COMPUTED CHECKS, DIAGNOSTICS AND TESTS
Existing analysis only. No refitting or new tests.
Formal reference-test df are distinct from descriptive dispersion divisors. No global all-checks-passed criterion exists.

SUMMARY TABLE
Check | Scope | Statistic / df / p | Criterion and interpretation
Missingness, duplication and calendar consistency | 731 rows / 16 source columns | missingCells=0; duplicateRows=0; duplicateDates=0; missingCalendarDays=0; negativeCounts=0; weekdayDateMismatches=0; monthDateMismatches=0; yearDateMismatches=0; workingdayDefinitionMismatches=0; reference df=not applicable; p=not computed / not applicable | Each reported issue count is checked against zero. These are descriptive audit outputs, not an implemented global acceptance gate. No issues detected by these specific structural checks.
Source fingerprint | Original day.csv | sha256=a6bcf826782d3c0fbfdcbeead17cd0884185a0dafe8ff10cd48a874ee7ba18be; reference df=not applicable; p=not computed / not applicable | Exact SHA-256 equality; mismatch raises ValueError before fitting. Matches the expected source fingerprint.
Target-component leakage | 731 rows | targetComponentMismatches=0; matchingRows=731; excludedColumns=['casual', 'registered']; reference df=not applicable; p=not computed / not applicable | Check cnt == casual + registered on every row. Do not use either target component as a predictor. Identity holds on all 731 rows; both columns excluded.
Zero humidity | 731 rows; sensitivity n=730 | count=1; date=2011-03-10; humidityEffectPrimaryPct=-5.0505466165488251; humidityEffectWithoutFlagPct=-5.9445123972644289; reference df=not applicable; p=not computed / not applicable | Flag hum == 0; retain in primary fit and separately refit without this day. No threshold for accepting the sensitivity difference was specified. One flagged record retained; humidity effect changes from -5.0505% to -5.9445% per +10 percentage points.
Temperature redundancy | 731 rows | pearsonR=0.99170155322946363; reference df=not applicable; p=not computed / not applicable | No numerical redundancy cutoff was encoded. At r=0.9917015532294636, the analyst retained temp and excluded atemp. Near redundancy; atemp excluded.
Weather-category support | 731 rows | category1=463; category2=247; category3=21; category4=0; reference df=not applicable; p=not computed / not applicable | Count actual category support. Category 4 cannot be estimated because it has no observations; no universal minimum category size was declared. Sparse category 3 and absent category 4.
Full design-matrix rank | Full NB2 mean design, n=731 | rank=26; columns=26; reference df=not applicable; p=not computed / not applicable | rank(X) == number of mean-design columns (26). Recorded as a diagnostic; not used as an explicit assertion in the script. 26/26: full column rank at NumPy numerical tolerance.
NB2 numerical validity | Training, full-data and sensitivity NB2 fits | trainingConverged=True; fullModelConverged=True; capturedFitWarningCount=0; trainingAlpha=0.041187056331103783; fullModelAlpha=0.053486821254930809; reference df=not applicable; p=not computed / not applicable | Every NB2 fit must report converged=True, finite coefficients, finite standard errors and alpha > 0; otherwise RuntimeError. Final JSON export disallows NaN/Infinity. The NB2 stop rules did not trigger; no captured fit warnings.
Poisson Pearson dispersion | Training, n=547 | pearsonSumSquaresReconstructed=56893.287514124357; divisor=521; meanParameterCount=26; dispersion=109.20016797336729; alpha=0; reference df=not applicable; p=not computed / not applicable | Dispersion = sum(Pearson residual squared) / (n - 26). Training Poisson > 1 is one explicit NB2-choice condition. No calibrated pass band around 1 and no formal dispersion-test p-value were computed. Large extra variation relative to Poisson.
Negative Binomial (NB2) Pearson dispersion | Training, n=547 | pearsonSumSquaresReconstructed=465.7114453866497; divisor=521; meanParameterCount=26; dispersion=0.89387993356362705; alpha=0.041187056331103783; reference df=not applicable; p=not computed / not applicable | Dispersion = sum(Pearson residual squared) / (n - 26). Training Poisson > 1 is one explicit NB2-choice condition. No calibrated pass band around 1 and no formal dispersion-test p-value were computed. Extra-variance model summary; not a pass/fail result.
Negative Binomial (NB2) Pearson dispersion | Full association fit, n=731 | pearsonSumSquaresReconstructed=531.90654398929632; divisor=705; meanParameterCount=26; dispersion=0.75447736736070403; alpha=0.053486821254930809; reference df=not applicable; p=not computed / not applicable | Dispersion = sum(Pearson residual squared) / (n - 26). Training Poisson > 1 is one explicit NB2-choice condition. No calibrated pass band around 1 and no formal dispersion-test p-value were computed. Extra-variance model summary; not a pass/fail result.
Training AIC comparison | Same 547 rows and mean formula | poissonAIC=66197.62698196336; nb2AIC=8815.6517856835653; aicDifferencePoissonMinusNB2=57381.975196279796; poissonParameters=26; nb2Parameters=27; reference df=not applicable; p=not computed / not applicable | Choose planned NB2 only if training Poisson dispersion > 1 AND NB2 AIC < Poisson AIC; otherwise stop and review. The implemented selection condition holds.
Ljung-Box, lag 7 | Full NB2 Pearson residuals, n=731 | Q=251.91044086566615; maximumLag=7; modelDfAdjustment=0; reference df=7; p=1.0862846732672167e-50 | No pass/fail p-value threshold is encoded. statsmodels is called with model_df=0 (default), so chi-square reference df equals the lag. Small p is evidence of remaining residual dependence. Strong residual-dependence signal.
Ljung-Box, lag 14 | Full NB2 Pearson residuals, n=731 | Q=258.99118696116716; maximumLag=14; modelDfAdjustment=0; reference df=14; p=3.9574122927910813e-47 | No pass/fail p-value threshold is encoded. statsmodels is called with model_df=0 (default), so chi-square reference df equals the lag. Small p is evidence of remaining residual dependence. Strong residual-dependence signal.
Residual autocorrelation | Full NB2 Pearson residuals, n=731 | lag1ACF=0.38744187878904401; lag7ACF=0.087276716333449592; reference df=not applicable; p=not computed / not applicable | Report the computed ACF; no significance band or cutoff was calculated. Positive short-lag autocorrelation remains.
Chronological holdout errors | 184 dates, 2012-07-01 to 2012-12-31 | poissonMAE=688.60131154559849; poissonRMSE=927.85447364676452; nb2MAE=863.4496756573235; nb2RMSE=1122.7938832209572; reference df=not applicable; p=not computed / not applicable | Report MAE/RMSE without test tuning. No deployment threshold or paired model-comparison significance test was specified. Poisson has lower holdout MAE and RMSE than NB2.

SELECTED COEFFICIENT WALD TESTS (FULL MODEL N=731)
Standard-normal reference; df not applicable. Two-sided p; HAC7 pointwise intervals; no multiplicity adjustment. Not model-assumption acceptance tests.
Contrast | z | exact stored p | change % [95% HAC CI]
Temperature / Normalized temperature 0.50 → 0.60; other covariates held fixed | 5.121842683926757 | 3.0256427138339952e-07 | 7.8402495% [4.7699404, 11.000535]
Humidity / +10 percentage points in recorded relative humidity | -4.1628215333642631 | 3.1433904958171295e-05 | -5.0505466% [-7.3393502, -2.7052075]
Wind speed / +0.10 in normalized wind speed (6.7 original units; unit unspecified in archive) | -6.9924559547725664 | 2.7011532954971292e-12 | -9.272694% [-11.713945, -6.7639384]
Mist / cloudy weather / Weather category 2 versus clear/few clouds (category 1) | -3.8459155780326966 | 0.00012010311318663571 | -8.9716063% [-13.229418, -4.5048647]
Light rain / snow / Weather category 3 versus clear/few clouds (category 1) | -6.3000744578931931 | 2.9750269738225754e-10 | -50.367518% [-60.086642, -38.281734]
Holiday / Holiday versus non-holiday, with weekday held fixed | -2.39700329936898 | 0.016529774870995884 | -16.847518% [-28.491218, -3.3078852]

IMPORTANT INTERPRETATION
Ljung-Box uses model_df=0 by default; chi-square reference df=7 and14, without a fitted-NB2 parameter adjustment. Report as a descriptive residual-dependence reference check, not definitive specification validation.
Pearson dispersion is sum of squared Pearson residuals divided by n-26 (521 training /705 full). No formal dispersion p-value or pass band around1 was computed.
AIC and holdout MAE/RMSE have no computed p-values. The explicit NB2 choice checks training Poisson dispersion>1 and lower NB2 AIC. Poisson nonetheless has the smaller holdout errors.
No hard redundancy cutoff was encoded; r=.9917015532294636 supports an analyst decision to omit atemp. Full rank26/26 does not establish low collinearity.
HAC7 is a covariance estimator, not a test and not a cure for residual dependence.
Numerical convergence and finite coefficients establish numerical usability, not scientific validity.

SOURCE LOCATIONS
analyse.py: fit_models; model_row; comparison guard; coefficient; temp_contrast; acorr_ljungbox; quality payload; matrix_rank.
analysis.json: quality; model.comparison; model.fullData; model.coefficientRows; evaluation; diagnostics.
Existing9-page PDF generator: work/evidence-demo/build_report.py. No source/analysis/report/checkout edits in this extraction.
