/home/claude/binzagr/docs')
from bzpdf import *  # noqa
from reportlab.platypus import Image, KeepTogether

OUT = sys.argv[1] if len(sys.argv) > 1 else '/mnt/user-data/outputs/Binzagr_Data/06_Model_Review/Binzagr_Model_Review_Before_After.pdf'
WK, CH = '/home/claude/binzagr/boost/work', '/home/claude/binzagr/boost/charts'
Wd = A4[0] - 32 * mm
N = pickle.load(open(f'{WK}/report_numbers.pkl', 'rb')); LV = N['lever']
fc = pd.read_csv(f'{WK}/fc_summary.csv').set_index('level')
rr = pickle.load(open(f'{WK}/reco_race.pkl', 'rb')); RT = rr['table'].set_index('strategy')
rf = pickle.load(open(f'{WK}/reco_final_v2.pkl', 'rb')); RES = rf['res']
ch = pickle.load(open(f'{WK}/churn_race.pkl', 'rb')); CT = ch['table'].set_index('model')
ad = pickle.load(open(f'{WK}/adoption_race.pkl', 'rb')); AT = ad['table'].set_index('model')
p10c, p10a = N['p10']; p5c, p5a = N['p5']; hitc, hita = N['hit']; ci = N['ci']
PICK = N['pick']; pct1 = lambda x: f'{100 * x:.1f}%'; pct0 = lambda x: f'{100 * x:.0f}%'
sb = lambda x: '0%' if round(100 * x) == 0 else f'{round(100 * x):+d}%'
et_ci = ch['ci']['Extra trees']


def img(name, width=Wd, ratio=None):
    from PIL import Image as PI
    w, h = PI.open(f'{CH}/{name}').size
    return Image(f'{CH}/{name}', width=width, height=width * h / w)


def H2(text):
    return [CondPageBreak(45 * mm), H(text, 1)]


story = [NextPageTemplate('main')]
story += title_block('BINZAGR PoV  ·  MODEL REVIEW', 'Can better models boost the dashboard figures?',
                     'Before and after figures for every model on binzagr.snop.live, how each "after" was reached, and whether it is more accurate.',
                     ['Version 1.0  ·  23 October 2026  ·  Site read on 23 Oct 2026: Value cockpit to Value roll-up (14 pages)',
                      'Data: 02_Cleaned v3, the same files the site uses. Our hold-out sales match the site to the cent.',
                      'Companion data: <b>06_Model_Review/data</b> (before_after_figures.csv / .json, every test table, improved picks and churn scores)'])
story.append(kpis([(f'+{CT.loc["Extra trees", "test_auc"] - CT.loc["LightGBM", "test_auc"]:.3f}', 'churn AUC on unseen outlets (0.772 to 0.797), picked before the test'),
                   (f'+{100 * (p10a - p10c):.1f} pts', f'next-best SKU picks bought ({pct1(p10c)} to {pct1(p10a)}) on the site\'s own test'),
                   (f'-{100 * (0.6686 - fc.loc["sku", "picked_at_site_bias"]):.1f} pts', 'SKU forecast error at the same bias (66.9% to 65.5%)'),
                   (f'+SAR {LV["total_change"] / 1e3:.0f}K', 'a year on the SAR 3.45M base case (under 1%)')], Wd))
story.append(Spacer(1, 4 * mm))
story.append(H('1. The answer', 0))
story.append(P('Better models lift three figures a little and leave the rest where they are. The forecast ensemble is already close to the best this '
               'data allows, and some headline figures go down when they are measured fairly. The limit is the data: eight months of history and '
               'four test weeks, not the choice of algorithm.'))
story += bullets([
    f'<b>Churn warning: real boost.</b> Extra trees, picked on validation, scores AUC {CT.loc["Extra trees", "test_auc"]:.3f} against {CT.loc["LightGBM", "test_auc"]:.3f} for '
    f'LightGBM (the site\'s model type) on the same unseen outlets: +{et_ci[1]:.3f}, 90% interval +{et_ci[0]:.3f} to +{et_ci[2]:.3f}. Its riskiest 20% holds '
    f'{pct0(CT.loc["Extra trees", "test_top20_capture"])} of the outlets that went quiet, against {pct0(CT.loc["LightGBM", "test_top20_capture"])}.',
    f'<b>Next-best SKU: small boost.</b> On the site\'s own test (picks on 25 Aug, 4,097 outlets, 8 weeks) a LightGBM ranker over new scores (EASE, RP3beta) '
    f'raises the share of picks bought from {pct1(p10c)} (our copy of the site scorer; the site reports 12.0%) to {pct1(p10a)}, 90% interval '
    f'+{100 * ci[0]:.1f} to +{100 * ci[1]:.1f} points. The top 5, which the value uses, barely moves.',
    f'<b>Cluster adoption: boost in our copy, but the site figure needs checking.</b> Blending logistic regression with XGBoost lifts the within-cluster AUC from '
    f'{AT.loc["LightGBM", "test_auc_within_mean"]:.2f} to {AT.loc[ad["pick"], "test_auc_within_mean"]:.2f}. With a time-honest split no model reproduces the site\'s 0.90 '
    f'(we reach {AT.loc["LightGBM", "test_auc_pooled"]:.2f} to {AT.loc[ad["pick"], "test_auc_pooled"]:.2f}).',
    f'<b>Demand forecast: keep it, show the bias.</b> At SKU level one equal blend of the best three models is 1.4 points better than the site at the same bias. '
    f'At van, principal and cluster level the site\'s blends are 0.6 to 1.4 points better than anything we could build without its foundation models.',
    '<b>Goes down when measured fairly:</b> the "22% less SKU error" falls to about 15% if the forecast must be unbiased; the principal-level ML loses to a '
    '4-week weekday average on the page as shown; the company-total comparison cannot rank models at all in four weeks.',
    f'<b>Value:</b> model accuracy drives only about SAR 217K of the SAR 3.45M. All upgrades together add about SAR {LV["total_change"] / 1e3:.0f}K a year. '
    'The large levers (fleet SAR 1.39M, labour SAR 639K, price SAR 397K) rest on benchmarks and assumptions that better data, not better models, would firm up.'])
story.append(Spacer(1, 2 * mm))
story.append(callout('<b>How to read "before" and "after".</b> Before = the figure on the site. After = the same test re-run on the same data with the best '
                     'model or combination, chosen on earlier validation windows only; the test weeks were used once. Where the site\'s exact model could not be '
                     'run (its three foundation models), we compare against the site\'s own published result and show our copy of its method beside it.', Wd))

# ---------------------------------------------------------------- replacement table
REP = pickle.load(open(f'{WK}/replacements.pkl', 'rb')); TYPES = REP['types']
TCOL = {'Replace': '#2a78d6', 'Fix': '#B3261E', 'Add': '#1B7F3B', 'Verify first': '#B26A00', 'Recompute': '#6250d6', 'Keep': '#7C8591'}
story.append(PageBreak())
story += H2('2. Before and after: every figure to change on the site')
story.append(P('One row per figure, page by page, in the words the site uses. "On the site now" is the text to find; "Replace with" is the text to put in its '
               'place. The same table is in data/before_after_figures.csv and .json, with the numbers in separate columns and the basis of every after value.'))
leg = [['Type', 'What to do']] + [[f'<font color="{TCOL[k]}"><b>{k}</b></font>', v] for k, v in TYPES.items()]
story.append(table(leg, [26 * mm, 152 * mm], markup_cols=(0,), font_size=7.8))
story.append(Spacer(1, 3 * mm))
rt = [['#', 'Page and place', 'On the site now', 'Replace with', 'Type']]
for r_ in REP['rows']:
    rt.append([str(r_['no']), f"<b>{clean(r_['page'])}</b><br/>{clean(r_['where_on_page'])}", r_['on_site_now'], r_['replace_with'],
               f'<font color="{TCOL[r_["type"]]}"><b>{r_["type"]}</b></font>'])
story.append(table(rt, [7 * mm, 36 * mm, 55 * mm, 58 * mm, 22 * mm], markup_cols=(1, 4), font_size=7.5))
story.append(P('Replace rows use numbers measured on the same test once the new model is live on the site; until then the current figures stand. '
               'Verify-first rows are lower in our time-honest re-test, so they should be re-checked before the demo, not raised.', 'small'))

# ---------------------------------------------------------------- method
story += H2('3. How the after figures were reached')
story.append(P('Every test copies the site\'s own: the same files, the same series, the same test windows and the same comparison baselines. We first checked '
               'that our copy reproduces the site\'s baselines; only then did we try new models.'))
base = [['Check', 'Site', 'Our copy'],
        ['Hold-out van sales, 23 Sep - 20 Oct (company total)', 'SAR 5,060,755', 'SAR 5,060,755 (all 92 vans within SAR 0.02)'],
        ['SKU error: repeat last week / 4-week weekday average', '87.5% / 76.0%', f'{pct1(fc.loc["sku", "naive7"])} / {pct1(fc.loc["sku", "wd4"])}'],
        ['Van error: repeat last week / 4-week weekday average', '53.3% / 43.9%', f'{pct1(fc.loc["van", "naive7"])} / {pct1(fc.loc["van", "wd4"])}'],
        ['Principal error: repeat last week / 4-week weekday average', '35.8% / 30.4%', f'{pct1(fc.loc["principal", "naive7"])} / {pct1(fc.loc["principal", "wd4"])}'],
        ['Next-best SKU: outlets judged / best-sellers / random', '4,097 / 3.95% / 0.66%', f'4,097 / {pct1(RT.loc["Best-sellers (popularity)", "test_p10"])} / {pct1(RT.loc["Random SKUs", "test_p10"])}'],
        ['Next-best SKU: the site scorer', '12.0%, 58.2% of outlets hit', f'{pct1(p10c)}, {pct1(hitc)} of outlets hit'],
        ['Cluster adoption: test adoption rate', '4.5% (242 of 5,335 pairs)', '5.7% (220 of 3,885 pairs; Personal Care left out)']]
story.append(table(base, [86 * mm, 40 * mm, 52 * mm], font_size=8))
story.append(Spacer(1, 2 * mm))
story += bullets([
    '<b>Chosen before the test.</b> Every model, weight and combination rule was picked on validation windows that end before the test starts '
    '(forecast: origins 28 Jul and 25 Aug; next-best SKU: picks at 30 Jun; churn: cut-offs 16 and 30 Jun; adoption: May - Jul). Combination rules were nested: '
    'members picked on one window, the rule scored on the next. The test was then run once.',
    '<b>Compared fairly.</b> Forecasts are compared at the same bias, because weekly error rewards forecasting low (section 4). Classifiers are compared on the '
    'same held-out outlets, with paired bootstrap intervals for the gain.',
    '<b>What we could not run.</b> The site\'s foundation models (Chronos-2, TimesFM 2.5, TiRex) cannot be downloaded in our environment. Our forecast pool is '
    'LightGBM (daily and weekly direct models) plus seven statistical methods (ETS, Theta, SES, ADIDA, IMAPA, TSB, SBA) on daily and weekly data and the two '
    'baselines, so our forecast "after" is a lower bound for what the '
    'site could reach by changing its combination rule.'])

# ---------------------------------------------------------------- forecast
story += H2('4. Demand forecast (Forecast proof, van loads)')
story.append(P('The site blends the best one to three models per series, picked by a backtest race (the ratio is not shown; the race leads are foundation '
               'models 31, LightGBM 30 and statistical 31 of 92 vans). We tested ten combination rules on the same 28 days.'))
story.append(img('fc_levels.png'))
fl = [['Level', 'Site today', 'Our copy of the site style', 'Best rule (picked on validation)', 'As fitted', 'Same bias as site', 'Verdict']]
for lv, nm, vd in [('sku', 'SKU (548)', 'Small boost'), ('van', 'Van (92)', 'Keep site'), ('principal', 'Principal (16)', 'Keep site'), ('cluster', 'Cluster (6)', 'Keep site')]:
    r_ = fc.loc[lv]
    fl.append([nm, f'{pct1(r_.site_best_run)} at {sb(r_.site_bias)} ({r_.site_run} run)', f'{pct1(r_.replica_site_style)} at {sb(r_.replica_bias)}',
               r_.picked_rule.replace(' (picked on validation)', ''), f'{pct1(r_.picked_wape)} at {sb(r_.picked_bias)}', pct1(r_.picked_at_site_bias), vd])
story.append(table(fl, [20 * mm, 30 * mm, 28 * mm, 34 * mm, 24 * mm, 20 * mm, 22 * mm], font_size=7.8))
story.append(Spacer(1, 2 * mm))
story += bullets([
    '<b>Composition.</b> At SKU level, racing models per SKU overfits: with two 4-week windows to choose from, the race picks whatever was lucky. Our copy of '
    f'that rule scored {pct1(fc.loc["sku", "replica_site_style"])}, worse than the 4-week average. One blend for all SKUs of the three models that were best '
    f'on validation ({fc.loc["sku", "level_top3_members"].replace("_w", " weekly").replace("_d", " daily").replace("LGBM", "LightGBM")}) scored '
    f'{pct1(fc.loc["sku", "picked_wape"])}.',
    '<b>Ratio.</b> Equal weights. Weights fitted on validation (LightGBM daily 57%, repeat-last-week 17%, Theta 10%, LightGBM weekly 7%, SBA 6%, ETS 3%) '
    'scored 63.4% against 63.0% for equal weights: fitted ratios do not beat a plain average with this little data.',
    '<b>Van, principal, cluster.</b> The site\'s blends, with foundation models, beat everything we could build by 0.6 to 1.4 points at the same bias. Keep them.',
    '<b>Company total.</b> The four test weeks were SAR 1.08M, 1.38M, 1.56M and 1.05M. Any flat forecast between 1.08M and 1.38M a week scores exactly 16.2%, '
    'because the misses cancel, so no model can be shown to beat another at total.'])
story.append(KeepTogether([img('sku_tradeoff.png'),
    P('Weekly error (WAPE) is lowest for a forecast below the mean, because SKU sales are lumpy: most weeks are small, a few are large. Both the site\'s runs and '
      'our blend forecast 17% to 33% too low. Held at the site\'s bias, our blend is 1.4 points better. Forced to zero bias it scores 74.1%: 15% less error than '
      'repeating last week (not 22%) and 3% less than the 4-week weekday average. For ordering, a forecast 17% too low means ordering 17% too little unless '
      'safety stock covers it, so the page should show the bias next to the error.', 'small')]))

# ---------------------------------------------------------------- next-best SKU
story += H2('5. Next-best SKU')
story.append(P('The site ranks SKUs an outlet has not bought by what similar outlets buy, neighbourhood and van penetration, and popularity. We added two '
               'stronger recommenders, EASE (a closed-form item model) and RP3beta (a graph walk), and two ways to combine them with the current scores.'))
story.append(img('nbs.png'))
nb = [['Model', 'Validation (picks 30 Jun)', 'Test: share of 10 picks bought', 'Test: top 5', 'Test: outlets with a hit']]
for k_, lbl in [('Best-sellers (popularity)', 'Best-sellers'), ('Site-style blend (similar baskets + van + popularity)', 'Copy of the site scorer'),
                ('EASE (item model, closed form)', 'EASE alone'), ('RP3beta (graph walk)', 'RP3beta alone'),
                ('Blend: EASE + RP3beta + site blend + van', 'EASE + RP3beta + current scores'), ('LightGBM ranker on all scores', 'LightGBM ranker (picked)')]:
    x = RT.loc[k_]
    val = rf['fair_blend_val'] if k_.startswith('Blend') else x.val_p10
    nb.append([lbl, pct1(val) + (' (weights from 2 Jun)' if k_.startswith('Blend') else ''), pct1(x.test_p10), pct1(x.test_p5), pct1(x.test_outlets_hit)])
story.append(table(nb, [52 * mm, 36 * mm, 34 * mm, 24 * mm, 32 * mm], font_size=8))
story.append(Spacer(1, 2 * mm))
story += bullets([
    f'<b>Picked fairly.</b> On validation the ranker scored {pct1(rf["ranker_val"])} and the blend {pct1(rf["fair_blend_val"])} (blend weights tuned on an earlier '
    f'cut-off, so both are out of sample). The ranker was picked; on the test it scored {pct1(p10a)}. The blend would have scored {pct1(RES["blend"]["test_p10"])}: '
    'the honest choice is not always the best one on the test.',
    f'<b>Gain.</b> +{100 * (p10a - p10c):.1f} points over our copy of the site scorer, 90% bootstrap interval +{100 * ci[0]:.1f} to +{100 * ci[1]:.1f}. '
    f'Expect the site\'s 12.0% to reach about {100 * (0.1198 + p10a - p10c):.1f}%.',
    f'<b>Ceiling.</b> In the test weeks the 4,097 outlets started buying 5.6 new SKUs each on average, so a perfect list would have 42% of its 10 picks bought. '
    'Models reach 12% to 13% because adoption depends on what the rep offers and what is on the van, which the files do not record.'])

# ---------------------------------------------------------------- churn
story += H2('6. Outlet churn warning')
story.append(P('Same protocol as our 05_Predictions churn model: will an outlet active in the last 13 weeks make no purchase in the next 8? 30% of outlets are '
               'held out; seven model types and every 2- and 3-model blend were compared on two validation folds, then the winner was tested once.'))
story.append(img('churn.png'))
cm = [['Model', 'Validation AUC (mean of 2)', 'Test AUC', 'Test: quiet outlets in riskiest 20%']]
for k_ in ['Extra trees', 'Random forest', 'Blend: all seven', 'Blend: LightGBM + Logistic regression', 'Logistic regression', 'CatBoost', 'LightGBM', 'XGBoost',
           'HistGradientBoosting', 'Recency rule']:
    x = CT.loc[k_]
    cm.append([('Days since last order' if k_ == 'Recency rule' else k_) + (' (picked)' if k_ == ch['pick'] else ''), f'{x.val_auc_mean:.3f}', f'{x.test_auc:.3f}',
               pct0(x.test_top20_capture)])
story.append(table(cm, [70 * mm, 38 * mm, 30 * mm, 40 * mm], font_size=8))
story.append(Spacer(1, 2 * mm))
story += bullets([
    f'<b>Gain.</b> Extra trees beats LightGBM by +{et_ci[1]:.3f} AUC (90% interval +{et_ci[0]:.3f} to +{et_ci[2]:.3f}) on {ch["n_test"]:,} unseen outlets, '
    f'{ch["pos"]} of which went quiet. LightGBM itself beats the days-since-last-order rule by +{ch["ci_rec"][1]:.3f} AUC, but its riskiest 20% catches no more '
    'quiet outlets than the rule; Extra trees does (59% vs 53%).',
    'The site\'s own churn label and test period differ from ours (its rule scores 0.68, ours 0.73), so its 0.77 is not directly comparable; expect a similar '
    'gain of about +0.02 to +0.03 AUC. New scores for every outlet as of 20 Oct are in data/churn_scores_v2.csv.'])

# ---------------------------------------------------------------- adoption
story.append(CondPageBreak(120 * mm))
story += H2('7. Cluster adoption')
story.append(P('Will an outlet start buying from a sales team it has never bought from? We trained on adoptions up to 31 Jul and tested on August - October '
               'adoptions at held-out outlets, for the five clusters sold under distributor 1000 (Personal Care sells under its own outlet codes, which the site '
               'merges by location and name).'))
story.append(img('adoption.png'))
am = [['Model', 'Validation: AUC within cluster', 'Test: AUC pooled', 'Test: AUC within cluster', 'Test: adoptions in top 10%']]
for k_ in [ad['pick'], 'XGBoost', 'Logistic regression', 'CatBoost', 'LightGBM', 'Channel popularity', 'What neighbours buy']:
    x = AT.loc[k_]
    am.append([k_ + (' (picked)' if k_ == ad['pick'] else ''), f'{x.val_auc_within_mean:.3f}', f'{x.test_auc_pooled:.3f}', f'{x.test_auc_within_mean:.3f}',
               pct0(x.test_top10_capture)])
story.append(table(am, [62 * mm, 32 * mm, 26 * mm, 30 * mm, 28 * mm], font_size=8))
story.append(Spacer(1, 2 * mm))
story.append(callout('<b>Check before the demo.</b> The site reports a pooled AUC of 0.90 and 65% of adoptions in the top 10%. With a split that only trains on '
                     f'adoptions before the test window, no model we tried gets above {AT.test_auc_pooled.max():.2f} pooled or {pct0(AT.test_top10_capture.max())} in the top 10%, and '
                     'LightGBM alone catches fewer adoptions in its top 10% than ranking by channel popularity. If the site model trained on August - October '
                     'adoptions of other outlets, it has seen the test period: re-run it on adoptions before 1 August. Within one cluster its own AUC is 0.67 to '
                     '0.86; show that next to the pooled 0.90, which mostly reflects that Unilever HPC adopts far more often than the other teams.', Wd, kind='warn'))

# ---------------------------------------------------------------- pages without a model + value
story += H2('8. Where a better model does not help')
story.append(P('Most of the SAR 3.45M rests on benchmarks, simulations and rules. A different algorithm cannot move these; only better data can.'))
nm = [['Page', 'What the figure is', 'What limits it'],
      ['Van sales (fleet SAR 1.39M)', 'Re-zoning simulation and top-quartile drops per van-day', 'Straight-line distances from outlet GPS; no van tracks or cost per van-day'],
      ['Warehouse labour, P&L (SAR 639K)', 'Benchmark against the best quarter of shifts, Monte Carlo range', 'No pay rates, roster or attendance'],
      ['Shift supervisor (198 sessions)', 'Rules plus an IsolationForest anomaly model', 'No attendance, so the flags cannot be scored at all'],
      ['Receivables, Collection list (SAR 174K)', 'Rules on a rebuilt ageing', 'No payment history to train a model on'],
      ['Returns (SAR 315K), price (SAR 397K)', 'Measured rates against cluster medians', 'Descriptive; promotion and trade-spend data would sharpen them'],
      ['Cluster harmonisation (SAR 321K)', 'Own-best neighbourhood benchmark', 'The lever does not use the adoption model\'s accuracy']]
story.append(table(nm, [46 * mm, 66 * mm, 66 * mm], font_size=8))
story.append(Spacer(1, 3 * mm))
vl = [['Lever (a year unless stated)', 'Site base', 'After', 'Basis'],
      ['ML forecast to safety stock', 'SAR 65.2K', f'SAR {LV["safety_same_bias"] / 1e3:.1f}K', f'Same bias. SAR {LV["safety_zero_bias"] / 1e3:.1f}K if forecasts must be unbiased'],
      ['Inventory cash released (one-off)', 'SAR 296K', f'SAR {LV["safety_oneoff_same_bias"] / 1e3:.0f}K', f'Same bias. SAR {LV["safety_oneoff_zero_bias"] / 1e3:.0f}K if unbiased'],
      ['Whitespace / next-best SKU', 'SAR 104.5K', f'SAR {LV["whitespace"] / 1e3:.1f}K', 'Scaled by top-5 precision (the list the value uses)'],
      ['Churn early-warning and win-back', 'SAR 47.3K', f'up to SAR {LV["churn_upper"] / 1e3:.1f}K', 'Upper bound: whole lever scaled by churners caught'],
      ['Annual total', 'SAR 3.45M', f'SAR {(3454548 + LV["total_change"]) / 1e6:.3f}M', 'All model upgrades together']]
story.append(table(vl, [56 * mm, 26 * mm, 30 * mm, 66 * mm], font_size=8))

# ---------------------------------------------------------------- data asks
story += H2('9. What more data would unlock')
da = [['Model or page', 'Data to ask for', 'What it would change'],
      ['Demand forecast', '12-24 months of daily sales', 'Learns Ramadan, Eid and the summer peak; allows test windows long enough to rank models'],
      ['Demand forecast', 'Promotion and trade-loading calendar', 'The mid-October burst (one SKU at about 24,000 cases in a week) becomes a known input'],
      ['Forecast, van loads', 'Stock on hand; van load / unload by SKU', 'Turns forecasts into orders and loads; separates lost sales from low demand'],
      ['Next-best SKU', 'Rep visit logs and what was offered', 'Separates "not offered" from "refused"; the gap between 12% and the 42% ceiling is mostly here'],
      ['Churn', 'Visit logs or geo-fence flags; closure reasons', 'Tells closed shops from quiet ones'],
      ['Cluster adoption', 'More months of sales', '242 test adoptions put about +/-0.03 around the AUC; per-cluster AUCs rest on a handful'],
      ['Collection list', 'Customer ledger with payment dates; real credit limits', 'Allows a payment-risk model; explains the SAR 336K unmatched'],
      ['Shift supervisor, labour', 'Attendance, pay rates, roster', 'Lets the badge flags be scored; moves SAR 639K from Simulated to Corrected']]
story.append(table(da, [36 * mm, 56 * mm, 86 * mm], font_size=8))

# ---------------------------------------------------------------- appendix
story.append(PageBreak())
story += H2('Appendix: files and definitions')
story.append(table([['File (06_Model_Review/data)', 'What it holds'],
                    ['before_after_figures.csv, before_after_figures.json', 'Section 2 as data: page, place, text on the site now, replacement text, type, basis, numeric before and after'],
                    ['forecast_by_level.csv', 'Per level: site result and bias, our baselines, our copy of the site rule, the picked rule, error at the site bias'],
                    ['forecast_all_rules.csv', 'Every model and combination rule: validation (25 Aug) and hold-out error, per level'],
                    ['next_best_sku_models.csv', 'Every recommender on the site\'s test: validation and test precision, top 5, outlets with a hit'],
                    ['churn_models.csv', 'Every churn model and blend: two validation folds, test AUC, riskiest-20% capture'],
                    ['adoption_models.csv', 'Every adoption model and blend: pooled and within-cluster AUC, top-10% capture'],
                    ['next_best_sku_picks_v2.csv', 'Ten picks per outlet as of 20 Oct 2026 from the picked ranker (SKUs sold somewhere in the last 4 weeks only)'],
                    ['churn_scores_v2.csv', 'Churn risk per outlet as of 20 Oct 2026: Extra trees (v2) beside the LightGBM version (v1)']],
                   [58 * mm, 120 * mm], font_size=8))
story.append(Spacer(1, 3 * mm))
story += bullets([
    '<b>Weekly error (WAPE)</b> = sum of absolute weekly misses / sum of actual weekly sales, over every series and the 4 test weeks.',
    '<b>Bias</b> = forecast total / actual total - 1. Negative means forecasting too low.',
    '<b>AUC</b> = how often the model ranks a case that happened (an outlet that went quiet, a cluster adopted) above one that did not; 0.5 is a coin toss.',
    '<b>Precision@10</b> = share of the 10 recommended SKUs that the outlet bought in the 8 weeks after the list was made.',
    '<b>90% interval</b> = paired bootstrap over outlets (2,000 resamples) of the gain against the current model.',
    'Scripts: 06_Model_Review/scripts (fc_race, fc_eval, reco_race, reco_final_v2, reco_picks_v2, churn_race, churn_final_v2, adoption_race, charts, '
    'data_changes, build_report, check_report), run in that order with the 05_Predictions work files.'])

doc = BZDoc(OUT, 'Binzagr PoV - Model review: before and after', 'Model review: before and after figures', 'Binzagr PoV model review')
build(doc, story)
print('wrote', OUT)