Books, papers and documentation

statsmodels developers

Binomial/Poisson GLM APIs; Projects 003 and 007.

Checked 2026-10-11: statsmodels 0.15.0 page title; GLM families incl. Binomial and Poisson.

NIST (U.S. National Institute of Standards and Technology) and SEMATECH

Day 3 common distributions (section 1.3.6); later days may use its EDA and process-modeling chapters.

Checked 2026-10-11: Handbook landing page title on itl.nist.gov; section 1.3.6.6 'Gallery of Distributions' (eda366.htm) loads and lists the normal, lognormal, exponential, gamma and other continuous distributions plus the binomial and Poisson discrete distributions. No publication date shown.

statsmodels developers

Count-data regression APIs; Project 003.

Checked 2026-10-11: statsmodels 0.15.0 page title; lists count models incl. NegativeBinomial and ZeroInflatedPoisson.

statsmodels developers

APIs for power/MDE, multiple-comparison correction and binomial CIs; Projects 001, 005.

Checked 2026-10-11: statsmodels 0.15.0 page lists 'Power and Sample Size Calculations' (NormalIndPower, tt_ind_solve_power), multipletests and proportion_confint.

statsmodels developers

Checked 2026-10-12: Opened 2026-10-12. scale accepts 'X2', 'dev' or a float; the default is 1 for Binomial and Poisson; 'X2' estimates the scale as Pearson chi-square divided by df_resid and 'dev' as deviance divided by df_resid. The page does not itself call this a remedy for overdispersion; that interpretation is the standard quasi-likelihood use of an estimated scale.

statsmodels developers

Checked 2026-10-12: Opened 2026-10-12. Poisson zero-inflated model for a 1-d count response; exog_infl supplies explanatory variables for the binary inflation (mixing probability) model, a constant if omitted; inflation link logit by default or probit; exog needs an explicit constant; offset and exposure enter the Poisson mean part with coefficient fixed at 1.

statsmodels developers

Checked 2026-10-12: Opened 2026-10-12. offset is added to the linear predictor with its coefficient constrained to 1; exposure adds log(exposure) to the linear prediction with coefficient fixed at 1 and is only valid with the log link; scale is 1 for Binomial, NegativeBinomial and Poisson, and log-likelihood and deviance are not valid when variance weights make them quasi-likelihood.

LightGBM developers (Microsoft)

Practical rules for passing categorical features natively (integer codes, high-cardinality cautions) and how missing values are handled.

Checked 2026-10-11: Page title 'Advanced Topics' with sections Missing Value Handle and Categorical Feature Support on LightGBM docs (4.7.0.99, latest).

scikit-learn developers

Checked 2026-10-12: Opened 2026-10-12. States the one-standard-error rule as selecting the simplest model whose performance is within one standard error of the best model's, to avoid overfitting; implements it with a custom refit callable in GridSearchCV (code uses the standard deviation of test scores across 30 splits). The example's prose has an inconsistent 0.88 figure; the printed result selects 25 PCA components.

Eleanor Dillon, Jacob LaRiviere, Scott Lundberg, Jonathan Roth, Vasilis Syrgkanis

Checked 2026-10-11: Article in the SHAP docs: SHAP shows associations the model learned, not causal effects; unobserved confounders and redundant features distort or split credit; causal questions need causal designs (double ML for observed confounding, experiments or quasi-experiments otherwise).

Yandex / CatBoost developers

Reference for ordered target statistics on categorical features and the training parameters; pair with the CatBoost paper.

Checked 2026-10-11: Page title 'CatBoost' and description 'gradient boosting on decision trees' on catboost.ai/docs.

scikit-learn developers

Checked 2026-10-11: 2.3.2 k-means assumes convex isotropic clusters, needs n_clusters, inertia not normalized, local minima; 2.3.7 DBSCAN eps is crucial (too small -> noise, too large -> merged clusters); 2.3.11.5 silhouette coefficient; 2.3.11.1 adjusted Rand index needs ground truth.

imbalanced-learn developers

Checked 2026-10-11: imbalanced-learn 0.14.2: resampling the whole dataset before splitting or CV leaks and balances the test data; the imblearn Pipeline resamples only during fit so validation folds keep the true class distribution.

scikit-learn developers

Official guidance on leakage through preprocessing, fitting transformers on the full dataset, and controlling randomness; cite it for 'fit on train only, use a Pipeline'.

Checked 2026-10-11: Page title and sections 12.1 Inconsistent preprocessing, 12.2 Data leakage (12.2.1 How to avoid data leakage) on the stable docs (version 1.9.1).

scikit-learn developers

Checked 2026-10-12: Opened the stable User Guide page, grouped-data section: samples from the same group are dependent, so a flexible model can look good by recognising group-specific features; GroupKFold guarantees the same group is never in both the training and the testing set of a split.

scikit-learn developers

Reference for K-fold, stratified, group-aware splits and TimeSeriesSplit (expanding-window folds for temporal data).

Checked 2026-10-11: Page title and section 3.1.2.6.1 Time Series Split (TimeSeriesSplit) on the stable docs (version 1.9.1).

scikit-learn developers

Checked 2026-10-12: Opened 2026-10-12. Predictions are piecewise constant approximations and trees are not good at extrapolation (example: a tree fitted on x in 0..5 predicts the largest training target for x = 10); min_samples_leaf=5 suggested as an initial value, a very small value usually overfits though for classification with few classes 1 is often best; min_samples_leaf guarantees a minimum leaf size; predict_proba is the fraction of training samples of the class in the leaf; min_samples_split ignores sample_weight (use min_weight_fraction_leaf).

scikit-learn developers

Checked 2026-10-11: HistGradientBoosting sections: 1.11.1.1.1 early stopping enabled by default above 10,000 samples using validation_fraction; 1.11.1.1.2 native NaN routing learned per split; 1.11.1.1.4 native categorical support; 1.11.1.1.5 monotonic_cst per feature, constraints are marginal only, not supported for multiclass or categorical features.

scikit-learn developers

Hashing trick and character n-gram vectorizers.

Checked 2026-10-12: Section 8.2.2 Feature hashing: signed hash so collisions tend to cancel, alternate_sign default and use for n_features < 10000, power-of-two n_features advised because a modulo maps hash to column, no inverse_transform, MurmurHash3 32-bit. Section 8.2.3.8: character n-grams resilient against misspellings and derivations; char_wb vs char, char_wb less noisy for whitespace-separated languages.

LightGBM developers (Microsoft)

Explains histogram-based splits, leaf-wise growth and native optimal categorical splits; cite for why LightGBM is fast and how it treats categoricals.

Checked 2026-10-11: Page title and sections 'Leaf-wise (Best-first) Tree Growth' and 'Optimal Split for Categorical Features' on LightGBM docs (4.7.0.99, latest).

scikit-learn developers

Checked 2026-10-11: Section 8.4.6 Marking imputed values: MissingIndicator, add_indicator on SimpleImputer/IterativeImputer, missingness can be informative; 8.4.7 lists estimators that accept NaN natively, including HistGradientBoosting and trees.

scikit-learn developers

Checked 2026-10-11: 1.1.11 Logistic regression note: regularization is applied by default (common in ML, not in statistics); no regularization corresponds to a very large C; solver table marks sag/saga as not robust to unscaled data.

scikit-learn developers

Checked 2026-10-11: 3.4.6.1 R2 is dataset dependent and not meaningfully comparable across datasets, can be negative, constant mean model scores 0; 3.4.6.5 MAPE divides by max(epsilon, |y|) and returns a fraction; 3.4.8 dummy estimators as sanity baselines; 3.4.4.18 Brier score is strictly proper and mixes calibration, resolution and uncertainty; 3.4.4.4 balanced accuracy.

XGBoost developers

Checked 2026-10-11: monotone_constraints with 1/-1/0 per feature or keyed by feature name; with hist/approx the constraint can remove candidate splits and give shallow trees, raising max_bin may help.

scikit-learn developers

Checked 2026-10-11: Example explains that non-nested CV tunes and evaluates on the same data and is optimistically biased; nested CV wraps GridSearchCV in cross_val_score; cites Cawley and Talbot (2010).

XGBoost developers

Checked 2026-10-11: Handle Imbalanced Dataset: if only ranking/AUC matters, balance with scale_pos_weight; if predicted probabilities must be right, do not rebalance and set max_delta_step to a finite value; overfitting control section.

scikit-learn developers

Checked 2026-10-11: Outlier vs novelty detection; contamination parameter sets the threshold applied to raw scores (for LOF, a percentile of training scores); IsolationForest path-length scoring.

LightGBM developers (Microsoft)

Checked 2026-10-11: is_unbalance and scale_pos_weight entries warn that they can improve the overall metric but give poor estimates of individual class probabilities; monotone_constraints and monotone_constraints_method (basic over-constrains, intermediate, advanced); early_stopping_round and first_metric_only; min_data_per_group, cat_smooth, max_cat_to_onehot.

LightGBM developers (Microsoft)

Checked 2026-10-12: Opened 2026-10-12 (added by the insights-G reviewer; the claim in insight-1141 sits on this page, not on the Parameters page). Section 'Tune Parameters for the Leaf-wise (Best-first) Tree': setting num_leaves = 2^max_depth is called not good in practice because a leaf-wise tree is typically much deeper than a depth-wise tree for a fixed number of leaves; num_leaves should be smaller than 2^max_depth; min_data_in_leaf is 'a very important parameter to prevent over-fitting in a leaf-wise tree', a large value avoids growing too deep a tree but may underfit, and hundreds or thousands is enough for a large dataset; with max_depth set, also set num_leaves <= 2^max_depth.

scikit-learn developers

Checked 2026-10-11: PDP and ICE assume the features of interest are independent of the complement features and can create unrealistic points when correlated; ICE shows heterogeneity the PDP average hides; kind='both' and centered=True options.

scikit-learn developers

Shared record: also cited by project batches after their duplicate was removed.

Checked 2026-10-11: Warning that importances of a poorly scoring model are not meaningful; 5.2.2 impurity-based (MDI) importance is computed on training data and biased toward high-cardinality features; 5.2.3 correlated features lower each other's permutation importance, suggested fix is clustering correlated features; held-out vs training set computation.

scikit-learn developers

Pipeline and ColumnTransformer usage, so that preprocessing is fit inside cross-validation folds and applied per column type.

Checked 2026-10-11: Page title and section 8.1.4 ColumnTransformer for heterogeneous data on the stable docs (version 1.9.1).

scikit-learn developers

Scaling, sparse scaling, one-hot options, infrequent categories, polynomial and spline features.

Checked 2026-10-11: scikit-learn 1.9.1 page 8.3: 8.3.1 standardization note that RBF-kernel SVMs and l1/l2 regularizers of linear models assume centered, similarly scaled features; 8.3.1.2 centering sparse data destroys sparsity, MaxAbsScaler recommended, StandardScaler needs with_mean=False; 8.3.1.3 RobustScaler for data with outliers; 8.3.4 OneHotEncoder drop and collinearity in non-regularized regression, handle_unknown; 8.3.4.1 min_frequency / max_categories infrequent categories; 8.3.7 PolynomialFeatures interaction_only, SplineTransformer has no interactions.

scikit-learn developers

Reliability diagrams, sigmoid (Platt) versus isotonic calibration and CalibratedClassifierCV; cite for when predicted probabilities need to be trusted.

Checked 2026-10-11: Page title, section 1.16.1 Calibration curves and CalibratedClassifierCV on the stable docs (version 1.9.1).

scikit-learn developers

Checked 2026-10-11: scikit-learn 1.9.1 docs section 8.3.4.2: fit_transform uses cross fitting (cv folds, default 5); fit(X, y).transform(X) on the same data does not and can leak; shrinkage toward the global mean controlled by 'smooth'; unseen categories get the global target mean.

SHAP developers

Checked 2026-10-11: model_output defaults to 'raw' (log-odds for XGBoost binary classification); 'probability' requires feature_perturbation='interventional' with background data; tree_path_dependent uses leaf counts as the background.

scikit-learn developers

Checked 2026-10-12: Opened 2026-10-12. score_samples is documented as the opposite of the anomaly score defined in the original paper, so lower means more abnormal; decision_function = score_samples - offset_, negative values are outliers; with contamination='auto' offset_ is -0.5 as in the paper, with a float contamination offset_ is set so that the expected share of training samples with a negative decision function equals it.

scikit-learn developers

Checked 2026-10-12: discrete_features default 'auto' (False for dense X, True for sparse X); warning that treating continuous as discrete or vice versa will usually give incorrect results; nonparametric k-nearest-neighbour entropy estimation (Kraskov, Ross); n_neighbors default 3 (higher reduces variance, may add bias); random_state controls noise added to continuous variables to remove repeated values.

scikit-learn developers

Checked 2026-10-12: features parameter: 'missing-only' (default) covers only features with missing values at fit, 'all' covers every feature; error_on_new applies only with 'missing-only' and, if True, raises at transform for features with missing values that had none in fit.

scikit-learn developers

Checked 2026-10-12: Opened 2026-10-12. lasso_path computes the Lasso regularization path with coordinate descent and returns the alphas and the coefficients for every alpha (coefs of shape n_features x n_alphas); alphas are set automatically when None, with eps = alpha_min / alpha_max.

scikit-learn developers

Checked 2026-10-12: Opened 2026-10-12. AP summarises a precision-recall curve as the weighted mean of precisions at each threshold, weighted by the increase in recall from the previous threshold (sum over n of (R_n - R_{n-1}) P_n). The description says the implementation is not interpolated and differs from computing the area under the precision-recall curve with the trapezoidal rule, which uses linear interpolation and can be too optimistic.

scikit-learn developers

Checked 2026-10-12: Opened the page: Cohen's kappa measures agreement between two annotators on a classification problem, as (p_o - p_e) / (1 - p_e) with p_e the agreement expected by chance; range -1 to 1, 0 means no better than chance; symmetric in its two arguments.

scikit-learn developers

Checked 2026-10-12: Opened 2026-10-12. k restricts the ranking to the highest k scores; ties in y_score are averaged by default (ignore_ties=False); ignore_ties=True assumes no ties and gives wrong results if ties exist; docs example: ties averaged gives 0.75 against 0.5 with ignore_ties=True.

scikit-learn developers

Checked 2026-10-12: Opened 2026-10-12. Inputs are (latitude, longitude) pairs in radians, two columns; the result is the angular great-circle distance, and the page example converts degrees with math.radians for two airports and multiplies by an Earth radius of 6371 km to get kilometres (about 11,100 km between Buenos Aires and Paris in that example).

scikit-learn developers

Checked 2026-10-12: Opened 2026-10-12. groups gives group labels used while splitting and is only used with a Group cv instance such as GroupKFold; train_sizes are fractions of the maximum training-set size or absolute counts.

scikit-learn developers

Checked 2026-10-12: n_bins_ attribute: bins whose width is too small (<= 1e-8) are removed with a warning; strategy quantile, quantile_method, encode onehot / onehot-dense / ordinal, subsample default 200000.

scikit-learn developers

Checked 2026-10-12: clip parameter (default False, added 0.24) clips transformed held-out values to feature_range, inverse_transform may not restore data; note that clip=True does not prevent feature drift; note that MinMaxScaler does not reduce the effect of outliers but linearly scales them into a fixed range.

scikit-learn developers

Checked 2026-10-12: cv (default 5): for int/None, KFold if target_type continuous else StratifiedKFold; custom splitters or iterables allowed since 1.9 but each sample index must appear in a validation fold exactly once, otherwise ValueError; shuffle (default True, deprecated since 1.9); smooth default auto (empirical Bayes); fit_transform uses cross fitting and fit(X, y).transform(X) does not equal it; unseen categories get the target mean.

scikit-learn developers

Checked 2026-10-12: Opened 2026-10-12. class_weight: 'Set the parameter C of class i to class_weight[i]*C'; 'balanced' uses n_samples / (n_classes * np.bincount(y)); fit(sample_weight) rescales C per sample; regularization strength is inversely proportional to C.

scikit-learn developers

Checked 2026-10-12: Opened 2026-10-12. Fold-wise scores of models evaluated on the same 10x10-fold partitions are correlated; treating them as independent underestimates the variance of the test statistic and inflates false positives (Dietterich 1998); corrected resampled t-test of Nadeau and Bengio multiplies the naive variance term 1/(k*r) by (1/(k*r) + n_test/n_train); the worked example gives an uncorrected t of 2.611 (p = 0.005) against a corrected t of 0.750 (p = 0.227).

statsmodels developers

Checked 2026-10-12: Opened 2026-10-12. The Binomial family accepts a 1d array of proportions as endog, with the number of trials supplied through var_weights; the default link is the logit link. The page does not itself discuss fractional-response inference.

statsmodels developers

Checked 2026-10-12: Opened 2026-10-12. Ordinal model based on a logistic or normal distribution; the logistic case corresponds to the proportional odds model; the endogenous variable is assumed ordered but its labels have no numeric interpretation besides the ordering; a latent linear variable is observed only through a discretisation at cut points; K levels need K-1 thresholds besides the slope coefficients.

scikit-learn developers

Post-hoc threshold choice from a business cost; Projects 008, 010, 012.

Checked 2026-10-11: Page title in scikit-learn 1.9.1 docs; documents TunedThresholdClassifierCV and FixedThresholdClassifier and links a cost-sensitive learning example.

scikit-learn developers

Checked 2026-10-12: Opened 2026-10-12. 3.2.2: randomized search budget is chosen independent of the number of parameters, adding parameters that do not influence performance does not decrease efficiency; loguniform is the continuous version of a log-spaced parameter; 3.2.3: successive halving (HalvingGridSearchCV, HalvingRandomSearchCV) is experimental, candidates start with few resources and only the best survive, min_resources should be high if many samples are needed to tell good from bad candidates.

scikit-learn developers

Checked 2026-10-11: 3.5.1 validation curve under/overfitting patterns and the warning that a validation score used for tuning is biased; 3.5.2 learning curve: more data helps while a train-validation gap remains open, not when both converge to a low score.

XGBoost developers (DMLC)

Reference for parameters, early stopping, categorical support and the Python API.

Checked 2026-10-11: Page title and description 'optimized distributed gradient boosting library' on XGBoost 3.4.2 stable docs.

XGBoost contributors

Checked 2026-10-12: Opened 2026-10-12. min_child_weight (default 1) is the minimum sum of instance weight (hessian) needed in a child; the page notes that in linear regression this corresponds to a minimum number of instances per node; larger values make the algorithm more conservative. max_cat_to_onehot: below the threshold one-hot-based splits, otherwise categories are partitioned.

Nixtla

Python reconciliation library used in the hierarchical-forecasting project.

Checked 2026-10-11: Docs landing page: lists BottomUp, TopDown, MiddleOut, MinTrace, ERM, probabilistic methods (Normality, Bootstrap, PERMBU, Conformal), temporal hierarchies; Apache 2.0 license section.

Nixtla

Global ML forecasting with leakage-safe lag construction per series; reference for entity-safe lags.

Checked 2026-10-11: Docs landing page: lag features and lag transforms (rolling/expanding), date features, target transforms (e.g. differences) reversed at prediction, static and exogenous features, scikit-learn-compatible regressors, Dask/Ray/Spark scaling, conformal prediction intervals.

scikit-learn developers

Gap/embargo between training and test windows; Day 25.

Checked 2026-10-11: API page documents gap (samples excluded from the end of each training set before the test set, added in 0.24), test_size and max_train_size, with a worked example.

Nixtla

Day 23 forecasting API used in the lesson code (forecast and cross_validation).

Checked 2026-10-11: Opened the core API page: StatsForecast(models=[...], freq=..., n_jobs=..., verbose=...); forecast(h, df, X_df, level, fitted, prediction_intervals, id_col, time_col, target_col); cross_validation(h, df, n_windows, step_size, test_size, input_size, ...); input needs unique_id, ds and y columns; output has one column per model.

Nixtla

Fast statistical baselines (AutoARIMA, ETS, Theta, seasonal naive) and built-in time-series cross-validation; the library used in FPP Pythonic Way.

Checked 2026-10-11: Page title 'Statistical Forecast - Nixtla' with StatsForecast getting-started, cross-validation and AutoARIMA/ETS tutorial navigation on nixtlaverse.nixtla.io.

Nixtla

Day 23 model names and arguments (season_length) used in the lesson code.

Checked 2026-10-11: Opened the models page (first 100,000 characters): AutoETS(season_length=1, model='ZZZ', damped=None, ...), AutoARIMA(..., seasonal=True, season_length=1, ...), SeasonalNaive(season_length), Naive, RandomWalkWithDrift, HistoricAverage, WindowAverage(window_size), SeasonalWindowAverage; 'ZZZ' lets AutoETS search the error, trend and seasonal components; both AutoETS and AutoARIMA select by AICc by default (AutoARIMA exposes ic='aicc').

statsmodels developers

Day 23 likelihood-based ETS with AIC/AICc, for the AIC-is-not-accuracy demonstration.

Checked 2026-10-11: Opened the ETSModel page (signature ETSModel(endog, error='add', trend=None, damped_trend=False, seasonal=None, seasonal_periods=None, initialization_method='estimated', ...), fit by maximising the log-likelihood) and the linked ETSResults page (attributes aic, aicc, bic, llf, resid; methods forecast, simulate, get_prediction).

statsmodels developers

Day 22 Ljung-Box test of white noise and Day 23 residual checks; remember model_df on fitted-model residuals.

Checked 2026-10-11: Opened the stable-docs page: signature acorr_ljungbox(x, lags=None, boxpierce=False, model_df=0, period=None, return_df=True, auto_lag=False); default lags min(10, nobs // 5); model_df subtracts estimated ARMA parameters from the degrees of freedom; period sets max lag to min(2 * period, nobs // 5); returns a DataFrame with lb_stat and lb_pvalue.

statsmodels developers

Day 22 STL decomposition and seasonal adjustment in Python.

Checked 2026-10-11: Opened the stable-docs page: signature STL(endog, period=None, seasonal=7, trend=None, low_pass=None, seasonal_deg=1, trend_deg=1, low_pass_deg=1, robust=False, seasonal_jump=1, trend_jump=1, low_pass_jump=1); a NumPy array needs an explicit period; robust=True uses a weighted fit less sensitive to outliers; fit() returns a DecomposeResult.

statsmodels developers

Day 22 and Day 23: seasonal ARIMA fit and residual extraction in Python.

Checked 2026-10-11: Opened the SARIMAX class page (signature SARIMAX(endog, exog=None, order=(1, 0, 0), seasonal_order=(0, 0, 0, 0), trend=None, ...); seasonal_order = (P, D, Q, s); simple_differencing discards the first s*D + d rows), the SARIMAX.fit page (documented disp, maxiter, method arguments) and the SARIMAXResults page (attributes resid, aic, aicc, bic, llf; methods forecast, get_forecast, predict).

statsmodels developers

Day 22 sample ACF; note that the confidence intervals use Bartlett's formula, not the flat 1.96/sqrt(n) white-noise band.

Checked 2026-10-11: Opened the stable-docs page: signature acf(x, adjusted=False, nlags=None, qstat=False, fft=True, alpha=None, bartlett_confint=True, missing='none', *, result_object=None); output includes lag 0; confidence intervals use Bartlett's formula; default nlags formula min(10 * log10(nobs), nobs - 1).

statsmodels developers

Day 21 and Day 22: unit-root test paired with KPSS.

Checked 2026-10-11: Opened the stable-docs page: signature adfuller(x, maxlag=None, regression='c', autolag='AIC', store=False, regresults=False, *, result_object=None); null hypothesis of a unit root; regression options c, ct, ctt, n; autolag options AIC, BIC, t-stat, None; ADFullerResult with statistic, pvalue, lags, nobs, critical_values; tuple default scheduled to change.

statsmodels developers

Day 21 stationarity checks; use result_object=True to avoid the deprecated tuple return.

Checked 2026-10-11: Opened the stable-docs page: signature kpss(x, regression='c', nlags='auto', store=False, *, result_object=None); null hypothesis that the series is level or trend stationary; regression 'c' or 'ct'; p-values interpolated from the Kwiatkowski et al. table and a boundary value returned when the statistic is outside it; the default tuple return is deprecated in favour of KPSSResult.

statsmodels developers

Reference for ARIMA/SARIMAX, autocorrelation and stationarity tests, decomposition and exponential smoothing in Python.

Checked 2026-10-11: Page title and API listing (AutoReg, ARIMA, stattools, X-13 helpers) on statsmodels 0.15.0 stable docs.

PyTorch

L2 vs decoupled weight decay (days 33, 34).

Checked 2026-10-11: Stable URL redirects to the PyTorch 2.14 page "Adam". weight_decay is documented as weight decay (L2 penalty), default 0; decoupled_weight_decay=True makes it equivalent to AdamW.

PyTorch

AdamW defaults and parameter groups (days 33, 34, 40).

Checked 2026-10-11: Stable URL redirects to the PyTorch 2.14 page "AdamW". weight_decay coefficient default 1e-2; refers to Decoupled Weight Decay Regularization; params may be an iterable of dicts defining parameter groups.

PyTorch

How the dynamic computation graph records operations and saves tensors for backward; supports day 33.

Checked 2026-10-11: Stable URL redirects to the PyTorch 2.14 notes page titled "Autograd mechanics".

PyTorch

API reference for backward(), grad and gradient tracking; supports days 33 and 35 (autograd in the NN lab).

Checked 2026-10-11: Stable URL redirects to the PyTorch 2.14 page titled "Automatic differentiation package - torch.autograd".

PyTorch

Unscale before gradient clipping under AMP (day 34).

Checked 2026-10-11: Stable URL redirects to the PyTorch 2.14 notes page. Gradient clipping section states that clipping without unscaling uses scaled gradients, which invalidates the threshold, and shows scaler.unscale_(optimizer) before clipping.

PyTorch

Mixed precision caveats and gradient scaling (days 34, 40, 54).

Checked 2026-10-11: Stable URL redirects to the PyTorch 2.14 page. Gradient Scaling section states small float16 gradients can underflow to zero and that gradient scaling multiplies the loss by a scale factor to prevent this.

PyTorch

Batch statistics vs running statistics (days 34, 36, 40).

Checked 2026-10-11: Stable URL redirects to the PyTorch 2.14 page "BatchNorm2d". By default the layer keeps running estimates of mean and variance during training (momentum 0.1) and uses them for normalization during evaluation.

PyTorch

Numerically stable binary and multi-label losses (day 32).

Checked 2026-10-11: Stable URL redirects to the PyTorch 2.14 page "BCEWithLogitsLoss". Page states the class combines Sigmoid and BCELoss and is more numerically stable than a plain Sigmoid followed by BCELoss because it uses the log-sum-exp trick; documents pos_weight.

PyTorch

Shape bugs that do not raise errors (day 31).

Checked 2026-10-11: Stable URL redirects to the PyTorch 2.14 notes page. States PyTorch follows NumPy broadcasting; tensors are broadcastable when, from the trailing dimension, sizes are equal, one of them is 1, or one does not exist; expansion happens without copying data.

PyTorch

CNN shape and parameter arithmetic (day 36).

Checked 2026-10-11: Stable URL redirects to the PyTorch 2.14 page "Conv2d". Gives H_out = floor((H_in + 2*padding - dilation*(kernel_size - 1) - 1)/stride + 1) and weight shape (out_channels, in_channels/groups, kernel_size[0], kernel_size[1]).

PyTorch

Logits vs probabilities at the loss boundary; class weights for imbalance (days 32, 41).

Checked 2026-10-11: Stable URL redirects to the PyTorch 2.14 page "CrossEntropyLoss". Page states the input is expected to contain unnormalized logits per class, that the optional class weight tensor is useful for an unbalanced training set, that ignore_index targets do not contribute to the gradient, and that label_smoothing mixes targets with a uniform distribution.

PyTorch

GPU memory measurement and budgeting (days 35, 40, 54).

Checked 2026-10-11: Stable URL redirects to the PyTorch 2.14 notes page. Memory management section states PyTorch uses a caching allocator whose unused memory still appears as used in nvidia-smi, and names memory_allocated, max_memory_allocated, memory_reserved and max_memory_reserved.

PyTorch

Dropout at inference (day 34).

Checked 2026-10-11: Stable URL redirects to the PyTorch 2.14 page "Dropout". During training elements are zeroed with probability p and outputs scaled by 1/(1-p); during evaluation the module computes an identity function.

PyTorch

Parameter counting (day 31).

Checked 2026-10-11: Stable URL redirects to the PyTorch 2.14 page "Linear". Weight has shape (out_features, in_features); optional bias of shape (out_features).

PyTorch

Train vs eval mode (day 34).

Checked 2026-10-11: Stable URL redirects to the PyTorch 2.14 page "Module". eval() sets evaluation mode, affects only certain modules such as Dropout and BatchNorm, and is equivalent to train(False).

PyTorch

Checked 2026-10-12: Stable URL redirects to the 2.14 page. Page states that, unlike a plain Python list, a ModuleList registers the submodules it holds so every Module method can see them.

PyTorch

Padding masks in attention (days 39, 41).

Checked 2026-10-11: Stable URL redirects to the PyTorch 2.14 page. key_padding_mask of shape (N, S) marks key elements to ignore for attention (padding).

PyTorch

Checked 2026-10-12: Page text: a general checkpoint for resuming training stores epoch, model state_dict, optimizer state_dict and loss; the optimizer state_dict is saved because it holds buffers and parameters updated during training; call model.train() to resume training and model.eval() before inference.

PyTorch

Checked 2026-10-12: Stable URL redirects to the 2.14 page. Page states that mean and standard deviation are computed over all mini-batches of the same process group, supports DistributedDataParallel with one GPU per process, uses convert_sync_batchnorm before wrapping in DDP, and synchronizes only in training mode.

PyTorch

Gradient checking of custom operations (days 33, 35).

Checked 2026-10-11: Stable URL redirects to the PyTorch 2.14 page. Notes that default tolerances are designed for double-precision inputs, that the check will likely fail for lower precision, and that it may fail at non-differentiable points.

PyTorch

Reference for layers, activations, losses and normalization modules; supports days 31-35.

Checked 2026-10-11: Stable URL redirects to the PyTorch 2.14 page titled "torch.nn".

PyTorch

Checked 2026-10-12: PyTorch 2.14 page: padding_idx entries receive no gradient, so the vector is not updated during training; it starts as zeros and can be set to other values.

PyTorch

Checked 2026-10-12: PyTorch 2.14 page: a boolean attn_mask value of True means the element takes part in attention, the inverse of MultiheadAttention's key_padding_mask where True marks padding to ignore (the page says to invert with ~mask when migrating); the reference code fills positions where a boolean mask is False with negative infinity before the softmax. The page does not discuss fully masked rows or NaN. Marked beta.

PyTorch

Exploding gradients (days 34, 37).

Checked 2026-10-11: Stable URL redirects to the PyTorch 2.14 page. Clips the gradient norm of an iterable of parameters, computed as if all gradients were concatenated into one vector; modifies gradients in place.

PyTorch

Checked 2026-10-12: PyTorch 2.14 page: takes the true lengths of the sequences in the batch and returns a PackedSequence without the padding; enforce_sorted=True requires longest-to-shortest order, False sorts internally (True only needed for ONNX export). The page says nothing about the final hidden state.

PyTorch

Checked 2026-10-12: Stable URL redirects to the 2.14 page. Learning-rate scheduler section: scheduler.step() is expected after optimizer.step(); calling it first skips the first value of the schedule (behaviour changed in 1.1.0). swa_utils section: AveragedModel (SWA and EMA), SWALR, and update_bn, which recomputes BatchNorm activation statistics with one pass over a loader at the end of SWA/EMA training. Optimizer state_dict/load_state_dict exist.

PyTorch

Activation memory reduction (days 34, 40, 44).

Checked 2026-10-11: Stable URL redirects to the PyTorch 2.14 page. Describes activation checkpointing as trading compute for memory by not keeping forward tensors alive until backward and recomputing them.

PyTorch

Data loader seeding and throughput (day 35).

Checked 2026-10-11: Stable URL redirects to the PyTorch 2.14 page. "Randomness in multi-process data loading" states each worker's PyTorch seed is base_seed + worker_id but seeds for other libraries may be duplicated across workers; worker_init_fn can seed them. Memory Pinning section describes pin_memory=True.

PyTorch / torchvision maintainers

Checked 2026-10-11: Docs page: ResNet18_Weights.IMAGENET1K_V1 (alias DEFAULT); ImageNet-1K acc@1 69.758 / acc@5 89.078; 11,689,512 parameters; 1.81 GFLOPS; weight file 44.7 MB; inference transforms resize 256, centre crop 224, rescale to [0,1], normalise with ImageNet mean/std. No licence terms for the weights are stated on the page.

Keras team

Checked 2026-10-12: Guide text: BatchNormalization has non-trainable running mean and variance; setting trainable to False puts the layer in inference mode so statistics stop updating; when unfreezing a base model for fine-tuning, call it with training=False so BatchNorm statistics are not updated and the learned features are not disturbed.

Hugging Face

Models, even those fine-tuned from the same base, expect different chat formats and control tokens; wrong control tokens can make performance drastically worse. Supports insight 0508.

Checked 2026-10-11: Live docs page (links to v5.19.0 docs): text on per-model chat formats and control tokens, apply_chat_template, add_generation_prompt and the warning about duplicated special tokens. No publication date shown.

Anthropic

At check time the temperature description says results are not fully deterministic even at 0.0; top_p and top_k are marked for advanced use; the page also marks these sampling parameters as deprecated (restricted) for newer models. Vendor behaviour changes: re-check before citing specifics. Supports insight 0510.

Checked 2026-10-11: Messages API reference page: temperature, top_p and top_k parameter descriptions and the stop_reason enum. No publication date shown.

scikit-learn developers

TruncatedSVD on a tf-idf matrix is LSA; used in the Day 41 application code.

Checked 2026-10-11: Opened the User Guide page: the section is numbered 2.5.3 under 2.5 'Decomposing signals in components'; it states that TruncatedSVD differs from PCA in not requiring a centred matrix, that applying it to tf-idf matrices (rather than raw counts) is recommended for LSA, and suggests sublinear_tf and use_idf. No publication date shown.

Anthropic

Registered here during verification for the reasoning-token claims; also cited by another batch (insight 1639).

Checked 2026-10-12: Page opened 2026-10-12 (plus the Thinking overview page and the Messages API reference). Thinking tokens are billed as output tokens even when thinking text is not returned and count toward max_tokens alongside the response text; budget_tokens must be below max_tokens (except interleaved thinking); the budget is a target not a strict cap; usage.output_tokens_details.thinking_tokens reports the reasoning tokens; adaptive thinking may skip thinking on easy inputs at lower effort. Models 4.7 and later reject type enabled with a 400.

Hugging Face

Checked 2026-10-12: Opened the page (v5.19.0 rendering). Attention mask section: for the BERT tokenizer 1 marks tokens that should be attended to and 0 marks padded values; position IDs section: transformers are unaware of token position without them, and absolute positional embeddings are selected in [0, config.max_position_embeddings - 1].

Anthropic

Vendor API rule as read on 2026-10-12; other providers use different message formats and rules.

Checked 2026-10-12: Page opened 2026-10-12 (as read). Each tool_use block has an id used to match the later tool_result via tool_use_id; tool_result blocks must immediately follow their tool_use blocks in the history with no messages between, and must come first in the user message's content array; a history that breaks this fails with an error such as 'tool_use ids were found without tool_result blocks immediately after'.

Hugging Face

Free course on Transformers, tokenizers and fine-tuning; optional reading for days 41-44.

Checked 2026-10-11: Chapter 1 introduction page loads (title "Introduction · Hugging Face"; page text labels it "LLM Course documentation").

Liang Wang et al. (intfloat)

One concrete example of per-model input formatting and silent truncation; other embedding models have different rules, so read each model's card.

Checked 2026-10-12: Model card opened. Each input should start with 'query: ' or 'passage: '; the FAQ says the prefix is required because of how the model is trained and omitting it degrades performance; 'passage: ' for the retrieved side of asymmetric retrieval; long texts are truncated to at most 512 tokens.

JSON Schema

The schema language used for tool inputs and structured outputs; supports days 47 and 49 (validation of model output).

Checked 2026-10-11: Page loads ("JSON Schema - Specification") and links the 2020-12 Core and Validation documents.

Elastic

Per-language tokenisation and stemming for the lexical branch of hybrid retrieval.

Checked 2026-10-12: Page opened. Language analyzers are built-in analyzers aimed at specific languages; the list includes a turkish analyzer; most support custom stop-word and stem-exclusion lists and the page shows how to rebuild each as a custom analyzer.

Hugging Face

merge_and_unload folds LoRA weights into the base model to remove extra inference latency; adapters can also be loaded and switched separately on one base; merging support depends on the quantization backend. Supports insight 0550.

Checked 2026-10-11: PEFT developer guide page 'LoRA' (current docs): merge_and_unload, merge_adapter/unmerge_adapter, load_adapter with adapter_name, set_adapter, mixed-adapter batches. Linked PEFT Quantization guide checked for backend-specific merge restrictions (e.g., AQLM, torchao, INC). No publication date shown.

Sentence Transformers contributors

Checked 2026-10-12: Opened the losses reference 2026-10-12 (page text read in two chunks; the MultipleNegativesRankingLoss section is in the second). The loss takes anchor-positive pairs (or triplets, n-tuples) and pushes each anchor away from the other items in the batch (in-batch negatives; also called InfoNCE or cross-entropy with in-batch negatives); the section warns that a batch containing duplicates of the anchor or positive would treat them as incorrect negatives and advises the NO_DUPLICATES batch sampler.

Anthropic

Vendor documentation that changes with model releases. Supports insights 1510 and 1516.

Checked 2026-10-12: Reference page opened 2026-10-12. temperature: marked Deprecated, models released after Claude Opus 4.6 do not support setting temperature, a value of 1.0 is accepted for compatibility and all other values are rejected with a 400; range 0.0 to 1.0, closer to 0.0 for analytical or multiple choice and closer to 1.0 for creative tasks; even at 0.0 results are not fully deterministic. top_p: marked Deprecated, values at or above 0.99 accepted on those models and all others rejected with a 400. top_k: marked Deprecated, any value rejected on those models. stop_details with category reasoning_extraction describes requests to reproduce internal reasoning in the response text.

Qdrant

Vendor operations guide for embedding-model changes; supports day 48 and 52.

Checked 2026-10-12: Page opened. States that switching models requires re-embedding all vectors in a collection; warns that a new embedding model is not automatically better on your data, so measure it before moving traffic; describes a blue-green migration with a second collection, dual writes, a background re-embed of existing points and an atomic collection-alias switch. The page does not itself state that vectors from different models are not comparable.

OpenAI

Frames two axes: context optimisation (e.g., RAG) for missing, outdated or proprietary knowledge and LLM optimisation (fine-tuning) for consistent behaviour such as format, tone and reasoning; recommends starting with prompt engineering plus an evaluation set; techniques are additive. Supports insights 0539 and 0541.

Checked 2026-10-11: platform.openai.com/docs/guides/optimizing-llm-accuracy 301-redirects to the developers.openai.com page; title and body text checked. No date shown.

Hugging Face

Checked 2026-10-12: Opened the main-branch page. truncation=True or longest_first removes tokens one at a time from the longest sequence of a pair; only_first and only_second truncate just that sentence of a pair; False applies no truncation (the default); max_length defaults to the model's maximum input length.

Anthropic

Checked 2026-10-12: Opened 2026-10-12. Claude may emit several tool calls in one turn; the application chooses concurrent or sequential execution; independent read-only operations are usually safe in parallel while tools with side effects, shared state or ordering needs may be better sequential; one tool_result must be returned for every tool_use, and a call that was not run should be returned with is_error true and an explanation.

Hugging Face

Library docs for LoRA and other parameter-efficient methods; supports day 44 implementation details.

Checked 2026-10-11: Page loads with title "PEFT · Hugging Face".

LangChain

Framework documentation for one library; used to show that checkpoint-and-resume is a supported pattern, not as a general standard.

Checked 2026-10-12: Opened 2026-10-12. Checkpointers save a thread's graph state as checkpoints and are listed as enabling conversation continuity, human-in-the-loop workflows, time travel and fault tolerance; state is scoped by thread_id; an in-memory checkpointer loses data on restart. Details of super-steps and pending writes are on a separate page not read.

Sentence Transformers (UKP Lab / Hugging Face maintainers)

Choosing cosine vs dot product per embedding model (days 38, 48).

Checked 2026-10-11: Page title "Pretrained Models - Sentence Transformers documentation". States some multi-qa and msmarco models were tuned for dot-product scoring, while others produce normalized length-1 vectors usable with dot product, cosine similarity or Euclidean distance.

Anthropic

Vendor guidance on success criteria, evals before prompt iteration, and prompt techniques; supports day 47.

Checked 2026-10-11: docs.claude.com URL redirects to platform.claude.com page titled "Prompt engineering overview - Claude Platform Docs".

Anthropic

Vendor-reported figure; supports insight 1505.

Checked 2026-10-12: Page opened 2026-10-12 (the older long-context-tips URL redirects here). Long-context section: place long documents and inputs near the top of the prompt, above the query, instructions and examples; a note says queries at the end can improve response quality by up to 30 percent in tests, especially with complex multi-document inputs. The page does not publish the test setup. Page also states, for batched tool calls, the system prompt line 'Only batch tool calls that are independent of each other' (see the parallel tool use page).

Nils Reimers and contributors

Cost structure of reranking: one forward pass per candidate pair.

Checked 2026-10-12: Page opened. Bi-encoders embed queries and passages separately so passage vectors can be precomputed; a cross-encoder scores each query-passage pair jointly and is more accurate; applying it to a whole collection would be too slow, so a retriever supplies a small candidate set (the page's example is about 100) that the cross-encoder reorders.

Sentence Transformers contributors

Checked 2026-10-12: Opened 2026-10-12. NoDuplicatesBatchSampler builds batches in which no value repeats within the batch, including across columns (an anchor text cannot recur as a positive or negative); matters for losses that treat other batch items as negatives, where a duplicate becomes a false negative; optional precompute_hashes (needs xxhash).

sentence-transformers

Checked 2026-10-11: Model card and Hub API: licence apache-2.0; maps text to a 384-dimensional vector; input longer than 256 word pieces is truncated by default; fine-tuned from nreimers/MiniLM-L6-H384-uncased with a contrastive objective on more than 1 billion sentence pairs; model.safetensors 90,868,376 bytes.

Sentence Transformers maintainers

Bi-encoder embeddings and cross-encoder re-rankers in practice; supports days 41, 48 and 51.

Checked 2026-10-11: sbert.net loads with title "SentenceTransformers Documentation - Sentence Transformers documentation".

Hugging Face

Defaults and flag names change between TRL releases; cite the mechanism, re-check the flag.

Checked 2026-10-12: Page 'SFT Trainer' (main docs). Loss is token-level cross-entropy over non-masked tokens; for prompt-completion datasets the loss is computed on the completion only by default (completion_only_loss), for plain language-modelling datasets on the full sequence; assistant_only_loss=True restricts loss to assistant messages and needs a chat template with generation markers; EOS token must match the chat template so responses terminate.

Elastic

Production search engine's BM25 configuration (k1, b); supports day 48 hybrid retrieval in practice.

Checked 2026-10-11: Page loads with title "Similarity settings | Elasticsearch Reference".

scikit-learn developers

TF-IDF baseline implementation (smoothing, sublinear tf, n-grams, fit on train only); supports day 41.

Checked 2026-10-11: Page loads with title "TfidfVectorizer - scikit-learn 1.9.1 documentation" (stable at check time).

Anthropic

Every successful response carries a stop_reason; max_tokens means the output was cut off and is incomplete (raise the limit or continue; a truncated tool call needs a retry). Supports insight 0512.

Checked 2026-10-11: Docs page (title now 'Stop reasons and fallback'): stop_reason description, max_tokens section and list of values; cross-checked against the Messages API reference enum.

Anthropic

Checked 2026-10-12: Opened 2026-10-12. Setting strict true constrains sampling so tool inputs match the JSON Schema; without it the model may return wrong types (for example a string for an integer) or omit required fields; the example uses an integer passengers field with an enum and a date-format string. The page guarantees schema conformance only, not business validity.

OpenAI

Second vendor's documentation of JSON-Schema-constrained output; supports day 47 comparison of schema enforcement vs prompt-only JSON.

Checked 2026-10-11: platform.openai.com/docs/guides/structured-outputs redirects to developers.openai.com page titled "Structured model outputs | OpenAI API".

Anthropic

Vendor documentation for schema-constrained responses; supports day 47 (structured outputs). Feature details change; re-check before citing specifics.

Checked 2026-10-11: docs.claude.com URL redirects to platform.claude.com page titled "Structured outputs - Claude Platform Docs".

Anthropic

Vendor documentation; behaviour may change by model and SDK version. Supports insights 1509 and 1510.

Checked 2026-10-12: Page opened 2026-10-12. Not supported: numerical constraints (minimum, maximum, multipleOf), string constraints (minLength, maxLength), array constraints beyond minItems 0 or 1, additionalProperties other than false, recursive schemas; unsupported features return a 400. SDK helpers transform the schema: remove unsupported constraints, add them to the field description, add additionalProperties false, and validate the response against the original schema if the helper validates; the TypeScript jsonSchemaOutputFormat helper does not validate responses. stop_reason refusal: HTTP 200, billed for generated tokens, output may not match the schema; a property asking for the model's step-by-step reasoning may trigger a reasoning_extraction refusal, so ask for a short explanation instead. max_tokens stop: output may be incomplete and may not match the schema.

Hugging Face

Checked 2026-10-12: Opened the page. Padding side section: LLMs are not trained to continue generation from padding tokens, so padding_side must be left for batched generation; the page's right-padded example produces a degenerate continuation and the left-padded one a sensible one.

Hugging Face

Checked 2026-10-12: Opened 2026-10-12. Subword tokenization (and the special tokens) creates a mismatch between word-level labels and tokens; the guide realigns by mapping tokens to words with word_ids, assigning -100 to special tokens so the PyTorch loss ignores them, and labelling only the first token of each word with -100 on the other subtokens. Its metric function drops positions whose label is -100 before calling seqeval.

Anthropic

Vendor documentation; numeric thresholds are vendor guidance, not an independent benchmark.

Checked 2026-10-12: Opened 2026-10-12. States that loading every tool definition up front causes context bloat (a typical five-server setup can use about 55k tokens of definitions) and that tool selection accuracy degrades once available tools exceed roughly 30-50; tool search loads only the few tools needed; recommends it for 10 or more tools, and keeping the 3-5 most used tools non-deferred. Thresholds are the vendor's statement for its own models.

Anthropic

Vendor documentation of tool definitions with JSON Schema inputs and the tool-call/tool-result loop; supports day 49.

Checked 2026-10-11: docs.claude.com tool-use URL redirects to platform.claude.com page titled "Tool use with Claude - Claude Platform Docs".

Hugging Face

Library docs for loading pretrained models, tokenizers and pipelines; supports days 42-44 code examples (docs-only verification).

Checked 2026-10-11: Page loads with title "Transformers · Hugging Face".

Unicode Consortium · UAX #15, revision 58 for Unicode 18.0.0 · 2026

Checked 2026-10-12: Opened the report. Canonically equivalent sequences represent the same abstract character yet can have different binary forms (precomposed C-cedilla vs C plus combining cedilla); NFD fully decomposes, NFC decomposes then recomposes; normalising both sides to one form makes binary comparison reliable.

OpenAI

Vendor doc; model names, defaults and limits change, re-check before citing.

Checked 2026-10-12: Page opened. text-embedding-3-small default length 1536 and text-embedding-3-large 3072; the dimensions parameter shortens the vector; recommended at creation time; a manually shortened vector must be re-normalised.

Anthropic

Supports insight 1518.

Checked 2026-10-12: Page opened 2026-10-12. Image cost is ceil(width/28) x ceil(height/28) visual tokens; per-tier limits: standard 1,568 px long edge and 1,568 visual tokens, high-resolution tier (Claude 4.7 and later) 2,576 px and 4,784 visual tokens; larger images are downscaled; the page states that high-resolution images can use up to roughly three times more visual tokens than the same image on a standard-tier model and advises downsampling when extra fidelity is not needed; lossy compression can introduce artifacts that hurt performance, especially with multiple passes; Files API file_id avoids resending base64 on every turn. Tier values change with model releases.

Anthropic

Offline evaluation and enrichment workloads; supports day 54.

Checked 2026-10-11: Page title 'Batch processing'. The Message Batches API processes requests asynchronously, each handled independently; suited to workloads that do not need immediate answers such as large evaluations; documented as lower cost and higher throughput than standard calls; most batches finish in under an hour.

NVIDIA

Checked 2026-10-12: Opened 2026-10-12. Dynamic batching combines requests on the server to raise throughput for stateless models; max_queue_delay_microseconds lets the batcher hold requests briefly and no request waits longer than it; preferred_batch_size is advised only when one size is much faster; the guide advises increasing the delay until latency exceeds the budget to see how much throughput it buys.

Microsoft

Checked 2026-10-12: Opened 2026-10-12. Pattern blocks calls to a remote resource after failures reach a threshold instead of retrying; states closed, open (fail immediately) and half-open (limited trial requests); the Retry pattern expects eventual success while the circuit breaker prevents operations likely to fail, and the two can be combined if retry logic stops when the breaker signals a non-transient fault; recommends separate breakers when independent providers sit behind one resource.

Anthropic

Supports day 52 request lifecycle and error taxonomy.

Checked 2026-10-11: Page title 'Claude API errors'. Lists 400 invalid_request_error, 429 rate_limit_error, 500 api_error, 504 timeout_error, 529 overloaded_error and others; every response carries a request id; the official SDK retries transient failures with exponential backoff twice by default, honouring retry-after; streaming errors can arrive after a 200 response; recommends streaming or batches for long requests.

Microsoft

Checked 2026-10-12: Opened 2026-10-12. Multi-step eventually consistent operations record each step and how to undo it, and rewind through completed steps on failure; compensation is application-specific and need not restore the exact prior state (a cancelled flight may not give a full refund); compensating steps can fail themselves so should be idempotent and resumable; define points of no return and place irreversible steps after critical validations; include a human for high-impact decisions.

Kubernetes authors

Checked 2026-10-12: Opened 2026-10-12. Liveness probes detect broken containers and trigger a restart; readiness probes tell whether a container can serve traffic (for example while loading large data at startup) and on failure the pod is marked not ready and Services stop routing to it without a restart; startup probes protect slow-starting containers and hold off liveness checks for failureThreshold x periodSeconds.

Evidently AI

Concrete defaults to compare against your own drift thresholds; shows that 'drift detected' depends on method, sample size and threshold choices.

Checked 2026-10-11: Docs page: default method selection (numerical <=1000 rows KS, >1000 rows normed Wasserstein; categorical chi-square/Z-test vs Jensen-Shannon), PSI available with default threshold 0.1, p-value methods drift at p < 0.05, dataset drift share default 0.5.

OpenAI

Second vendor example next to docs-anthropic-model-deprecations; re-check dates before quoting.

Checked 2026-10-12: Opened 2026-10-12 (redirected from platform.openai.com/docs/deprecations). Customers actively using a model are notified by email; each deprecation table gives a shutdown date per model and some name a recommended replacement; stated minimum notice differs by model class (about six months for generally available models, shorter for specialised variants and previews) and faster retirement is possible for safety or compliance reasons. Values as read on 2026-10-12.

DVC / Iterative

Versioning data and models alongside Git, and reproducible pipelines (dvc.yaml stages).

Checked 2026-10-11: dvc.org/doc redirects to doc.dvc.org; page title 'Home | Data Version Control · DVC'.

Evidently AI

Data drift, model quality and LLM evaluation reports; cite for monitoring lessons.

Checked 2026-10-11: docs.evidentlyai.com redirects to /introduction, title 'What is Evidently?', describes an open-source (Apache 2.0) framework for evaluating and monitoring ML and LLM systems.

Great Expectations

Supports the claim that data validation is a set of explicit, runnable assertions about data run before training and scoring (Day 19).

Checked 2026-10-11: Opened the GX Core overview page: an Expectation is a verifiable assertion about data (like an assertion in a Python unit test) used to validate a Batch; a Validation Definition links a Batch Definition to an Expectation Suite; a Checkpoint runs Validation Definitions in production and can trigger Actions on the results.

Prometheus authors

Checked 2026-10-12: Opened 2026-10-12. Quantiles of summaries cannot be aggregated across instances (averaging them is statistically nonsensical); histogram bucket counts can be summed across instances and passed to histogram_quantile; classic histograms need a bucket boundary at the SLO threshold for an accurate within-objective fraction, and estimate error is limited by the width of the bucket containing the value.

Pallets

Supports insight 1515.

Checked 2026-10-12: Page opened 2026-10-12. The default Undefined can be printed, iterated and used as a boolean, and its string form is empty; other operations raise UndefinedError. StrictUndefined raises an error on printing, iteration, boolean tests and comparisons; only the defined test is safe; the page notes it is the closest to standard Python behaviour.

Prometheus authors

Checked 2026-10-12: Opened 2026-10-12. Every unique combination of label key-value pairs is a new time series, which can greatly increase storage; the page cautions against using labels for dimensions with high cardinality such as user ids or email addresses, or other unbounded sets of values.

MLflow project

Checked 2026-10-11: Page 'ML Model Registry': incrementing model versions, mutable aliases (e.g. champion) to point deployments at a version, tags, Markdown descriptions, and lineage from each version to the run that produced it.

MLflow project (Linux Foundation)

Experiment tracking, model registry and packaging; cite for reproducible runs and model versioning.

Checked 2026-10-11: Page title 'MLflow Documentation' and links to experiment tracking, model packaging, registry and deployment guides on mlflow.org.

Anthropic

Pinned model versions have an end date; supports day 52 versioning and rollback.

Checked 2026-10-11: Page title 'Model deprecations'. Lifecycle terms active, legacy, deprecated, retired; requests to retired models fail; at least 60 days' notice for publicly released models; advises testing applications on replacement models well before the retirement date.

scikit-learn developers

Checked 2026-10-12: Opened 2026-10-12. Pickle-based methods do not support loading a model trained with a different scikit-learn or dependency version (InconsistentVersionWarning is raised on mismatch); pickle, joblib and cloudpickle have security vulnerabilities by design and should be loaded only from trusted sources; recommends pinning dependencies in training and production, saving metadata (reference to an immutable data snapshot, source code, library versions, cross-validation score); skops.io and ONNX described as alternatives.

MLflow project

Checked 2026-10-12: Opened 2026-10-12. Signatures act as a contract for inputs, outputs and parameters; inputs are checked against the signature when loading as PyFunc, with deployment tools and when serving over REST. Required fields must be present, optional fields may be absent, extra fields are dropped, safe type conversions (int to long, int to double, float to double) are applied while long to double and string to int are rejected; an integer column with missing values becomes float and causes a mismatch; unknown parameters produce warnings.

Grafana Labs

Checked 2026-10-12: Opened 2026-10-12. Closed model: a new iteration starts only when the previous one finishes, so when the system under test slows down iterations take longer and the arrival rate of new iterations drops. Open model (arrival-rate executors): iterations start at a set rate independent of how long earlier ones take, so response time does not reduce the applied load. The page links the closed-model behaviour to the term coordinated omission.

OpenTelemetry

Status is Development: names may change. Supports day 52 tracing and day 54 latency metrics.

Checked 2026-10-11: Conventions moved from opentelemetry.io to the open-telemetry/semantic-conventions-genai repository (notice on opentelemetry.io/docs/specs/semconv/gen-ai/). Spans page marked Development; attributes include gen_ai.request.model, gen_ai.response.model, gen_ai.usage.input_tokens, gen_ai.response.finish_reasons; content attributes (output messages, system instructions, retrieval documents and query text) are Opt-In and flagged as possibly sensitive. Metrics page (Development) defines gen_ai.server.time_to_first_token, gen_ai.server.time_per_output_token, agent inference/tool-call counts and tool execution duration.

Anthropic

Vendor feature details (TTL, minimums, multipliers) change; cite only the mechanism unless re-checked. Supports days 52 and 54.

Checked 2026-10-11: Page title 'Prompt caching'. A hit requires an identical prefix up to the cache breakpoint; prefix order is tools, then system, then messages, and a change invalidates that level and everything after it; default lifetime 5 minutes refreshed on each hit (1-hour option at extra cost); prompts below a model-specific minimum are silently not cached; cache reads are billed at a fraction of base input price and reduce time-to-first-token; output is unchanged.

Anthropic

Vendor documentation, dated 'as read'; behaviour, lifetimes and isolation rules can change.

Checked 2026-10-12: Page opened 2026-10-12 (as read; first 100,000 of about 160,000 characters read). Cache limitations: a cache entry becomes available only after the first response begins, so parallel requests that need hits should wait for the first response before sending the rest. Lifetime clock starts when the request that writes or reads the entry begins, and default lifetime is 5 minutes, refreshed on use. Storage and sharing: organisations do not share caches even for identical prompts; within an organisation caches are isolated per workspace on some platforms and only per organisation on others; hits need an exact match of the prefix up to the cache_control block.

Anthropic

Supports insights 1517 and 1519. Vendor values change often; the insights state no prices or lifetimes.

Checked 2026-10-12: Page opened 2026-10-12. Usage fields cache_creation_input_tokens, cache_read_input_tokens and input_tokens; total input is their sum; both cache fields 0 means not cached, most likely below the minimum length, with no error. Minimum cacheable length varies by model (examples read: 512, 1,024, 2,048 and 4,096 tokens; the page lists which model has which). Hierarchy tools, system, messages; a change invalidates that level and all after it. Invalidation table: tool definitions invalidate all levels; a tool_choice change and adding or removing images anywhere in the prompt keep the tools and system levels valid and invalidate message blocks; thinking parameters always invalidate message blocks and, model-specific, may invalidate earlier levels. Other toggles such as web search, citations and speed also invalidate. Numeric minimums change with releases; re-read before quoting.

Anthropic

Supports day 52 (rate limits and retries) and day 54 (caching raises effective throughput).

Checked 2026-10-11: Page title 'Rate limits'. Limits measured in requests per minute, input tokens per minute and output tokens per minute per model class; token-bucket replenishment, so short bursts can trip limits; exceeding returns 429 with a retry-after header; a monthly spend-cap 429 carries no retry-after and keeps failing until access resumes; for most models cache reads do not count toward input-token limits.

PyTorch contributors

Checked 2026-10-11: Page 'Reproducibility' (stable redirects to 2.14): completely reproducible results are not guaranteed across releases, commits or platforms, nor between CPU and GPU even with identical seeds; torch.use_deterministic_algorithms and cudnn.benchmark=False trade speed for determinism.

dbt Labs

Checked 2026-10-12: Opened 2026-10-12. A data test is a select that returns failing rows (zero rows = pass); singular tests are one-off SQL files, generic tests are parameterized and attached to models, columns, sources, seeds or snapshots; built-in generic tests unique, not_null, accepted_values, relationships; severity warn/error with warn_if/error_if thresholds; store_failures saves failing rows to a table for inspection.

PostgreSQL Global Development Group

Checked 2026-10-12: Opened 2026-10-12. count(*) counts all input rows and count(expr) only non-null values; except for count, aggregates return NULL when no rows are selected (sum is NULL, not 0; coalesce substitutes a default); avg, sum, min, max skip NULL inputs; array_agg keeps NULLs; percentile_cont/percentile_disc/mode ignore NULL inputs; order-sensitive aggregates (array_agg, string_agg) have unspecified input order unless ORDER BY is given inside the call.

Apache Airflow project

Checked 2026-10-12: Opened 2026-10-12. Tasks should produce the same outcome on every re-run because Airflow can retry; avoid now() for critical computation inside a task; read a fixed partition keyed by data_interval_start rather than the latest data; use upserts instead of plain INSERTs so reruns do not duplicate rows. The page does not explicitly recommend overwriting partitions.

SQLite

Checked 2026-10-11: Official page: avg() averages non-NULL values and returns NULL with no non-NULL input; count(X) counts non-NULL X while count(*) counts all rows; sum() returns NULL when there are no non-NULL inputs while the non-standard total() returns 0.0; DISTINCT and FILTER clauses described.

PostgreSQL Global Development Group

Checked 2026-10-12: Opened 2026-10-12. Ordinary comparisons return NULL when either input is NULL (7 = NULL and 7 <> NULL are NULL); IS DISTINCT FROM / IS NOT DISTINCT FROM treat NULL as an ordinary value (NULL IS NOT DISTINCT FROM NULL is true); IS TRUE / IS NOT TRUE / IS UNKNOWN never return NULL.

dbt Labs

Checked 2026-10-11: Official page: without unique_key most adapters append every returned row, so re-runs can duplicate rows; with unique_key existing rows are updated or replaced; the example filter uses >= on the latest date so a re-run reprocesses the last period; --full-refresh rebuilds the table when logic changes.

PostgreSQL Global Development Group

Checked 2026-10-12: Opened 2026-10-12. By default two NULLs are not equal for a UNIQUE constraint, so duplicates with NULL in a constrained column are allowed; NULLS NOT DISTINCT changes this; the SQL standard leaves the default implementation-defined; a CHECK constraint passes when its expression is true or NULL, so use NOT NULL to forbid NULLs.

Vercel

How scheduled jobs are declared (vercel.json crons) and invoked as HTTP requests to a route.

Checked 2026-10-11: Page title 'Cron Jobs', 'available on all plans' and definition as time-based scheduling on vercel.com/docs.

SQLite

Checked 2026-10-12: Opened 2026-10-12. SQLite has no dedicated date/time storage class; dates can be stored as TEXT (ISO8601), REAL (Julian day) or INTEGER (Unix time); NULL is less than every other value, so it sorts first in ascending order; TEXT-affinity columns compare as text (the page's example: a TEXT column holding '500' is not less than 60 numerically but a < 600 is true as text comparison results 0|1|1).

PostgreSQL Global Development Group

Checked 2026-10-12: Opened 2026-10-12. date_trunc on timestamptz truncates in the session time zone unless a zone is given as a third argument; AT TIME ZONE converts timestamp to timestamptz (reading input as local time in that zone) and timestamptz to the wall-clock timestamp; ISO weeks start on Monday, week 1 contains January 4, and isoyear can differ from the calendar year (2012-12-31 is in week 1 of 2013); adding interval '1 day' keeps local time of day across a DST change while interval '24 hours' adds elapsed hours (America/Denver 2005-04-03 example).

The PostgreSQL Global Development Group

Checked 2026-10-11: Official page: timestamp with time zone is stored internally in UTC and displayed in the session TimeZone; timestamp without time zone silently ignores a zone in the input; full zone names (IANA, e.g. America/New_York) carry daylight-saving rules while abbreviations such as EST are fixed offsets.

Great Expectations

Declarative data-quality tests ('Expectations') for pipelines; cite for data contracts and validation gates.

Checked 2026-10-11: Page title 'Home | Great Expectations', GX Core docs version 1.24.0 on docs.greatexpectations.io.

Apache Spark project

Checked 2026-10-12: Opened 2026-10-12 (first 100,000 characters). Windowed aggregation is keyed by event time and a late record updates its window; a watermark (max event time seen minus threshold) bounds state; data later than the threshold may be dropped; deduplicating on an id without a watermark keeps unbounded state, with a watermark on id plus event time state is bounded; dropDuplicatesWithinWatermark deduplicates on id alone.

Resend

Cite for exactly-once email sending in retries: pass an idempotency key so duplicate requests within 24 hours are not sent twice.

Checked 2026-10-11: Page title, supported endpoints (Send Email, Send Email Batch and inbox send/reply/forward) and the rule that Resend checks for the same key within the last 24 hours.

Stripe

Checked 2026-10-12: Opened 2026-10-12. Client-generated idempotency key lets a request be retried safely without performing the operation twice; API v1 saves the status code and body of the first request and returns them for retries with the same key; keys can be pruned after 24 hours; parameters are compared with the original and a mismatch errors.

PostgreSQL Global Development Group

Checked 2026-10-12: Opened 2026-10-12. ON CONFLICT DO UPDATE guarantees an atomic INSERT-or-UPDATE outcome per row even under concurrency; the command may not affect the same existing row twice (cardinality violation error), so rows proposed in one command must be unique on the conflict target; EXCLUDED refers to the proposed row; DO NOTHING skips conflicting rows.

Vercel

Cite for designing scheduled jobs to be idempotent and reconciliation-based, and for securing cron endpoints.

Checked 2026-10-11: Page title and sections on securing with CRON_SECRET, duration, concurrency, and 'Cron job delivery and idempotency' (runs can be missed or delivered more than once; make jobs idempotent).

PostgreSQL Global Development Group

Checked 2026-10-12: Opened 2026-10-12. real and double precision are inexact and stored values may show small discrepancies; comparing floating-point values for equality may not behave as expected; for exact storage and calculations such as monetary amounts use numeric, which is much slower than integer and floating-point types.

OpenLineage project

Checked 2026-10-12: Opened 2026-10-12. Core entities Job (unit of defined work that consumes or produces datasets), Run (one occurrence of a job, client-generated id), Dataset (discrete collection of data, name unique in a namespace); facets are extensible metadata on jobs, runs and datasets; run state-update events carry timestamps and may add input or output datasets.

pandas development team

Join-cardinality checks that make a fan-out join fail loudly (scenario-003; Day 7 join grain).

Checked 2026-10-11: API page for pandas 3.0.6: the validate parameter checks whether the merge is of the specified type - 'one_to_one' (keys unique in both), 'one_to_many' (unique in left), 'many_to_one' (unique in right), 'many_to_many' (no check). The page does not name the exception raised on failure.

pandas development team

Checked 2026-10-12: Returns a DataFrame with columns year, week and day according to ISO 8601; the example maps 2010-01-01 to year 2009, week 53, day 5. Python date.isocalendar run locally also gave 2024-12-30 -> ISO year 2025 week 1 and 2021-01-01 -> ISO year 2020 week 53.

Feast project

Day 2 point-in-time correctness; Day 7 event vs load time; Day 19 training-serving skew.

Checked 2026-10-11: Page loads with title 'Point-in-time joins'; describes scanning backward from each entity row's timestamp up to the feature view TTL, and the option to also filter on created timestamp so backfilled or corrected values do not leak into training data.

Protocol Buffers project (Google)

Checked 2026-10-12: Opened 2026-10-12. 'Updating A Message Type': changing a field number is wire-unsafe; adding fields and removing fields without reusing numbers are wire-safe; some type changes (int32/uint32/int64/uint64/bool) are wire-compatible but lossy. 'Consequences of Reusing Field Numbers': field numbers should never be reused, reuse can cause parse errors, data corruption or leaked data. 'Deleting Fields': add the deleted number (and name) to a reserved statement so the compiler rejects reuse.

Confluent

Checked 2026-10-11: Page lists BACKWARD (default), FORWARD, FULL, their transitive variants and NONE; BACKWARD lets new-schema consumers read old data, allows adding optional fields and deleting fields, and requires upgrading consumers before producers.

PostgreSQL Global Development Group

Checked 2026-10-12: Opened 2026-10-12. DISTINCT ON keeps the first row of each group and the first row is unpredictable unless ORDER BY fixes it; default NULLS LAST for ASC and NULLS FIRST for DESC (nulls behave as larger than non-nulls); LIMIT/OFFSET without ORDER BY returns an unpredictable subset and different LIMIT values can change plan and order; TABLESAMPLE BERNOULLI scans all rows and keeps each independently, SYSTEM samples blocks and may be less random because of clustering, REPEATABLE(seed) repeats a sample while the table is unchanged; UNION ALL is usually much faster than UNION.

dbt Labs

Checked 2026-10-12: Opened 2026-10-12. Snapshots implement type-2 slowly changing dimensions over mutable tables; dbt_valid_from / dbt_valid_to mark version validity and current rows have a NULL dbt_valid_to (or a configured value such as 9999-12-31); timestamp strategy uses updated_at, check strategy compares check_cols; unique_key must be truly unique; hard_deletes options ignore / invalidate / new_record (v1.9+); changes are captured only when dbt snapshot runs, so it must be scheduled.

dbt Labs

Checked 2026-10-11: Official page: freshness checks are run with dbt source freshness (not part of dbt build) to tell whether data meets its SLA; recommends running freshness jobs at least twice as often as the tightest SLA (e.g. every 30 minutes for a 1-hour SLA). Threshold syntax (warn_after / error_after) is on the linked sources page, not this one.

SQLite

Checked 2026-10-11: Official page: '/' on two integers uses integer arithmetic (result truncated); if either operand is real, floating point is used. Operators return NULL when an operand is NULL; IS and IS NOT treat NULLs as comparable and never return NULL.

SQLite

Checked 2026-10-11: Official page: window functions added in SQLite 3.25.0 (2018-09-15); lag() documented as a built-in window function; ROWS frame type with n PRECEDING / CURRENT ROW.

The PostgreSQL Global Development Group

Checked 2026-10-11: Official page: if there are no equal right-hand values and at least one right-hand row yields null, NOT IN returns null, not true, following SQL's three-valued logic.

The PostgreSQL Global Development Group

Checked 2026-10-11: PostgreSQL 18 docs, 4.2.7 aggregate expressions: most aggregates ignore NULL inputs; count(*) counts rows, count(f1) counts non-null f1. 4.2.8 window function calls: default frame is RANGE UNBOUNDED PRECEDING (= RANGE BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW); with ORDER BY the frame extends to the current row's last peer; without ORDER BY all rows are peers.

PostgreSQL Global Development Group

Checked 2026-10-12: Opened 2026-10-12. Frame modes ROWS, RANGE and GROUPS; peers are rows the ORDER BY sorts as equivalent; with ORDER BY the default frame is RANGE UNBOUNDED PRECEDING through the last peer of the current row, without ORDER BY all partition rows are peers; RANGE with an offset needs exactly one ORDER BY column (interval offset for date/time columns); EXCLUDE CURRENT ROW / GROUP / TIES.

PostgreSQL Global Development Group

Checked 2026-10-12: Opened 2026-10-12. first_value, last_value and nth_value use the window frame, whose default is likely to give unhelpful results for last_value and sometimes nth_value; rank() has gaps and returns the row_number of the first peer, dense_rank() has none; peers get the same rank; lag/lead offset defaults to 1 and default value to NULL; ntile divides the partition as equally as possible.

pandas development team

Rolling, expanding and exponentially weighted windows; cite for lag/rolling feature construction and the need to shift before rolling to avoid look-ahead.

Checked 2026-10-11: Page title 'Windowing operations' in the pandas 3.0.6 User Guide.

National Institute of Standards and Technology (NIST)

Landing page for the AI RMF, Playbook and revision status; for the framework document itself cite docs-nist-ai-rmf-1.

Checked 2026-10-11: Page title, AI RMF 1.0 released January 26, 2023 for voluntary use, links to the AI RMF Playbook and Resource Center, and a note that AI RMF 1.0 is being revised under the White House AI Action Plan.

National Institute of Standards and Technology (NIST) · NIST AI 100-1 · 2023

Govern/Map/Measure/Manage functions and trustworthy-AI characteristics; supports day 53 governance vocabulary.

Checked 2026-10-11: NIST publication page confirms title and publication date 2023-01-26; Crossref record for DOI 10.6028/NIST.AI.100-1 confirms report title and date. NIST AI RMF hub page www.nist.gov/itl/ai-risk-management-framework also loads.

National Institute of Standards and Technology (NIST) · NIST AI 600-1 · 2024

GenAI-specific risk list and suggested actions mapped to the AI RMF; supports days 53 and 55.

Checked 2026-10-11: NIST publication page confirms title and publication date 2024-07-26; Crossref record for DOI 10.6028/NIST.AI.600-1 confirms the report and date.

OpenAI

Retention differs per endpoint and feature, so a vendor review must be per endpoint. Re-check before quoting numbers.

Checked 2026-10-12: Opened 2026-10-12 (redirected from platform.openai.com/docs/guides/your-data). Data sent to the API is not used to train models unless the customer opts in (as of 2023-03-01); abuse-monitoring logs kept up to 30 days by default; Zero Data Retention and Modified Abuse Monitoring exist for approved customers; endpoints differ: some store no application state, others (files, fine-tuning jobs, batches, vector stores, conversations) are stored until deleted; Responses API objects 30 days by default; prompt caching expires after 24 hours. Values are as read on 2026-10-12 and change.

Pallets

Supports insight 1701; says nothing about LLM prompts.

Checked 2026-10-12: Page opened 2026-10-12. The sandbox intercepts attribute access, method calls, operators, mutation and string formatting to render less trusted templates; its Security Considerations section says the sandbox alone is not a solution for perfect security, that a small template can produce very large output (use CPU and memory limits), and to pass only the data that is needed.

Hugging Face

Checked 2026-10-11: Page 'Licenses': license is declared in the repo card metadata; table of license identifiers including permissive, non-commercial, RAIL and model-specific community licenses; 'other' requires a LICENSE file; advises users to seek out and respect a project's license.

Model Context Protocol project

Tool metadata from a third-party server is input to the model and can change at runtime.

Checked 2026-10-12: Opened 2026-10-12. Tool definitions carry name, description, inputSchema and optional annotations; clients MUST consider tool annotations untrusted unless they come from trusted servers; servers declaring listChanged send a notification when the tool list changes; Security Considerations: servers MUST validate inputs, enforce access control, rate limit and sanitize outputs; clients SHOULD prompt for confirmation on sensitive operations, show tool inputs before calling, validate results before passing them to the LLM, apply timeouts and log tool usage. Dated spec version; later versions may differ.

Python Software Foundation

Checked 2026-10-12: Opened 2026-10-12. Top warning: the pickle module is not secure; only unpickle data you trust; malicious pickle data can execute arbitrary code during unpickling; never unpickle data from an untrusted source or that could have been tampered with; consider signing data with hmac; use a safer format such as json for untrusted data.

Presidio project (originally Microsoft; now community-owned under Data Privacy Stack)

Reference open-source redaction pipeline used as a candidate in project-050.

Checked 2026-10-11: microsoft.github.io/presidio redirects to data-privacy-stack.github.io/presidio, which loads and describes analyzer (regex, rule-based and NLP recognizers) and anonymizer modules and a project-transition notice; GitHub API reports repository data-privacy-stack/presidio under MIT.

Presidio project (Data Privacy Stack)

Complements docs-presidio, whose registry text predates this check.

Checked 2026-10-12: Opened 2026-10-12 (redirected from data-privacy-stack.github.io/presidio). A warning under Main features states that Presidio's automated detection cannot guarantee that all sensitive information is found, so other systems and protections should be used alongside it; modules listed: analyzer, anonymizer, image redactor, structured.

European Parliament and Council of the European Union · 2016

Cite only for the principles checked; this is not legal advice and other jurisdictions differ.

Checked 2026-10-11: EUR-Lex official text: Article 5(1)(c) data minimisation (adequate, relevant, limited to what is necessary) and 5(1)(e) storage limitation (identifiable form no longer than necessary, with safeguarded exceptions for archiving, research and statistics).

Supabase

How Postgres RLS policies restrict rows per user (auth.uid()); cite for per-user data isolation.

Checked 2026-10-11: Page title, policy examples with 'using' clauses and the auth.uid() helper on supabase.com/docs.

Erika McCallister, Tim Grance, Karen Scarfone (NIST) · 2010

Defines PII and impact-level thinking for confidentiality safeguards; supports PII category design in project-050. Educational, not jurisdiction-specific legal advice.

Checked 2026-10-11: NIST CSRC page loads with title, three authors and Date Published April 2010; Crossref DOI 10.6028/NIST.SP.800-122 matches title and authors.

Alexander Nelson, Sanjay Rekhi, Murugiah Souppaya, Karen Scarfone · NIST SP 800-61 Rev. 3 · 2025

General incident-response guidance; GenAI-specific actions are in the NIST AI 600-1 profile (docs-nist-ai-600-1-genai-profile).

Checked 2026-10-12: Opened the NIST CSRC page 2026-10-12: title, four authors, published April 2025, supersedes Rev. 2 (2012); stated goals are to prepare for incidents, reduce their number and damage, and improve detection, response and recovery within CSF 2.0 risk management. Document body not read.

Unicode Consortium

Documents that the block exists and mirrors ASCII; the invisibility and misuse claim is a practitioner observation.

Checked 2026-10-12: Downloaded the chart PDF and extracted its text 2026-10-12: block named Tags, range E0000 to E007F, containing U+E0001 and code points from E0020 upward; that these mirror printable ASCII is from secondary write-ups. The chart does not itself say how renderers draw them; that tag characters are usually invisible and have been used to hide prompt instructions comes from secondary security write-ups seen in a web search, not cited as a claim.

Amazon Web Services

Vendor-specific example of a general pattern (short-lived, scoped credentials).

Checked 2026-10-12: Opened 2026-10-12. Temporary credentials are short-term (minutes to hours), generated dynamically and not stored with the user, so long-term credentials need not be embedded in an application; after expiry they are no longer accepted, so no explicit revocation is needed.

PyTorch contributors

Checked 2026-10-12: Opened the 2.14 page 2026-10-12 (the unversioned stable URL redirects to it). torch.load uses an unpickler under the hood and says never to load data from an untrusted source; weights_only restricts the unpickler to tensors, primitive types, dictionaries and types registered with torch.serialization.add_safe_globals; the signature on that page shows weights_only defaulting to True. The default differs in older releases, which this check did not cover.

Eberly Center for Teaching Excellence & Educational Innovation, Carnegie Mellon University

Basis for the academy's objective-activity-assessment alignment in lesson design.

Checked 2026-10-11: Page title and the explanation of why assessments, learning objectives and instructional strategies should be aligned on cmu.edu.

Anthropic

A contract, not technical documentation; terms differ by vendor and change. Not legal advice; the team's counsel decides what the clause covers.

Checked 2026-10-12: Page opened; effective date shown June 17, 2025. Section D.4 (Use Restrictions) bars customers from accessing the Services to build a competing product or service, including to train competing AI models.

Python Software Foundation

Checked 2026-10-12: str.lower: 'The lowercasing algorithm used is described in section 3.13.2 Default Case Conversion of the Unicode Standard'; str.casefold follows section 3.13.3 Default Case Folding. The page does not mention Turkish or locale rules; the behaviour used in the insight ('I'.lower() == 'i', len('İ'.lower()) == 2) was observed by running Python 3 locally.

Eurostat

License basis for Eurostat dataset use in projects.

Checked 2026-10-11: Page title loads; reuse of statistical data authorised provided the source is acknowledged, editorial content CC BY 4.0, exceptions for data from non-Eurostat sources (commercial reuse), policy implemented by Commission Decision of 12 December 2011 (2011/833/EU).

Mozilla contributors

Secondary reference for HTTP semantics (RFC 9110, section 13.1.1).

Checked 2026-10-12: Opened 2026-10-12. For unsafe methods such as PUT, If-Match makes the request conditional on the stored resource still matching a client-supplied ETag; this prevents the lost update problem, and when the condition fails the server answers 412 Precondition Failed instead of applying the change. Weak ETags never match under the strong comparison.

JSON Schema organization (IETF draft authors) · 2022

Supports insight 1514. Libraries may add default-filling as their own feature.

Checked 2026-10-12: Specification page (draft 2020-12) opened 2026-10-12. Section 9.2 describes default as supplying a default JSON value associated with a schema; keywords of section 9 are annotations for documentation and UI display, not validation assertions; the keyword has no validation behaviour; the value is recommended, not required, to be valid against its schema. The page does not say applications may use default to fill in missing values.

Python Software Foundation

Documents the mechanism only. Supports insight 1701.

Checked 2026-10-12: Page opened 2026-10-12. A replacement field's name can be followed by any number of attribute lookups (.name, resolved with getattr) and index lookups ([index], resolved with __getitem__); examples include '{0.weight}' and '{players[0]}'. The page contains no warning about user-supplied format strings (confirmed absent in the Format string syntax and Template strings sections), so the risk statement in the insight is the course's inference from the documented lookups.

Board of Governors of the Federal Reserve System, FDIC, Office of the Comptroller of the Currency · 2026

Supersedes SR 11-7; cite SR 26-2 for current wording.

Checked 2026-10-12: Transmittal page: SR 26-2 issued 17 April 2026, supersedes SR 11-7 (4 April 2011) and SR 21-8, risk-based and tailored. Attachment PDF (SR2602a1.pdf) text extracted: section 'Model Inventory' says maintaining a comprehensive set of information for models under development or in use is common industry practice and an effective inventory has enough information to understand model risk individually and in aggregate; 'Documentation' section says adequate documentation supports continuity of operations. Applicability stated as most relevant to banking organisations over 30 billion USD in assets.