Changes in version 0.6.0 Breaking changes - stepwise() is linear(), named after the model it fits as additive() is, and its defaults are biomod2's GLM: every column as x + I(x^2), searched both ways by AIC as MASS::stepAIC() searches, with no bound on the terms kept. The search is the select argument ("both", "forward", "backward" or "none"), formerly direction. The old defaults were the study's forward arm, which is now linear(select = "forward", terms = "column", max_terms = 3). The learner registers and reports as linear. - maxnet() is maxent(), named after the model rather than the package whose formulation it follows. Its arguments, defaults and numbers are unchanged; it registers and reports as maxent. - boosting(newton = TRUE) is boosting(method = "xgboost"). The flag switched the whole algorithm, the tree growth, the split criterion, what depth and min_leaf count and the default set, where its name described the leaf step alone, which gbm's trees take as well. method = "gbm", Friedman's gradient boosting machine, is the default as before. - year_start defaults to "01-01", the calendar year, in grain_matrix(), grain(), multigrain(), grains() and coverage(), where it was the Schrankogel deposit's hydrological year, "09-01". A season or year bin now starts on 1 January unless a boundary is given; inst/reproduce/schrankogel.R passes "09-01". - timesift() takes learners for models, n_inner for inner and choose for rule; select_grain() takes n_inner, choose, and cut_rule for threshold, and its result carries choose and cut_rule where it carried rule and threshold; tune() takes n_inner. A fit's fitted models are still fit$models. - elasticnet() defaults to cv.glmnet()'s ten inner folds and s = "lambda.1se", where it took five folds and "lambda.min". The reproduction driver and the selection benchmark pass their own settings, so their numbers are unchanged. - mars() and discriminant() take max_terms for nk and min_gain for thresh; elasticnet() and maxent() take tol for thresh. A convergence tolerance and the least gain a forward step is kept for were sharing one name. - preset = "package" is preset = "default" on tree(), forest() and boosting(). - kappa_score() cuts at "prevalence" by default, the rule the registered kappa metric reads, where it cut at "youden". - train_control(early_stopping) defaults to Inf, which never stops. It is read only where val_frac holds a validation set back, and at val_frac = 0 it was never read. - decision_threshold() on a fit takes perc for the "mpa" rule in both languages; in Python candidate and perc are keyword-only, so the R call decision_threshold(fit, "ensemble") errors there instead of reading "ensemble" as predictions. New - perceptron() fits one network of a single hidden layer per response, as nnet fits it and biomod2 fits ANN: logistic hidden units, a bias and optional skip-layer connections, the weights fitted by Nash's variable metric (BFGS) method on the head's loss plus a decay on the squared weights. It runs on a new core, src/ts_perceptron.cpp over the minimiser in src/ts_quasi_newton.cpp, that both languages call, and from the same starting weights it ends at nnet's weights to the bit on the fixture design. preset = "default" is biomod2's default option set (two hidden units, nnet's decay = 0, rang = 0.7, maxit = 100) and preset = "bigboss" its tuned one; tune() searches biomod2's ANN grid of hidden and decay. Under a presence-absence head it fits the cross-entropy, nnet's entropy = TRUE, where biomod2 leaves nnet's least squares on the logistic output. threads fits that many responses at once, each network the one it gets alone, and standardise = TRUE centres and scales the columns first, which nnet and the default do not. - maxent(threads) fits that many responses at once under either formulation, each the fit it gets alone; it reached only the absence formulation's inner folds, so a background fit ran on one thread. The designs held at once share max_design. - elasticnet(), linear(), tree(), forest(), boosting(), mars() and discriminant() hand their core every response at once, as perceptron(), maxent() and additive() do, and threads fits that many responses at once. With one response to fit it goes where it went before: the inner folds, a step's candidate fits, a forest's trees or the column search. tree() takes threads. Each response is the fit it gets alone, so no number moves. - train_control() takes optimizer ("adamw", the default, "adam" or "sgd"), a penalty with its alpha, cito's lambda and alpha with alpha read as elasticnet() reads it, and a schedule ("cosine", the default, "constant" or "plateau" with plateau_factor and plateau_patience). biomod2's DNN is cito's dnn(), and both its option sets are now settings of mlp(); the biomod2 article spells them out. The defaults train as before. Bug fixes - A nested run whose ensemble sets min_score no longer fails with "numbers of columns of arguments do not match" when an outer fold's stack leaves a candidate out. fold_weights has a column per candidate in every fold, 0 where that fold left it out; in Python every row carries every candidate, where a fold's row named only the candidates it kept. - The map of plot() on a range_change() labels its classes lost, kept, absent and gained, and colours a map holding only some of them by the class each cell is in. - Subsetting the summary() of a run, or an occlusion() result, returns a plain data frame where it printed empty headers. - plot() on a response curve takes legend, its position or NULL, and lays more than four responses out in columns. - A learner whose own inner cross-validation asks for more folds than a grouped or blocked split leaves it groups, such as the default elasticnet() under block_cv(v = 5), is told to lower that learner's n_inner. - The refusals of a record name one empty cell or one non-finite reading in the singular, and a prediction on a record with fewer bins than the fit says why a calendar grain needs the same period. Documentation - Five articles on the package site: representing a record (grains, statistics, calendars, gaps, partial bins, lookbacks), the learners, finding the grain on simulated records with a known answer, ensembles, thresholds and maps, and troubleshooting. - Examples for predict(), plot() and summary() of a run, build_representation(), and the plots of a response curve and of a change in range. - mlp() is described as what it is, a fully connected encoder reading the bins flattened into one block; only cnn() and rescnn() read them as a sequence. - The lookback's edge rule is stated as the core applies it: a bin lying wholly past the record is an error, and one the record reaches in part is summarised over the readings it holds. Other changes - Every learner whose defaults follow a reference package says that under the shipped presence-absence head it is fitted under positive_weights(), so a default tree() is rpart's specification under those case weights. hierarchical() says that at spatial = "none" and random = FALSE it is a penalised logistic model. - hierarchical() finds the mode of its hyperparameters with the variable metric minimiser perceptron() uses, where it carried a second BFGS of its own with a capped step. The mode, and with it the nodes and the fit, can move at the size of the search's tolerance. - In Python a representation, a learner and a training control print as they do in R, where they printed as their dataclass fields and a learner's functions. - The site's Python side has a Get started article, the R walkthrough section by section and run against the package, and a reference laid out under the sections of the R reference with an index page, so the two language menus carry the same pages. - The reproducing-schrankogel vignette is gone. inst/reproduce/README.md is where the reproduction is documented: how to run each stage, what it costs, and the comparison with the paper, which the vignette carried an older copy of. Changes in version 0.5.0 New - A count response. response = "count" is whole numbers of zero or more, fitted under the Poisson deviance through a log link and an exponential output, and read by neg_poisson_deviance (regression_metric(y, p, "poisson_deviance")). Every core that follows a head's loss takes the family: the penalised fit (glmnet's family = "poisson"), the stepwise search and the MARS refit (glm(family = poisson), MASS's stepAIC() with the Poisson likelihood, earth's glm = list(family = poisson)), the additive model (mgcv's gam(family = poisson), whose scale is known, so the criterion is the unbiased risk estimator), the tree (rpart's method = "poisson", with the gamma shrinkage of a leaf's rate as tree(shrink = )) and the boosted trees (gbm's poisson distribution, and xgboost's count:poisson under newton = TRUE). A forest cuts a count on its variance and a leaf holds the mean count, the encoders train under the deviance, and the combiner minimises it. maxnet(), envelope() and discriminant() refuse a count head as they refuse any numeric one. Both languages, pinned core by core in the fixtures, with the response in count_response.csv. - cv(repeats = ) and grouped_cv(repeats = ) draw the outer split several times, each with its own seed, as biomod2's nb.rep does. timesift() makes the run once per repeat: a score stays one fit's held-out score, scores and cells carry a repeat column and every repeat's folds under numbers of their own, a response is averaged over its folds and its repeats, estimate is read off all the repeats' held-out predictions, and the stack is fitted on the repeats' out-of-fold predictions together. oof is a target's mean prediction over the repeats, and the models, the per-fold fits and folds are the first repeat's. Both languages. - pseudo_absences() draws units where a response is zero from a pool of background units, by bm_PseudoAbsences()'s strategies: "random", "sre" (outside the envelope of the presences) and "disk" (a distance band from the nearest presence, planar or on the sphere), with repeated draws. The units a strategy admits are the same in both languages and pinned in pa_candidates.csv; which are drawn depends on the language's generator. - Three response heads for a numeric response: "continuous", "abundance" (not negative) and "ordinal" (whole-number classes), fitted under squared error through an identity output by every learner but maxnet(), envelope() and discriminant(). regression_metric() carries biomod2's RMSE, MSE, MAE and Max_error and R^2 and the correlation, and ordinal_metric() its Accuracy, Recall, Precision and F1, all registered by name. Both languages, pinned in numeric_metrics.csv. - project() applies a fit to one target per cell of a raster and returns the prediction as a raster with a layer per response: type = "binary" the binary map at the fit's own cuts and type = "spread" the ensemble's disagreement, which is BIOMOD_Projection() and BIOMOD_EnsembleForecasting(). Rasters are terra in R (in Suggests) and xarray in Python (the grid extra). range_change() counts the cells lost, kept and gained between two such maps, as BIOMOD_RangeSize() does, and returns the map of codes. - tune() wraps a learner so that, whenever it is fitted, it searches a grid of its settings by cross-validation inside the units it is handed and fits the best one. Inside a run each outer fold chooses from its own training units. What was chosen is on the fitted model and in the new settings column of the candidate table. This is biomod2's BIOMOD_Tuning(). Both languages. tune(learner) with no grid searches the one registered for it, which for forest(), boosting(), mars(), discriminant(), maxnet(), envelope() and mlp() is the grid BIOMOD_Tuning() searches, and register_tuning() adds a grid for a learner of one's own. - response_curve() varies one predictor of a fitted candidate across the range it takes while the others are held at the mean, median, minimum or maximum of their cells, and returns the prediction, as biomod2's response curves do. A predictor is a channel (moved in every bin, as a static column is), a bin of a one-channel representation, or one cell; a second predictor gives a surface, and for the ensemble spread = TRUE adds the members' standard deviation and interval. plot() draws it. Both languages. - occlusion() reports importance beside the fall in score: one minus the correlation of the held-out predictions with and without the part, which is biomod2's variable importance and needs no response. occlusion(fit, "ensemble", over = "channel") reads the ensemble, withholding a channel from every member that carries it and combining their predictions by the run's stack. Both languages. - block_cv() and env_cv() hold a block of targets out whole, cut on columns of targets by halving at the median of the widest column, in space or, on scaled columns, in predictor space. Nothing is drawn, so R and Python return the same blocks, pinned in blocks.csv. cv(by = ) stratifies on a column of targets instead of the richness of the response, which is biomod2's strat strategy. - table_metric() reads biomod2's evaluation statistics off the table of decisions against observations: pod, pofd, far, sr, accuracy, bias, or, orss, csi and ets, at the cut of any rule of decision_threshold() or at one given. Each is registered by name, so grain_ladder(metric = "csi") reads it. boyce_index() is the continuous Boyce index, registered as "boyce", and decision_threshold(rule = "mpa", perc = ) is the minimum predicted area cut. Both languages, pinned in metrics.csv. - hierarchical() fits a Bayesian logistic model with an optional nearest-neighbour or Hilbert-space Gaussian-process field over the targets' coordinates, named in the new timesift(coords = ), and an optional intercept for each unit. Laplace's method over the coefficients, the field and the intercepts together, with the hyperparameters integrated on a grid around their posterior mode, in a C++ core both languages compile (src/ts_hierarchical.cpp), the nearest-neighbour precision factored by a sparse Cholesky decomposition under a minimum-degree ordering (src/ts_sparse.cpp) so the cost follows the number of locations. The case weights enter every configuration. The coordinates and the unit names travel with every representation and are split with its units. Pinned in the fixtures against tulpa's posterior mode, empirical-Bayes fit and each node of its nested grid, and in both suites against the same marginal written out densely. Both languages; tulpa is not a dependency. - mlp(activation = ) sets the nonlinearity between the hidden layers: "relu" (the default), "gelu" or "selu", in both languages. - The article "Coming from biomod2" maps every biomod2 algorithm and ensemble to a learner or an option of ensemble() and fits the whole set across weekly, monthly and seasonal grains on one simulated record. - additive() fits one generalised additive model per response, which is what biomod2 fits as GAM: mgcv's gam(y ~ s(x1) + s(x2) + ..., method = "GCV.Cp"), a thin plate regression spline of k = 10 basis functions per column with its smoothing parameters chosen by the unbiased risk estimator under presence-absence and generalised cross-validation under a squared-error head. Past 2000 distinct values a column's knots are the subsample mgcv draws, R's sample() after set.seed(1). A column of fewer distinct values than k takes as many basis functions as it holds, and one of two enters linearly, where mgcv refuses both. It runs on a new core, src/ts_additive.cpp, compiled into both languages: the basis from a Lanczos eigensolver, penalised IRLS over the Householder QR of src/ts_glm.cpp, and Newton's method on the smoothing parameters with the exact gradient and Hessian. On the contract's seven cases it settles where mgcv does to 5e-10 in the criterion and 1e-8 in the fitted means. Those cases are ones where the criterion has a single minimum. On the fixtures' 80 units of weekly columns, mgcv restarted from four smoothing parameters settled in more than one minimum on 15 of 20 binomial designs and 1 of 7 Gaussian ones; mgcv's own start reads its own parametrisation, so where there are several the two searches can settle in different ones. On 894 units and 20 columns a fit takes 5.2 s against mgcv's 3.1 s, and 2.6 s on four threads, which run the responses at once or, for one response, the pieces of its derivatives. Both languages. - discriminant() fits one flexible discriminant analysis per response, which is what biomod2 fits as FDA: mda's fda(method = mars), optimal scoring of presence and absence regressed on mda's own MARS basis, and the posterior of two normal classes on the one canonical variate. The basis is mda's Fortran marss, not earth's, and like it runs its forward pass unweighted; the head's case weights set the scores and the variate. calibrate = TRUE predicts through the probit recalibration biomod2 always applies to FDA, fitted on the fitting units. degree, penalty, nk, thresh and prune are mda's settings under mda's defaults, which biomod2 passes unchanged. It runs on a new core, src/ts_fda.cpp, compiled into both languages, with R's normal distribution in src/ts_normal.cpp and a probit link in the shared glm.fit; on the contract's five cases it keeps mda's terms exactly and its coefficients and predictions agree with mda's and with R's glm() to 1.4e-14. On 560 units and 150 columns a fit takes about what mda takes, and threads searches the columns at once. Both languages. - mars() fits one multivariate adaptive regression spline per response, as the earth package fits it and biomod2 fits MARS: earth's forward pass of hinge pairs with Fast MARS, leaps' backward pruning by generalised cross-validation, and the kept terms refitted as a logistic model under a presence-absence head (earth's glm = list(family = binomial)) or by least squares under a squared-error one. degree, penalty, nk, thresh, minspan, endspan, fast_k, fast_beta, prune and nprune are earth's settings under earth's defaults, which biomod2 uses under both of its option sets. It runs on a new core, src/ts_mars.cpp, compiled into both languages; on the contract's ten cases, weighted ones included, it keeps earth's terms exactly and its coefficients and predictions agree with earth's to 8e-14. A weighted fit reaches the residual sums earth refits by QR at every candidate knot with Friedman's running updates: on 715 units and 471 columns under presence weights it takes 0.1 s against earth's 17 s. Both languages. - tree() grows one classification or regression tree per response under rpart's rules: the Gini index under a presence-absence head and the sum of squares under a squared-error one, min_split, min_leaf, cp and max_depth as rpart.control() has them, and the cost-complexity table with its cross-validated error. The tree is pruned back by an inner cross-validation read by prune: "se_sum", the rule biomod2 prunes its classification tree by, "one_se", "min" or "none". preset = "bigboss" takes biomod2's tuned option set for every setting left open. It runs on a new core, src/ts_tree.cpp, compiled into both languages, so the two grow the same tree; on the contract's twelve cases it reproduces rpart's node table, complexity table, cross-validated error and pruned predictions to 1.1e-16. Both languages. - boosting() fits one gradient boosted model per response, gbm's trees by default and xgboost's exact greedy second-order trees under newton = TRUE, which are what biomod2 fits as GBM and XGBOOST. n_inner folds choose how many trees are kept as gbm's cv.folds does, and preset = "bigboss" takes biomod2's tuned option set for either. It runs on a new core, src/ts_boost.cpp, compiled into both languages. With every unit in the bag it reproduces gbm's predictions and cross-validated error to 6.7e-16, and xgboost's predictions to its single-precision storage; the subsample and the column draw come from the forest's generator and are pinned against a fit grown from the contract's text in R alone. On 894 units and 471 columns, one thread, biomod2's tuned gbm fit (2500 trees of seven splits) takes 56 s against gbm's 68 s, and xgboost's defaults 1.2 s against xgboost's exact method's 2.6 s. Both languages. - maxnet() fits one maxnet model per response, which is what biomod2 fits as MAXNET: maxnet's linear, quadratic, product, hinge and threshold features, chosen by the presence count as maxnet.formula() chooses them or named in classes, its regularisation of each feature times regmult, and a lasso over them on the penalised core elasticnet() runs on, so the two languages fit the same model. formulation = "background" is maxnet's presence-background model and predicts its cloglog output; formulation = "absence" reads the absences as absences, a logistic lasso over the same features under the head's case weights with the penalty chosen by an inner cross-validation. On the contract's twelve cases the features and penalty factors are maxnet's to 7.4e-16 and the objective glmnet's to 5.5e-14. The core is src/ts_maxnet.cpp. On 900 units a weekly three-channel representation, 47,100 features under maxnet's "lqh", fits in 8 s; a design above max_design gigabytes is refused with its size. Both languages. - forest(balance = TRUE) is the down-sampled forest biomod2 fits as RFd: each tree draws as many units from each class as the smaller class holds. forest() also takes preset ("package", randomForest's defaults, or "bigboss", biomod2's tuned set) and threads. Both languages. - ensemble() takes biomod2's ensemble set. ensemble("committee", rule =) is committee averaging: each member cuts each response at its own decision_threshold(), learned from its out-of-fold predictions, and the combination is the share of members voting presence; inside timesift() each outer fold's cuts are learned on its inner out-of-fold predictions, so the estimate reads no cut chosen on the targets it scores. min_score leaves out the candidates scoring below it before scope picks, and ensemble("weighted", decay = d) weighs each rank 1 / d of the one above it, as EMwmean.decay does. Both languages. - ensemble_spread() and predict(type = "spread") read how far an ensemble's members disagree: their weighted mean, standard deviation, coefficient of variation and t interval at alpha, biomod2's EMcv and EMci. On one target per map cell it is an uncertainty map. Both languages. - envelope() draws biomod2's surface range envelope, SRE, around each response's presences: every column's quantile and 1 - quantile quantiles over the presences, R's type 7 as bm_SRE() reads them, and a prediction of one where every column lies inside its band. On the contract's six cases its bounds and projections are bm_SRE()'s to the last bit. The core is src/ts_envelope.cpp. Both languages. - stepwise() takes direction = c("forward", "both", "backward", "none") and terms = c("column", "power"). "both" and "backward" are MASS's stepAIC() step for step, and terms = "power" makes each power of a column a term of its own, as biomod2's quadratic formula writes them; direction = "none" with it is the model biomod2 fits as GLM, x + I(x^2) over every column. max_terms = Inf lifts the budget, and threads runs one step's candidate fits at once. On the contract's thirteen cases the core chooses MASS's, glm()'s and the R oracle's terms, in the same order, with the deviance to about 1e-14. The defaults are the forward search over column terms it always ran. Both languages. Changed - stepwise() runs on a new core, src/ts_stepwise.cpp, compiled into both languages: R's glm.fit iteration over LINPACK's dqrdc2, which sets the rank R reports, and MASS's step rules. It replaces stats::glm() in R and a separate NumPy fitter in Python, which differed on what they refused: R took any fit that settled within 25 iterations, while Python also refused one whose fitted probabilities came within 1e-8 of zero or one. Both now refuse only a fit that has not settled, as R did. On the fixture's design and on simulated records the core chooses the columns R's glm() search chose, with predictions equal to about 1e-15. A fit saved from an earlier version does not predict under this one; refit it. - A stepwise() fit names every response whose final model did not settle in stopped. - A "weighted" ensemble in Python gives every member the same weight where none scores above zero, as R's always did, where it used to stop; ensemble(metric =) in Python now recomputes the scores it weighs and filters by, as R's does, and a scope's best candidate is the first offered among any on its score on both sides. - forest() is grown by the core src/ts_tree.cpp on both sides, in place of ranger in R and scikit-learn in Python. Its split search is the tree's, and its bootstrap and column draws come from one generator the contract defines (SplitMix32 seeding xoshiro128**, one stream per tree), so R and Python grow the same forest, bit for bit, on any number of threads. The forest is asserted node for node against one grown from the contract's text in R alone. ranger leaves Suggests and the Python sklearn extra is gone. This changes forest()'s numbers on both sides. Scored by TSS on the vignette's simulated record, on the vignette's folds, the old forest and the new one read 0.718 and 0.733 weekly and 0.737 and 0.752 monthly on 60 plots, and 0.762 and 0.760 weekly and 0.748 and 0.742 monthly on 300, every difference inside one standard error. On simulated presence-absence data its held-out AUC is level with randomForest's under the same settings (a mean difference of -0.0025 over ten draws, paired t -0.62). One forest on 894 units and 471 columns takes 5.3 s on one thread against ranger's 3.3 s. Nothing in the reproduction or the selection benchmark fits a forest. - min_node is the fewest units each side of a split keeps, randomForest's nodesize; it had been ranger's smallest node split in R and scikit-learn's smallest leaf in Python. - Python's internal module timesift.tree is timesift._tree. Importing the submodule replaced the exported tree() on the package with the module. - The Python source distribution carries src/ts_tree.cpp and src/ts_tree.h, which a wheel built from it needs. Changes in version 0.4.0 New - predict() on a fit takes type = "binary", cutting each response into presence and absence at the threshold learned under rule ("youden", "kappa" or "prevalence") from the same candidate's out-of-fold predictions of the fit's own targets. For the ensemble those are the members' combined under the refitted stack's weights. A binary map is this prediction on one target per map cell. Both languages. - decision_threshold() takes a fit in place of the response and returns one cut per response, the cut type = "binary" applies. In R it is an S3 generic with a method on a fit. - A learner's fit that declares a weights argument is handed the response head's case weights, a matrix of the response's shape, as one declaring head, control or group is handed those. A learner of one's own now fits a rare response under the weight every shipped learner fits it under. Both languages. - as.matrix() on a representation lays it out as the block of predictors the tabular learners read, one column per bin of each channel, named channel@bin, and one column per static predictor. It is Python's flatten(), which was already exported there. The example in ?learner used to average the bins away and now reads this block. - The Python side carries everything the R side does. grain_contrasts() fits score ~ grain + (1 | variable) + (1 | fold) by restricted maximum likelihood, reads each contrast's degrees of freedom as lmerTest does and adjusts by the multivariate t as emmeans' "mvt" does, written out in numpy with scipy's integrator and optimiser beside it; on the pinned scores it returns R's differences to 1e-6 and R's intervals and p-values to the integrator's error. simulate_records() builds the same design as R's, anchors, weights, driver scale and link, and draws the units from numpy's stream. plot() draws a ladder, a run or a selection on matplotlib in R's colours and returns the table it drew from. scipy and matplotlib are the contrasts and plot extras of the wheel, and the numpy-only install is unchanged. - grain_contrast_cells.csv, grain_contrast.csv and simulate_design.csv join the contract's fixtures, and both suites assert them. Changes in version 0.3.2 Reproduction - inst/reproduce/schrankogel.R runs the paper's headline arm, the eleven-member ensemble with its window chosen inside every outer training set, as the ensemble_selection stage. The eleven members pinned to one window are one learner, whose prediction is their mean probability, and the stage is select_grain() over the eleven rungs the study chose among: the window mean at all seven windows and the coldest-day, mean and warmest-day reading from the weekly window up, on the study's five inner folds by AUC. It is compared with the published 0.889 AUC and 0.729 TSS. The ensemble's members keep the paper's seeds at --seed=1 and move from them by seed - 1 at any other, so a run per seed refits the ensemble too. The reproducing-schrankogel vignette carries the same call (#82). Changes in version 0.3.1 (2026-10-02) New - calendar_channels() gains cycles. "day" adds day_sin and day_cos, the place of each bin's midpoint in the UTC day, beside or instead of the year's pair, in the order named. It is read in UTC for the reason the year is: a zone's offset and a site's longitude are the same shift for every bin, where a clock's summer time would move the phase by an hour twice a year. It is refused on bins a day or more apart, where every bin would sit at one place in the day. One definition in the shared core serves both languages, and the fixtures pin it as they pin the year's. Fixed - The encoders read the channels of calendar_channels() at their own amplitude, where they used to standardise them like a reading. The channels lie on the unit circle already, and being identical across units they reach a network as a bias, whose step under AdamW grows with the scale of the input carrying it; standardising multiplied them by 1.41. At the hourly rung they are four of the fully connected encoder's five channels, and on the Schrankogel record they cost it 0.044 TSS against the study, which fed them unscaled. calendar_channels() records its channels in a position attribute (a field in Python), and bind_channels() carries position and static from every argument; R's used to drop static and Python's to keep the first argument's alone (#81). - plot() of a select_grain() result drew empty axes. The inner scores were matched to their candidates on a label joined by a space where the candidates' was joined by a bar, so every line and every circle was placed at no position. One helper now writes the label wherever a selection's tables are matched, the bottom margin is sized to the longest candidate name, and the table returned says where each score was drawn and which one its fold selected (#79). - occlusion() with substitute = "permute" permuted each channel of a held-back bin separately, so a unit could be shown the coldest day of one unit beside the warmest day of another, a reading no unit made. A bin is now held back whole, by one permutation per draw. A channel that is the same for every unit, as the calendar channels are, is left in place when a bin is held back; unit_mean used to replace it too, which took the bin's place in the year away along with its reading (#79). - The Python occlusion() scored every cell, including the ones the scorable-cell mask leaves out, so its weights could rest on cells R's did not. It now reads the mask as every score does (#79). - grain_contrasts() read each grain's name back out of the contrast label emmeans writes, which puts parentheses around a level holding a character it treats as an operator: a grain called month-mean came back as (month-mean). The name is now taken from the factor's levels (#79). - grain_contrasts() stopped with "undefined columns selected" on a ladder of more than 3000 scored cells, which the Schrankogel grid is at 7070 per architecture: emmeans stops computing Satterthwaite's degrees of freedom past that size and names its interval columns differently. The limit is now the ladder's own size (#79). - Recovery tests for grain_contrasts(): a planted effect per grain is recovered and its intervals cover it at their stated rate over twenty seeds, and with no effect the contrast finds none (#79). The reproduction - inst/reproduce/schrankogel.R compares the stepwise arm and the network grid with the paper, where it used to write them without a reference (#80). The stepwise arm is fitted unweighted, as the study's forward selection was, under a head that differs from the shipped one in that alone. The eleven-member ensemble is the study's: each member at its own window, channels, kernel, dropout and seed, where the driver used to run one generic set at every window. The hourly rung reads the five channels the study's hourly networks read. The grain contrast is read in AUC against the window the study took as each architecture's best, as its table is. Changes in version 0.3.0 Changed - elasticnet() is one implementation in both languages. The elastic net is now src/ts_penalised.cpp, compiled into the R package and into the Python extension beside the binning, in place of glmnet::cv.glmnet() on one side and scikit-learn's LogisticRegressionCV on the other. Those were different algorithms: a different optimiser, a different parametrisation of the penalty, ten penalty values against a hundred, and different conventions for standardisation, so the same call returned different coefficients in the two languages. The penalised arm is what the networks are measured against, which made it the worst place for the twin to diverge. glmnet and scikit-learn drop out of the learner's needs; scikit-learn is still what forest() fits on (#76). - The core's conventions are glmnet's, and the replacement was held to reproducing it rather than to being an elastic net of its own. At a matched convergence threshold of 1e-14 the path's length and its penalties agree exactly, the cross-validated lambda.min and lambda.1se are the same point of the same path in all twelve fixture cases, and the penalised objective sits between 3e-13 and 2e-7 of glmnet's. Coefficients agree to 2e-6 where no two columns are collinear and to 1e-4 on the collinear fixture design, where the split between two nearly identical columns is not determined to the precision the fit is. inst/spec/representation.md states all of it, and inst/spec/fixtures/penalised_*.csv hold glmnet's own numbers for both suites to assert against. - elasticnet() gains threads, how many fits of one response's inner cross-validation run at once. The path on every fitting unit and the path of each inner fold are one independent fit each, reading the design and sharing nothing, so they parallelise without any of them seeing another: the result is bit-identical to the serial one, which both suites assert. The default is serial. - elasticnet() gains n_lambda and thresh, and s now reads on the Python side too: the whole penalty path is kept on both, so "lambda.1se" or a penalty of a caller's own naming a point between two of the path's is read the way predict(s =) reads one. - A fitted penalised model is arrays rather than a fitter's object, so it round trips through saveRDS() and pickle and predicts on a machine that has neither glmnet nor scikit-learn. - The reproduction's elastic-net row reads the same under the shared core as it did under glmnet: inst/reproduce/schrankogel.R --stages=baseline against the deposit gives 0.68605 on the 188 aggregates, 0.00095 from the published 0.687 and inside the 0.002 the driver compares at. - The penalised core's binomial fit is at glmnet's speed on one thread, at the shipped thresh = 1e-8, one notch tighter than glmnet's default. The path reads the columns once per penalty rather than twice, a binomial fit ends each penalty on the one reweighting the next penalty opens with, and a reweighted least squares step is judged on its own move, which is glmnet's test (#77). A Gaussian fit folds the root of each case weight into the design, so a coordinate's step reads two streams where it read three. - The descent's inner cycle is Anderson-extrapolated every five passes (Bertrand and Massias, 2021), and the extrapolated point is taken only where it lowers the penalised objective. That is what an ill-conditioned design needed: a tighter threshold there multiplied the passes, and now costs 2.8 to 6.9 times fewer of them. - Timings, one cross-validated path, installed builds, against cv.glmnet at its default on the same run: 0.47 s against 0.47 to 0.51 s for a binomial fit on the Schrankogel shape (894 units, 942 columns, five inner folds), 264 ms against 228 ms for a Gaussian one there (190 ms at glmnet's own threshold of 1e-7), and 0.12 s against 0.25 s for the four-response learner on a small ill-conditioned design (60 units, 18 columns). The reproduction's elastic-net row reads 0.68606 and runs in 21 minutes. - A penalised fit that does not settle at a penalty ends its path there and returns the points before it, which is what glmnet does with the same event. A fit that does not settle at the first penalty is an error. The path records where it stopped in stalled, a cross-validation records each fold's in fold_stalled, and elasticnet() names every response whose path ended that way in stopped, beside unfitted (#78). The package is renamed from timegrain to timesift. With it go the C++ prefix (tg_ to ts_ and the four files that carried it), the S3 classes (timegrain_matrix and its siblings to timesift_matrix), timegrain_set() to timesift_set(), and the Python module. No function signature, default or return type changed, and the fixture digests are untouched. What the user chooses is the temporal grain of a climate representation, and the name now carries the thing whose grain it is. The title reads "Temporal Climate Resolution for Ecological Prediction". Changes in version 0.2.0 Changed - train_control() holds no inner validation set back by default (val_frac = 0, was 0.15): an encoder trains every fitting target for the whole epoch budget and keeps the last epoch, where the cosine schedule has annealed the learning rate to zero. On the Schrankogel weekly three-channel arm that epoch reads 0.8803 AUC against 0.8729 for early stopping on a 15 percent split and 0.8813 for the best epoch read on the test folds. early_stopping = Inf with a split held back still restores the epoch of lowest validation loss (0.8760 there). The study's rule is train_control(val_frac = 0.15, early_stopping = 10), which the reproduction driver passes. On both sides. - The encoders read the inner validation loss that early stopping watches under the head's case weights, as they read the fitting loss, so the epoch kept is the one the fit's own objective prefers. The weights are still read off the fitting units alone: a response head's weights now takes (y, fitting), with fitting marking the rows the model is fitted on, and positive_weights() gains the same argument. Read unweighted, the validation loss of a presence-absence fit is dominated by the absences of the rare responses the fit was told to weigh up and stops at an earlier epoch: on the Schrankogel weekly three-channel arm the change is worth about 0.003 AUC, which is the gap to the analysis pipeline's encoder (#75). On both sides. - The inner validation set val_frac holds back is now drawn by a plain random permutation, not by one unit from each of n_val equal-count strata of the response total. Five seeds each of the fixed Schrankogel weekly three-channel arm, package recipe otherwise unchanged: the plain draw reads 0.8721 AUC against the stratified draw's 0.8674 (+0.0047, Welch p = 0.019), the same margin the validation-loss weighting is worth. The stratified draw guaranteed a rare response a presence in every validation set; the plain draw can still miss one, and no longer guarantees it. train_control(val_frac = 0.15, early_stopping = 10) is unaffected by name, only by what the split under it now does (#75). On both sides. - timesift() now estimates the procedure it runs. Inside each outer fold of resampling the training targets are split again into inner folds (5 by default), every candidate is cross-validated there, rule chooses one on its inner score and the stack's weights are fitted on the inner out-of-fold predictions; the choice and the weights then predict the outer test fold once. estimate holds the held-out score of the selected candidate and of the stack under every registered metric, selected and inner the choice in every outer fold, fold_weights the weights, and predictions both held-out prediction matrices. The candidates' scores on the outer folds are unchanged and are reported apart from the estimate, because the best of them was picked out on the folds it is scored on. inner = NULL compares the candidates without an estimate. timesift() and select_grain() share one inner search and one refit, so on the same candidates, fold map, inner count and seed the selected arm is select_grain()'s selection. On both sides. - The ensemble row of the report is the stack's held-out score under per-fold weights. Before, the weights were fitted on the outer out-of-fold predictions and the combination was then scored against the responses they had been fitted to. summary() prints the candidates and the procedure as two tables; the Python side's ensemble_row() is replaced by procedure_table(). - predict(candidate = "selected") predicts with choice, the candidate the rule takes on every target. - The presence-absence head is scored by roc_auc unless a metric is named, and score_predictions() defaults to it too. TSS read at its best cut is inflated by an amount that depends on how a model's predictions are distributed, so it stays beside AUC rather than ahead of it. On both sides. New - average_precision(), registered as a metric and pinned in the metric fixtures. On both sides. - The reproduction driver compares the demonstration species by species (#74). inst/reproduce/reference.csv carries, per species, the analysis pipeline's five-inner-fold selection and its weekly series elastic net under AUC and TSS, and the spread of a single fitted encoder at that species over the pipeline's eleven runs of the fixed weekly arm; reference_runs.csv carries those runs' levels. Every tolerance is derived from those two files rather than written into the script: a level or a margin of the encoder is two fitted runs against each other, so its tolerance is three times sqrt(2) times the spread of the level, and a species is held to the same rule at its own spread, with the check being how many of the 101 sit inside. Run on the deposit with the study's fold map and inner partition on a graphics processor (dev_notes/repro-lisc/ holds the run's files): the series elastic net reproduces to +0.001 TSS and +0.0002 AUC with every species inside; the selection lands on a weekly candidate in nine of the ten outer folds and on a monthly one in the tenth; the procedure's level is 0.0065 AUC and 0.0086 TSS below the pipeline's, inside the tolerances of 0.0087 and 0.0145, with 100 and 99 of 101 species inside their own bands; and its margin over the series elastic net is +0.002 AUC against the pipeline's +0.009, inside the tolerance but, unlike the pipeline's, not separated from zero across species. Documentation - The pairing and inflation text no longer says a paired TSS difference cancels the inflation of a self-selected threshold, or that a level is an upper bound: the inflation is an expectation, and two arms of equal skill can carry different amounts of it. - The width of interval = "nested_cv" is the mean squared error of the bias-corrected estimate it is centred on, from a nested cross-validation one level further down (#73). The paper's width is the mean squared error of the plain cross-validation estimate and adds its bias correction to the centre as a shift with no spread of its own; in the package's benchmark that correction moved from sample to sample by more than the estimate it corrects where the sample was small or the signal absent, and the paper's interval covered a nominal 95% at 0.86 to 0.93 on the two penalised cells rerun at ten repetitions (300 units, no signal and the weekly lag; 174 and 183 replicates) and at 0.84 to 0.96 across the twelve cells at one repetition. Inside every outer training set the same nested cross-validation now runs once more, over the unordered triples of outer folds, which gives the bias-corrected estimate that training set alone would report, and the squared gap between that and the held-out fold's score is what the width is read off; its root is held between the corrected centre's own jackknife standard error and sqrt(K) times it, where the paper holds the plain estimate's root between the plain estimate's. On the package's cheap recovery design (150 units, five outer folds, one repetition, 200 replicates) the interval covers a nominal 95% at 0.96 with a signal and 0.94 without, against 0.955 and 0.895 for the paper's width around the same centre. The paper's width is kept beside it as se_bates in nested_cv, and the benchmark writes both intervals and reports both coverages. Nested cross-validation needs four outer folds now rather than three. On both sides. The binning and the reduction are now one implementation, src/ts_core.cpp and src/ts_calendar.cpp, compiled into the R package by R itself and into the Python extension by CMake. The two languages agree by construction rather than by two implementations being checked against each other after the fact. What each side keeps above it is the boundary: resolving the columns, resolving the zone, and wrapping the result. Four bugs that existed twice, once per language, are closed by that (#1, #2, #4, #5), and a fifth on the Python side with them (#3). The calendar - Bin starts are computed by proleptic Gregorian arithmetic on local time rather than by writing a local midnight and parsing it back. A zone that moves its clock at midnight, such as America/Sao_Paulo before 2019, no longer produces NA bin starts and a failure from inside the reduction, and a year_start landing on such a night is an argument rather than an error (#1). - A bin start is a local time, so reporting it as an instant now has a stated rule: one the clock skipped resolves to the instant the clock jumped to, one the clock repeated to the first of the two. - grain_matrix() in Python takes a tz argument. The same instants and the same zone now give the same answer in both languages, and the fixtures pin it rather than leaving it assumed (#5). - Instants are read at whole seconds in both languages, so two readings a fraction of a second apart are the same reading twice. Guards - Consecutive bin starts must be one bin apart on the grain's own calendar. A bin no unit reaches is never built, so a month missing from the whole record used to pass as four adjacent monthly bins with one gone, in both languages (#4). Not asserted for native, whose bin is the reading itself, nor for a supplied calendar, which declares its own bin lengths. - A day-level statistic requires every calendar day to lie inside one bin, decided from the bins rather than from the grain's name. A supplied calendar cutting inside a day used to give a mostly-NA array in R and a numpy IndexError in Python (#2). - Python no longer rejects a unit called nan and no longer misses a genuinely missing id (#3). Evidence - The pure-R and pure-NumPy implementations are kept as test oracles, tests/testthat/helper-oracle.R and python/tests/oracle.py. Neither package reaches them at runtime; both suites check them against the core on the fixtures and on random series. The NumPy one was written from inst/spec/representation.md rather than from the R source, which is what makes it evidence that the document is complete. - A third fixture series and a tz column in digests.csv: every grain of the aligned series read as a Europe/Vienna clock, and a short series across the night America/Sao_Paulo moved its clock at midnight. The two languages The names, the defaults and the extension points had drifted, so a script moved from one side to the other met them one at a time (#8). One name per concept now, and where the two still differ the difference is a row in inst/spec/representation.md rather than something to be discovered at a call site. - glmnet_learner() is elasticnet_learner(), and its nfolds is n_inner. The name says which model is fitted rather than which package fits it, which is also what the Python side already called it. It is registered as "elasticnet". - Python carries the response and metric registries, so on both sides the response head and the metric are registry entries and the fitting path holds no list of names. learners(), metrics() and responses() list what a session has, in C collation. - stepwise_learner() and select_grain() are on the Python side. The selector's orthogonal polynomial basis and its logistic fits agree with R's poly() and glm() to twelve decimals on the same design; the selection procedure is the same nested one, reporting its estimate under every registered metric. - Python's grain_matrix() returns one representation when one grain is named, whether as a string or as a sequence of one, and a timesift_set() for two or more. Its fold map carries the units it was drawn for, so aligning one is by name there as it already was in R. - The torch encoders take swa and swa_start on both sides, and both refuse a setting the learner does not have rather than ignoring it. A misspelled argument used to be dropped in silence. - select_grain() searches its candidates in the order they were declared. It read them off a join on the names before, which is the collation the rest of the package stopped depending on in this version, so an exact tie on the inner score could fall to a different candidate on a different machine. Build - LinkingTo: cpp11, and flat .cpp under src/, so R compiles the core with no Makevars. - pyproject.toml and CMakeLists.txt sit at the repository root rather than under python/, because a Python source distribution cannot reach above its own project directory and the shared sources must not be vendored into a second copy. The Python build is scikit-build-core and nanobind; the wheel carries python/timesift as timesift. - The wheel depends on tzdata on Windows, which ships no IANA database of its own, so a zoned record bins there as it does everywhere else. - R-CMD-check, pytest and contract run on push and on pull requests, the last of them running both fixture suites against one inst/spec/fixtures/ in a single job. Documentation - One site for both languages at https://gillescolling.com/timesift/: the R reference from the Rd files, the Python reference written from the Python sources by tools/python_reference.py, and inst/spec/representation.md rendered as a page of its own, so the document the two answer to is read where the calls are. The pytest workflow rewrites the Python pages and fails if what is on disk differs, so a docstring cannot change without the page changing with it. - Every public class, method and property of the Python package carries a docstring. Changes in version 0.1.1 New - select_grain(rule = "coarsest_adequate") takes, inside each outer fold, the coarsest candidate whose inner score lies within one standard error of the highest, where "argmax", still the default, takes the highest. Coarser is fewer bins, then fewer channels. The standard error is the spread over the inner folds of each fold's score, so the rule reads nothing the selection did not already compute. Where the inner profile is flat it returns the least storage the record can be kept at without a measured loss inside the training data; where one candidate separates by more than a standard error the two rules agree. Every outer fold now reports the highest inner score (inner_best) and its standard error (inner_se) beside the chosen one, and inner carries each candidate's standard error. On both sides. - select_grain(threshold = "youden") learns a presence-absence cut inside the training data. Each outer fold takes one cut per variable from the inner out-of-fold predictions of the candidate it selected, which the inner search already made for every outer training unit and for no other, by decision_threshold() under the rule named, and reads its test fold at that cut, frozen. The estimate gains a row tss_inner_cut, and the selection carries thresholds and the per-cell cut_scores. tss() takes the cut it is read at as threshold, left NULL for the maximum over cuts as before. On a binormal design with a planted skill of 0.60 the learned cut reads it back within Monte Carlo error, where the maximum over cuts on the same cells does not. On both sides. - An interval for the procedure's risk, by the nested cross-validation of Bates, Hastie and Tibshirani (2024). select_grain(interval = "nested_cv") cross-validates each outer training set again over the remaining folds of the same map, over repeats maps, and reports the estimate's mean squared error the way the paper's Algorithm 1 does, rescaled and bounded as its section 4.3.2 states and centred on its bias correction; final is then the procedure fitted on every unit, whose risk the interval is for. grain_ladder(interval = "nested_cv") does the same for every arm, and paired_contrast(interval = "nested_cv") reads it on the difference between two. The paper's error is a mean of per-unit losses, so a fold's score here is the mean over the variables scorable in it and the variance of that score is its delete-one jackknife variance, which for a mean of per-unit losses is exactly the paper's var(e) / |I_k|. - The estimate and the contrast now name what their interval is for. The across-variable interval is still reported and now carries interval = "variables", lower and upper: it is the spread across the response variables of one dataset, all of them fitted and scored on the same units and folds, and not an interval for a new sample. On a simulated design with a measured truth (150 replicates, five outer folds, AUC) it covered 0.925 of the time with an error-SD to standard-error ratio of 1.23, where the nested interval covered 0.955 at one repetition (#73). - inst/reproduce/schrankogel.R runs the demonstration's own procedure through the public interface: a selection stage searching the study's 33 candidates, a window and a summary each, with select_grain() on the convolutional encoder, over the study's outer fold map and its own inner partition, which ships beside the script as inner_folds.csv. Beside it a series arm fits the penalised model on the weekly coldest-day, mean and warmest-day reading of the record rather than on summaries of it, which is the comparison the demonstration is read against. Every comparison with a published number states its tolerance before anything is fitted and lands in checks.csv; run.meta now records the torch and libtorch versions and the device; and --smoke=, runs the stages on a few of each, names its output smoke_ and compares nothing (#74). - grain_matrix() documents what a day is in a zone that keeps daylight saving time: a wall-clock day, 23 hours on the day the clock goes forward and 25 on the day it goes back, with the week and the month holding them an hour shorter or longer, and 24-hour days on a record kept on a fixed offset or carried in one such as "Etc/GMT-1". Nothing in the binning changed; a test on both sides pins the Europe/Vienna transitions of 2021 at both offsets. - coverage() lays the binning out as a count of readings per (unit, bin), over every bin the calendar tiles the record with. grain_matrix() refuses a record where a unit misses a bin rather than padding it, and this is the table that decision is made on: a logger that started late, stopped early or lost a month is a row with zeros in it, and a bin the whole record skips is a column of zeros. Nothing here fills a cell. On both sides, and pinned by inst/spec/fixtures/coverage.csv. Paired contrasts - paired_contrast() takes each arm whole, as "grain|learner". A learner named alone used to take its best grain, chosen on the held-out scores the contrast was then read off, so the difference was one between two maxima and its p-value optimistic by an amount nothing in the output recorded. It is refused now, with a message naming the whole arm; select_grain() chooses a grain on inner folds and contrasts the selection through compare. occlusion() still reads a learner named alone at its best grain, since it describes a fitted model rather than testing a difference. On both sides (#67). - The interval on a contrast is Student's t on one degree of freedom fewer than there are variables, in place of the normal quantile, which made it about a quarter too narrow at six variables and more than six times too narrow at two. The ladder plot draws its interval the same way (#68). - The signed-rank p-value is read by a method chosen rather than fallen back on, and the row says which: p_method is "exact" below fifty per-variable differences holding no zero and no tie, and "normal", the approximation with continuity and tie corrections, otherwise. The Python side took the exact distribution where the differences held a zero and R the approximation; both take the approximation now. inst/spec/fixtures/contrast.csv pins the interval and the method, and the Python side carries its own Student's t quantile, which agrees with qt() to within about 1e-14 (#68). The response head - A head carries a loss and an activation, and they are what every shipped learner fits toward. The encoders train under the loss and predict through the activation, the per-response learners take the family the loss names, and the combiner minimises the same loss. A registered head other than presence-absence is therefore fitted as itself rather than as clamped binary cross-entropy under another name. - A fit that declares a head argument is handed the head, as one that declares control is handed the control. Both are read from the fit's own signature, at the one point every fold loop fits through, so a learner whose fit takes (x, y) is called with (x, y). - elasticnet(), stepwise() and forest() fit the Gaussian family where the head's loss is squared error. - ensemble() left without a response takes the run's head; one naming a different head is refused before a candidate is fitted. Fitting - elasticnet() deals its inner folds for each response and stratified on it, so a rare outcome is spread over them as evenly as its count allows rather than wherever a plain deal dropped it. A presence-absence response whose inner training sets cannot each hold two of each outcome, the fewest a logistic path is fitted to, is predicted its share among the fitting units, as a response holding one outcome already was, and the fit names it in unfitted. Such a response used to stop the whole run from inside cv.glmnet(), which cost a benchmark cell a replicate it could not recover by rerunning. The inner fold draw is not the one before, so a penalised fit's chosen penalty can differ from an earlier run's. On both sides (#72). - On the Python side the elastic net chooses its penalty on the held-out log loss, weighted by the head's case weights, as cv.glmnet() chooses on the weighted deviance. It was choosing on accuracy, scikit-learn's default, a step function of the penalty whose choice moved by orders of magnitude with the fold draw. - Predicting from a fit whose new targets are missing a static predictor names the column that is missing. An absent column read as one that is not numeric, and the message sent the reader to encode something their table does not hold. - On the Python side models takes one learner as well as a list of them, which is the form R has always taken; a single learner was reaching the fitting layer as a sequence of nothing. - A learner is handed the whole response matrix whether it declares joint or separate, and the three that fit one model per response already did that inside their own fit. The fitting layer was splitting the responses as well, so the flattened block of predictors was rebuilt once for every response: on the published grid, 101 copies of a 47 MB matrix per fold and per arm. multi is now what the learner says it does with the matrix and what a report says of the candidate. The wrapper that reassembled the columns is gone from both sides, and with it the Python CandidateFit. - One fold loop on the Python side, which the run, the ladder and the inner search of a selection all call, in place of three copies; one _check_bins and one _check_grain; and one baseline prediction per fold in the occlusion, which was rebuilding an encoder from its arrays once per response. - A fit refers to the code that made it rather than carrying a copy of it. R writes a closure's body out beside the closure, so saveRDS() on a fit copied every function that produced it and a fit read back after an upgrade rebuilt the old encoder and loaded the new weights into it. A fitted encoder now stores the name of its module builder, and a fit stores the name and the settings of its learner wherever the registry can rebuild it; a learner defined outside any registry still travels whole, because there is nowhere else for it to come from. The Python side stores the encoder's name for the same reason. - A candidate that holds no number on a scorable cell stops the run and names itself. A threshold metric returns NA on a cell whose predictions are NaN or infinite, which is the same NA a one-class cell returns, so a network whose training diverged was reported as an unscorable cell and dropped from the combiner without a word; the stack was then fitted on fewer cells than the mask said and nothing showed it. The combiner is now fitted on every scorable cell and refuses to drop one. score_predictions() refuses the same predictions on both sides. - grain_ladder() and select_grain() take a control, so the training settings of a ladder are given once for the run rather than restated on every neural learner in it. A selection hands the same control to the inner search and to the refit, and a learner carrying settings of its own still overrides it on the ones it names. - A response is fitted under a seed of its own, taken from its name. A learner that covers the responses one at a time is handed one column per call, so a seed spent from one shared stream made a response's fit depend on which responses were fitted before it, and a seed read off the column's position inside its own call gave every response the same one. Either way the model of a response depended on how the columns were batched, and elasticnet() and forest() both did. Fitting a response alone, or beside others, or in a different order now gives the same model, which tests/testthat/test-variable-seeds.R holds to. This changes the numbers those two produce; stepwise() spends no randomness and the encoders cover the responses jointly, so neither moves. - metric takes a registered name or a function of (y, p) everywhere the contract says it does. What scores and what a report prints travel with the fit, so a function reads as in a report rather than as whatever each language calls an anonymous one, and the occlusion profile rescores under the metric the fit was scored with. select_grain() is the one door that takes a name only, because it reports its estimate under every registered metric and selects on a row of that table. - A static predictor enters the design once rather than once per bin. It reaches the array as a channel that does not move across the bins, which is what an encoder reads; flattening the bins was emitting it once per bin, so a penalised fit saw it as often as the grain had bins and a forward search could pick it repeatedly. - A resampling given as an unnamed vector follows the targets into the order they are fitted in. timesift() sorts the targets by identifier before it builds the fold map, so such a vector was landing on whichever unit had taken that position. - elasticnet() takes s, the point of the penalty path to predict at. It also raises what cv.glmnet() raises rather than turning any failure into the response's own mean, and the forward search leaves out a column holding a single value rather than reaching a fitter's error on it. - On the Python side elasticnet() honours its mixing -- the scikit-learn floor is 1.8, where l1_ratios alone selects the elastic net -- and standardises its design before penalising it, as glmnet does. The scaler travels with the fit. The representation boundary - calendar_channels() reads the shared core. The year fraction, the midpoint of the record a bin holds and the calendar the year is read on used to be written once per language, and the two were written differently at the two points that decide the number: R took the phase through sinpi() and the midpoint as a double, Python took sin(2 * pi * x) and truncated the midpoint on datetime64. inst/spec/representation.md now defines all of it, and channels_digests.csv and channels_guards.csv pin it: what is digested is the year fraction, which is arithmetic on the calendar, with a stated tolerance of 1e-12 on the sine and the cosine, which are the platform's library. Both suites read the fixtures. - calendar_channels() refuses a lookback, whose bins are placed relative to a target rather than on the calendar and have no position in the year. It used to reach for a bin_start that is not there. - bind_channels() reads every argument as a representation rather than only the first, and both languages number the arguments from one and raise the same message. A bare array reached the R side and came back carrying attributes describing a binning it had never been through. - A supplied calendar is checked before its bins are read as bins: a bin begins at or before every reading it holds, and a bin's readings are a stretch of the record. A calendar shifted by one boundary, and one interleaving consecutive readings, used to produce an array that looked like any other. - An identifier is written by one rule in both languages. A whole number is its digits, a character id is itself, a factor is its label, and anything else is refused naming the column. - A time column carrying a zone names the calendar to bin by, in grain_matrix(), coverage() and lookback_matrix() alike; a tz naming a different zone beside it is refused rather than silently preferred. - The lookback's target table is read by name, id and at, as every other alignment in the package is, and a lookback reaching the fitting layer without an anchor is refused with a message that says so. - A reading that is not a finite number is refused by the shared core, naming the unit and the instant of the first one, so the rule is written once rather than once per wrapper. digest_array() refuses an array that is not finite for the same reason the contract gives. - A record and any permutation of its rows now reduce to the same bytes. Both reductions walk the record by unit and then by instant; the calendar reduction accumulated in the caller's order, which moved a mean in its last bits under a shuffle that changed nothing about the record. - The native grain is read on the instant rather than on the local clock, so the two readings of the hour a zone repeats when it sets its clock back are two bins. They used to share a local second and were merged into one bin holding their mean, on a grain the contract defines as the record unreduced; the Europe/Vienna native fixture pinned 9599 bins for 9600 readings, and is regenerated. The walk order both reductions accumulate in carries the instant as a third key, so those two readings are no longer a tie the sort resolved as it liked, and a zoned record already in order is no longer sorted again. - A supplied calendar that returns no bin start for a reading, NA in R or NaT in Python, is refused naming the first such reading. It used to reach the core as a bin at the beginning of time, holding the reading its real bin then lacked, with nothing raised; the guard fixture carries the case as missing. - A time column carried in a zone the database does not know is an error in R, as it was in Python. R's zone resolution used to warn and read the clock in UTC. - The contract now says that a lookback's length is measured on the local clock, as everything below the zone boundary is: a one-day lookback ending at a local midnight holds 25 hours of record on the night a zone sets its clock back and 23 on the night it sets it forward. That is what keeps a calendar day whole inside a bin for the day-level statistics; both suites pin it. - The guards above the core name what they refuse the same way on both sides: a missing identifier or instant names its column in Python as it did in R, a duplicated reading is named by its instant in UTC in R as it was in Python and the noun agrees with the count on both, and Python reads bins = 3.0 as R reads bins = 3. A zoned record from before 1970 no longer raises OSError on Windows in Python. A lookback's rows are named by the row names of at whatever kind of row names the frame carries, which is what the contract already said. - The zoned digests have a witness that is not the core. Each oracle reads a zoned series as a clock in its zone and bins that clock with its own calendar, and both suites assert every Europe/Vienna and America/Sao_Paulo digest against it. The core's negative-lag guard, its out-of-range unit index and its bound on the day table are exercised by both suites as well. - The two orderings among the seven statistics, min <= mean_daily_min <= mean <= mean_daily_max <= max and min <= cold_day <= mean <= warm_day <= max, are asserted by both test suites on the core's output and the oracle's. They used to sit in a block of the core that neither build compiled. - grain_matrix() and lookback_matrix() on the Python side refuse a call that names no value column before reaching the compiled core, and a lookback's label reads a zero lag as zero however it is spelled. Names - as_sift() is the coercer to a set of representations on both sides, in place of R's timesift_sift(); the class keeps its name. - The last section of the contract records every public name of both languages, on the side it is on, and each suite reads that section against its own exports, so a name added to one side without a line there fails the suite. Python exports resolve_metric beside get_learner, as the contract had said it did. What a rare response weighs - The response head owns it. A head's weights(y) returns one case weight per cell of the response, and the encoders, the penalised fit, the forest and the forward search all fit under it: the encoders as an elementwise weight on the loss, the penalised fit and the forward search as case weights, the forest as the probability a unit is drawn into a tree's bootstrap. The shipped presence-absence head carries positive_weights(), exported on both sides: each presence weighs the ratio of absences to presences among the fitting units, capped at 50, and each absence weighs one. A head registered with weights = function(y) positive_weights(y, cap = 20) weights every learner by that cap, and a head without weights fits unweighted. - Three learners used to decide this for themselves, three ways. The encoders capped the weight at pos_weight_cap from train_control(), the elastic net and the forest weighted uncapped under weight_positives, and the forward search did not weight at all, so a response present in one target of a hundred weighed 99 in two learners, 50 in a third and 1 in the fourth, in the rare regime the network-against-aggregates comparison is about. pos_weight_cap and weight_positives are gone. On the study's data, whose rarest species has 26 presences in 894 plots, the cap never binds, so the reproduced elastic net is the same number. - Python's forest draws each tree's bootstrap by the weights itself, as ranger's case.weights does. scikit-learn's forest takes a sample weight into its impurities and leaf values, and a tree grown to pure leaves is the same tree under any weight, so the Python forest was fitting a rare response unweighted while the R forest was not. - The forward search fits the quasi-binomial family and reads its criterion off the deviance, which is the same number the binomial family reports for a 0/1 response and takes a case weight that is not a whole number without a warning. The Python forward search runs its iteratively reweighted least squares under the same case weights. Fitting under a grouping - A fit is checked once, on the timesift_fit, against the bins and the channels it was made on, before any learner sees the representation it is asked to predict. A calendar grain's bins are named by their instants, so a record from another period with the same number of bins is refused by the first bin that differs rather than read by position, as cnn() fitted on September to November used to read March to May. A lookback's bins are relative to each target and predict any period. The check that each learner carried is gone; every learner, including one of your own, is held to this one. On both sides. - The grouping an outer fold map keeps whole reaches every split drawn inside a fit. A [fold_map()] built with group carries it, a fit that declares a group argument is handed the grouping of the units it is fitted on, and the encoders' inner validation set, the elastic net's inner folds and select_grain()'s inner map are all dealt by it. Under grouped_cv() with target_time a unit kept whole across the outer folds used to be split across the inner ones, so the epoch and the penalty were chosen on rows the fit had already seen a near-copy of. On both sides. The elastic net deals its own inner folds now, through fold_map(), rather than leaving the draw to cv.glmnet() or to scikit-learn. Fixes - fold_map() deals with one round-robin counter that runs on from each stratum into the next, on both sides, so the folds are equal in size to within one unit whatever v is and leave-one-out is every unit in a fold of its own. A counter restarted in every stratum used the first labels once more than the last in every stratum, and a stratum smaller than v never reached the last labels at all: ten folds over thirty units came back holding two to five units each. The fold-map fixture, and the mask and the inflation built on it, are regenerated. - A fold map given as resampling that holds one fold is refused, and one built by fold_map() keeps its grouped, seed and strata when it reaches the run, so a grouped design reports as one. A run under which no (response, fold) cell is scorable stops before anything is fitted, and ensemble_fit() handed no scorable cell says so rather than fitting a combiner on nothing. - The threshold metrics check that the response is 0/1 before coercing it, on both sides. They used to coerce first, which read 0.6 as 0 and scored a response that was never binary. - A torch fit puts torch's generators back where it found them, as it does R's. - .simplex_weights() stops at a relative gain of 1e-14 rather than 1e-12, on both sides. Near the minimum the loss is flat to second order, so the stop settles the gradient to about the square root of the tolerance, and the looser one left the carried members' gradients apart by more than a millionth on some boards. - The within-variable-then-across-variables mean is one function, .cell_means(), read by the ladder, the report, the stack's member scores and the selection's estimate; the five copies had four spellings. - Small edges: models and learners take a character vector of registered names, as the docs said; a fit read back through its reference keeps a setting held at NULL, so a forest refits with its default mtry; train_control(early_stopping = Inf) is a patience that never runs out; predict() on a timesift fit holds the new targets to one row per identifier, as the fit's own were held; the elastic net refuses a design of one column in its own words; and grain_contrasts() names the reference it cannot find rather than failing on a length-zero test. - Python: a learner pinned to a representation that shares a sift member's label but not its definition is refused, as R refuses it, rather than fitted on the sift's array. - Python: auto_grains() counts the bins in the zone the time column carries, the clock the representation is then binned by. It used to probe in UTC, so a zoned record could be offered a grain it does not support or denied one it does. - Python: grain() takes a supplied calendar as R does, reported as custom; multigrain() refuses one on both sides. - Four differences between the languages are gone rather than recorded. grouped_cv() deals the groups unstratified on both sides. A lookback is labelled x lag on both, and a numeric span as a duration. lookback() refuses a span that is not positive when it is built. forest() and elasticnet() derive one seed per response from the response's name in Python as in R, so two species no longer share their bootstrap draws. - Python: feature_matrix() names unnamed rows from 1, as every other unit label does, and a fold map prints every fold level it holds. - A response given as a data frame with two or more non-numeric columns is refused, naming them, rather than read through the first and silently stripped of the rest. - A presence-absence response holding a missing value says so on the Python side, rather than reporting it as a value that is not 0 or 1. - An arm is matched whole against the ladder's labels rather than split at the |, so a learner or a sift whose own name holds one is still found, on both sides. - select_grain()'s inner is checked before it is coerced, so a value that is not a count is named in the error rather than printed as NA. The selection benchmark - The oracle and the single-loop truth are averaged over the outer training sets, as the procedure's truth always was. Every candidate is scored on the deployment sample through the ladder's own per-fold fits, and the three arms are read off that one grid, so the regret compares the procedure with the best fixed candidate under the same training draws. The oracle used to be fitted on the first outer training set alone, so the regret carried a maximum over twelve one-draw estimates, and the single-loop bias a truth read on one fold against a score read on five. inst/benchmark/design.R states the definition and what of the selection term remains. The interval whose coverage is read is Student's t, as the package's are (#65). Packaging - inst/CITATION cites the package, and citation("timesift") prints it; the methods article's entry is added once it has a DOI. codemeta.json is generated from DESCRIPTION with codemetar and regenerated at a release. DESCRIPTION declares Language: en-GB, and inst/WORDLIST holds the proper nouns and API names the spell check would otherwise flag. tests/spelling.R runs the check against it and fails on a word it does not know, and the R-CMD-check workflow sets NOT_CRAN so it runs there; it is skipped on CRAN, whose dictionaries are not this machine's (#70). - The Python side is checked as a distribution. The python-dist workflow builds the source distribution, installs it into a clean environment, and runs the suite it ships from the unpacked tarball, which is what holds the manifest in pyproject.toml to the sources. It builds wheels for Python 3.11 to 3.13 on Linux, Windows and macOS with cibuildwheel, each tested on an interpreter that did not build it, and passes every artefact through twine check. A published GitHub release uploads them to PyPI by trusted publishing (#69). - The committed site is held to its sources. build_site.R writes a digest over everything the site is rendered from into docs/site-digest.txt, and the site workflow recomputes it from the checkout on every push, so a push that changes the reference, the news or an article without rebuilding the site fails there rather than serving last week's pages under this week's sources. The site stays built locally, because the build post-processes its figures in a way a plain build in CI would not. - The Python floor is 3.11. The sklearn extra pins the scikit-learn release where the mixing parameter alone selects the elastic net, and that release has no build for 3.10, so the 3.10 job had failed at install on every push. The test extra carries pandas, so the suite runs timesift() on a data frame with a categorical identifier, a nullable-integer response and a zone-aware time column rather than skipping that test on every runner. - DESCRIPTION is where the version is written, and pyproject.toml reads it from there. - .gitattributes holds the repository to one line ending. - The wheel ships py.typed. Changes in version 0.1.0 timesift() is the whole entry point. A table of targets, a table of time-stamped series belonging to them, and one call builds every candidate representation, fits the learners that can read each one, scores them all on one set of held-out folds, and stacks the out-of-fold predictions. fit <- timesift(plots, logger, y = starts_with("sp_"), id = plot_id, time = datetime, models = c(elasticnet(), forest(), cnn()), sift = grains("day", "week", "month")) summary(fit) The version reads 0.1.0 because this is a first release: the package fits any time-varying record against any prediction target, where its predecessors fitted a climate record at a climate grain. Species distribution modelling from microclimate loggers is the application it ships defaults for, and the Schrankogel grid it was built on still reproduces from inst/reproduce/schrankogel.R. The four concepts - A target is one row to predict, a series is the long record belonging to those rows, a representation is how that record becomes an array, and a learner is a fit and a predict pair. A candidate is one (representation, learner) pair, and every candidate emits an out-of-fold prediction for every scorable cell over the same folds. Comparison, ensembling and importance read only those predictions. - y, x and static take tidyselect expressions over their own table, and starts_with(), ends_with(), contains(), matches(), all_of(), any_of(), everything() and where() are re-exported rather than redefined. A column of targets that is neither the response nor the identifier nor the anchor is a predictor only where static names it. - predict() on a fit rebuilds each member's representation for the new targets from the settings its own arm was built with, and combines them through the ensemble. Representations - native(), grain(), multigrain() and lookback() are what a representation is before any record has been read; grains() and lookbacks() are sets of them, and grains("auto") reads off the record every named grain it gives at least two bins. - c() combines learners, or representations, into a set: c(elasticnet(), forest()) and c(grains("day", "week"), lookback("30 days")). A set handed back to c() splices, so a set can be added to rather than rewritten. models and sift take that, a bare spec, or a list. - lookback() is a fixed span of record ending a fixed lag before each target's own instant, which is what a unit carrying several targets through time needs. It is one entry point in src/ beside the calendar reduction, so both languages read it from the same implementation, and lookback_matrix() exposes it directly. - build_representation() is the one place the fitting layer turns a representation and the two tables into an array, so a run and a prediction on new targets reach a record the same way. - A learner given data = grain("week") runs at that representation alone; left open it runs across the whole sift. A learner that reads a block of features and one that reads a sequence say so, and a pairing neither can carry is reported by name rather than fitted. Learners and training - elasticnet(), stepwise(), mlp(), cnn() and rescnn() drop the _learner suffix and gain data. forest() joins them, a probability forest on ranger in R and on scikit-learn in Python. reads and multi are learner()'s, where a learner of your own declares what it reads and whether one fitted model covers every response. - train_control() is the one place a training setting is defaulted. The architecture constructors carry architecture, a run gives one control to every neural learner, and a learner given its own control overrides that on the settings it names. - A learner declares whether one fitted model covers every response or one is fitted per response. Either way a candidate emits one [target, response] matrix, so nothing above the learner layer has to know which it was. Resampling and the ensemble - cv() and grouped_cv(); resampling also takes a fold vector or a fold_map() result, which is how a split the package has no constructor for reaches the same fitting path. - ensemble() fits non-negative weights summing to one on the out-of-fold predictions alone, minimising the response head's own loss over the scorable cells, solved by an exponentiated- gradient loop in the package. "mean", "median" and "weighted" combine without fitting. ensemble_fit() is handed the predictions, the response, the mask and the fold map, and never a model. - summary() reports each candidate's mean, how many responses it scored highest on, whether one model covered them, and the level the combination reached under the weights it reached it with. Names - native() rather than raw() and lookback() rather than window(), which would have masked base::raw() and stats::window(). occlusion() is one generic over a run and a ladder, and bin_occlusion() is gone. ensemble_learner() is gone with it: it fitted its members and averaged them, which the stack does over any candidates at all and with the weights fitted rather than assumed. - The bin's name is the grain throughout, in both languages. First release. The package builds the representation, fits at every grain, and reports where predictive skill saturates. Representation - grain_matrix(): reduces a long table of sensor readings to a [unit, bin, channel] array at one of seven temporal grains, from the unreduced record to a single value per hydrological year. Naming several grains returns one representation per grain; passing a function bins by a calendar the package does not carry, such as seasons cut at the equinoxes. - Bins follow the calendar, so a month is 28, 30 or 31 days and a week starts on a Monday, and every (unit, bin) cell is asserted to hold readings. - A bin the record does not cover for its whole calendar span is reported on bin_partial and kept or removed by the partial argument, so a record that begins away from a bin boundary says so rather than carrying a short bin that looks like any other. - Seven statistics: mean, min, max, the day-level cold_day and warm_day, which reduce each day to its own mean before taking the extreme over days, and mean_daily_min and mean_daily_max, which take the mean of the daily extremes. - calendar_channels() and bind_channels() supply the position of each bin in the year to an encoder that would otherwise pool it away. - feature_matrix() brings an already-reduced feature table in as an arm of the same ladder. - Gaps, duplicated (unit, time) pairs and missing values are errors rather than silent padding. Fitting and scoring - fold_map() and scorable_cells(): one split read by everything that scores, and the mask of cells a score is defined on, computed from the response and the fold map with no model involved, so every arm shares one denominator and every paired difference runs on matched cells. - grain_ladder() fits every learner at every grain, plot() draws the curve, and paired_contrast() compares two arms inside each cell both scored. - bin_occlusion() holds each bin of the record back and rescores, without refitting, so a fitted model says which part of the year its skill rests on. - grain_contrasts() fits score ~ grain + (1 | variable) + (1 | fold) and compares every grain against a learner's best by Dunnett's procedure. Needs lme4, lmerTest and emmeans. - tss(), roc_auc(), kappa_score(), model_agreement() and decision_threshold(). - tss_inflation() measures how much a self-selected threshold inflates a reported level at the user's own presence counts, and implied_skill() inverts that map to say what population skill a level actually read is consistent with. Learners - elasticnet_learner() and stepwise_learner() on the flattened representation, both redoing their selection inside whichever units they are handed. - mlp_learner(), cnn_learner() and rescnn_learner(), joint multi-label encoders on torch, sharing one training recipe. - ensemble_learner() averages several members' predicted probabilities before the threshold is chosen, and the torch learners take swa = TRUE to average the weights of their tail epochs. - learner() takes a fit and a predict pair of your own; register_learner(), register_response() and register_metric() extend the three registries the fitting path reads. Reproducing the study - inst/reproduce/schrankogel.R runs the published grid from the Zenodo deposit it was built on, asserting the plot count, the species count, the cell count and the bin count of every grain before fitting anything. See vignette("reproducing-schrankogel"). The cross-language contract - inst/spec/representation.md is normative for both the R and the Python implementation. - inst/spec/fixtures/ carries a synthetic series and the digest of every grain-by-statistic combination; both test suites assert against the same digests. - The Python side implements the representation, the folds, the mask, the metrics, the ladder and the same three encoders.