/Users/ronlevimacmini/Library/Python/3.9/lib/python/site-packages/statsmodels/base/model.py:607: ConvergenceWarning: Maximum Likelihood optimization failed to converge. Check mle_retvals
  warnings.warn("Maximum Likelihood optimization failed to "
/Users/ronlevimacmini/Library/Python/3.9/lib/python/site-packages/statsmodels/base/model.py:607: ConvergenceWarning: Maximum Likelihood optimization failed to converge. Check mle_retvals
  warnings.warn("Maximum Likelihood optimization failed to "
/Users/ronlevimacmini/Library/Python/3.9/lib/python/site-packages/statsmodels/base/model.py:595: HessianInversionWarning: Inverting hessian failed, no bse or cov_params available
  warnings.warn('Inverting hessian failed, no bse or cov_params '
/Users/ronlevimacmini/Library/Python/3.9/lib/python/site-packages/statsmodels/base/model.py:595: HessianInversionWarning: Inverting hessian failed, no bse or cov_params available
  warnings.warn('Inverting hessian failed, no bse or cov_params '
======================================================================== 
SECTION 1 - DATASET
allegations 1243 | files 727 | agencies 15
GBV: 92 allegations / 61 files / 14 agencies
concluded 2018+: all 61% GBV 88%
pathways (files): {'OI': 436, 'PC': 147, 'ID': 134, 'SH': 10}
======================================================================== 
SECTION 2 - MASTER COMPARISON (all rows with tests)
fired: GBV 35% vs rest 13% (p=1.2e-08)
job-level: GBV 87% vs rest 44% (p=5.1e-15)
susp days: GBV med 5 (n=43) vs 3 (n=282), MW p=4.3e-06
duration: GBV med 16 vs 12, MW p=0.0022
demotion/suspension without dismissal: GBV 48/92=52% vs rest 31% (p=7.7e-05)
criminal: GBV 50% vs rest 17% (p=1.2e-13)
no-charge: GBV 27% vs rest 4% (p=1.4e-17)
GBV share of no-charge: 25/76
no-charge GBV filed under: {'Discreditable Conduct': 16, 'Corrupt Practice': 9}
appealed: GBV 12% vs rest 15% (p=0.45)
trust: GBV 11% vs rest 13% (p=0.71)
welfare: GBV 5% vs rest 2% (p=0.17)
colleague target (NEUTRAL regex): GBV 27% vs rest 13% (p=0.00028)
biggest case: GBV 16/92=17% vs rest 26/1151, Fisher p=7.7e-09
---------------------------------------- 
S2 relationship table + DISJOINT test (referee M3)
  overlapping minor    :   6 alleg,  5 files, fired 83%
  overlapping partner  :  17 alleg, 12 files, fired 29%
  overlapping colleague:  25 alleg, 20 files, fired 36%
  overlapping via_work :  17 alleg,  2 files, fired 47%
  overlapping stranger :  10 alleg,  7 files, fired 50%
  DISJOINT (priority: minor>partner>colleague>via_work>stranger):
   {'colleague': {'n': 23, 'fired': 0.3}, 'partner': {'n': 17, 'fired': 0.29}, 'stranger': {'n': 1, 'fired': 0.0}, 'via_work': {'n': 14, 'fired': 0.43}}
  disjoint chi2(3)=1.3, p=0.73
  outside-GBV kind-of-misconduct chi2 p=1.4e-35
======================================================================== 
SECTION 3 - COUNTS + FAIR WINDOW (referee minor 7)
counts: kw19=19/13f | flag110=110 | gbv_set=92/61f (review-upheld 90 + deferred to OPCC label 2)
fair window (concluded 2016+): verified 85 vs labelled 19 = 4.5x  (all-years 90/19 = 4.7x)
label test: named 19 fired 47% | catch-all 38 fired 50%
filed under: {'Discreditable Conduct': 72, 'Corrupt Practice': 11, 'Neglect of Duty': 5, 'Public Trust Offence': 3, 'Abuse of Authority': 1}
vs OPCC label: 19 labelled alleg in 13 files | flagged step1 19 | kept 19 | declined 0
  labelled files represented: 13 of 13 | would have been declined on the text alone: 2
charged as Abuse of Authority (persons heading): force 89% | GBV 1/92 (p=5.7e-37)
charged as Abuse of Authority (persons heading): arrest 71% | GBV 1/92 (p=6.2e-22)
AoA availability (hand-read): {'public': 52, 'police': 39, 'none': 1}
  public-victim 52 alleg / 28 files | with duties nexus 37 / 16 files
  those 37 filed as: {'Discreditable Conduct': 22, 'Corrupt Practice': 11, 'Public Trust Offence': 2, 'Neglect of Duty': 1, 'Abuse of Authority': 1}
form of conduct (double-coded, n=92); kappa in CLAUDE.md, sheet in qualitative/:
  badge/position used to pursue     43 of 92
  sexual comments or messages       35 of 92
  physical contact or assault       35 of 92
  intimate images or video          20 of 92
  overlap: 37 carry >1 category, 5 carry none, 133 assignments total
no-charge probe validated (1243 of 1243 coded - FULL CENSUS, no extrapolation):
  precision 98% (41/42) | recall 54% (found 41, missed 35); blind re-code kappa 1.00 on this item
  CORRECTED: 76 no-charge allegations, 25 GBV vs 51 other (probe said 42 and 20)
  rate within GBV 27% vs rest 4% (probe said 22% vs 2%); ratio 6.1x, was 11.4x
  era: all 25 corrected GBV no-charge allegations decided 2018+, 0 before (probe: all 20)
  case level: GBV 16% vs rest 4%
  the probe undercounted EVERY category, not just GBV: corruption 3.5->8.0%, investigative neglect 1.4->6.3%, unlawful arrest 0->5.4%
  dendrogram now uses nocharge_coded; k=2 and k=3 partitions unchanged, k=4 moves GBV from its own branch into the deceit/corruption/off-duty family
---------------------------------------- 
CASE-LEVEL REPLICATION (unit check: 61 GBV cases vs 666 others)
  dismissal            GBV 31% vs rest 6% (p=1.1e-10)
  job-level            GBV 82% vs rest 34% (p=3.8e-13)
  criminal             GBV 38% vs rest 16% (p=6.2e-05)
  no-charge            GBV 16% vs rest 4% (p=0.00015)
  appealed             GBV 13% vs rest 11% (p=0.85)
  discredit heading    GBV 90% vs rest 36% (p=3.4e-16)
  abuse of authority   GBV 2% vs rest 14% (p=0.0039)
  cases with one allegation: 28 of 61; dropping the largest case: 30% vs 6%
  UNIT CHECK: dismissal status is uniform across all allegations in 222/242 (92%) multi-allegation files -> dismissal behaves as a file/officer-level event; case level is the conservative unit
  single-allegation cases only: GBV 5/28 = 18% vs rest 15/457 = 3% (p=0.0011)
  excluding all 65 files with any deceit allegation: GBV 13/53 = 25% vs rest 21/609 = 3% (p=2.2e-10)  -> the GBV gap does not run through deceit co-charging
======================================================================== 
SECTION 4 - SANCTIONS
  Msr_Suspension             all 380  gbv 44
  Msr_Written_Reprimand      all 324  gbv 10
  Msr_Training               all 232  gbv 26
  Msr_Dismissal              all 177  gbv 32
  Msr_Advice_Future_Conduct  all 154  gbv  1
  Msr_Verbal_Reprimand       all 135  gbv  1
  Msr_Close_Supervision      all  48  gbv  3
  Msr_Counselling            all  47  gbv  6
  Msr_Program_Activity       all  44  gbv  1
  Msr_Reduction_In_Rank      all  39  gbv  4
  Msr_Transfer_Reassignment  all  30  gbv  4
======================================================================== 
SECTION 5 - CONDUCT x OUTCOMES (primary label)
                                              fired   emp   rev    oi    n
Conduct_Primary                                                           
deceit_and_falsification                       57.0  90.0  18.0  73.0  122
gendered_and_sexualized_violence               35.0  87.0  11.0  85.0   99
corruption_and_misuse_of_position              17.0  61.0  14.0  87.0  113
off_duty_criminal_or_discreditable_conduct      9.0  69.0  19.0  92.0  137
improper_access_or_disclosure_of_information    8.0  38.0  14.0  75.0  110
investigative_neglect                           8.0  20.0   8.0  56.0  142
workplace_harassment_non_sexualized             8.0  41.0   8.0  35.0  111
firearms_handling                               6.0  17.0   0.0  83.0   69
insufficient_information                        6.0  51.0  16.0  37.0   51
property_and_equipment_misuse                   4.0  20.0   9.0  65.0   46
discourtesy                                     2.0  13.0  27.0  29.0   45
vehicle_operation                               2.0  11.0   2.0  66.0   62
excessive_or_unnecessary_force                  1.0  39.0  31.0  50.0   80
unlawful_arrest_detention_or_search             0.0  38.0  50.0  18.0   56
======================================================================== 
SECTION 6 - MODELS: FULL vs MINIMAL (referee M2)
  focal GBV, minimal dis: OR=3.2 [1.5-6.7] p=0.00193
  focal GBV, minimal emp: OR=7.8 [3.8-16.2] p=2.84e-08
  focal GBV, full    dis: OR=3.0 [1.3-6.7] p=0.00716
  focal GBV, full    emp: OR=7.4 [3.5-15.7] p=1.39e-07
  model B minimal deceit  : OR=22.5 [4.4-114.2]
  model B minimal gendered: OR=7.5 [1.6-34.6]
  model B full    deceit  : OR=14.7 [2.9-75.2]
  model B full    gendered: OR=8.4 [1.8-38.5]
======================================================================== 
SECTION 12 - DETECTION incl. UNSTATED (referee M5)
  det_criminal   all 210 (17%)  gbv  40
  det_public     all  79 (6%)  gbv   1
  det_dept       all  54 (4%)  gbv  17
  det_self       all  14 (1%)  gbv   0
  det_colleague  all   2 (0%)  gbv   0
  det_audit      all   1 (0%)  gbv   0
  NO ROUTE STATED: 933 (75%) | GBV 50
======================================================================== 
SECTION 16 - ROBUSTNESS (corrected)
A. per-definition complements (referee minor 1):
  kw19      : def 47% vs complement 14%
  gbv_set   : def 35% vs complement 13%
  flag110   : def 35% vs complement 12%
  review_only: def 36% vs complement 13%
B. era, decision year: pre18 GBV 9% (n=11) rest 10% | 2018+ GBV 38% rest 15%
B2. era, INCIDENT year (referee M6): pre18 GBV 30% (n=30) rest 15% | 2018+ GBV 36% (n=39) rest 9%
B3. deceit convergence check: deceit fired pre-2018 41% (n=32) | 2018+ 63% (GBV moved 9% -> 39%, B above)
B4. route by era (thesis coupling): GBV no-charge pre-2018 0/11 | 2018+ 25/81; criminal pre 2/11 | 2018+ 44/81
B5. era simple tests within GBV: dismissal 1/11 vs 31/81 (p=0.089); no-charge (p=0.032)
C0. injury check: 6 of 80 force stories mention hospital/serious injury; 0 dismissals among them
C. text-length stratification (referee M4):
  précis length vs job-level punishment, Spearman rho=0.05 (longer=harsher, confirms concern)
  short (GBV n=14): criminal 7% vs 4% (p=0.49) | no-charge 0% vs 1% (p=1)
  mid   (GBV n=47): criminal 55% vs 15% (p=3.4e-10) | no-charge 38% vs 6% (p=4.5e-12)
  long  (GBV n=31): criminal 61% vs 33% (p=0.0027) | no-charge 23% vs 7% (p=0.007)
  criminal ~ GBV controlling log(text length): OR=4.4, p=7.9e-10
  no-charge ~ GBV controlling log(text length): OR=7.1, p=3.6e-12
D. section 9 GBV contrasts, tested (referee minor 3):
  seriousness   : GBV 44% vs rest 27% (p=0.05)
  responsibility: GBV 15% vs rest 27% (p=0.18)
  proceeding    : GBV 59% vs rest 34% (p=0.01)
Done. Every number above feeds the HTML page.
