Changelog¶

0.5.1¶

This version includes the following changes:

full_ddm no longer discards the response times between t - st/2 and t (#1292; by @EItanm1999). The blackbox likelihood behind full_ddm (hddm_wfpt) reads st as the full width of the non-decision-time distribution, so its support starts at t - st/2, yet on 0.5.0 and earlier the log-likelihood was floored for every response in (t - st/2, t], biasing t downward (profile maximum 0.36 against a true 0.40 in a 2000-trial example). Refit any full_ddm model whose fastest response times fall within st/2 below t.
A likelihood can declare where its response-time support starts (#1292; the configurable edge was proposed by @EItanm1999 in #1344, which this supersedes). The optional "ndt_edge_shift": {"param": "<name>", "scale": s} entry (the hssm._types.NDTEdgeShift TypedDict) places the floor at t - s * param; it is accepted in a register_model likelihood dict, as a model_config override on HSSM(...), and as make_distribution(ndt_edge_shift=...). full_ddm declares {"param": "st", "scale": 0.5}; a likelihood without the key keeps the floor at t, and a model without t has no floor. See Declaring where the response-time support starts.
RLSSM decision processes carry the same declaration (#1358). register_ssm(..., ndt_edge_shift=...) stores it with the decision-process spec, RLSSMConfig.ndt_edge_shift picks it up from there (as well as from a ModelConfig decision process, and, when the config is built from an ssm-simulators preset, from the spec of the decision process the preset names), and RLSSM forwards it to the RL likelihood.
Default initial values that fall outside a parameter's bounds are clamped into them (#1293; by @EItanm1999). A default start outside the declared bounds previously gave a NaN starting point (100% divergences under numpyro, "Bad initial energy" under PyMC) for bounded Uniform priors, bounded safe Intercept priors on identity-link regressions, and models whose t bounds start above 0.025; such defaults are now moved just inside the bounds, logged at INFO with a pointer to sample(initvals=...). Initial values you supply yourself are used as given, and a regression with a log or gen_logit link keeps its link-scale default.
Documentation: how bounds act on an identity-link regression (#1343; by @EItanm1999). The link_settings docstrings, RegressionParam.validate and the priors how-to now say that a bounded parameter used in a regression with the identity link does not have its bounds built into the parameterization: a draw that puts any observation's value outside them is penalised by the likelihood (apply_param_bounds_to_loglik) rather than mapped back inside, which is usually harmless when the parameter sits well inside its bounds, and that "log_logit" composes the regression on the link scale instead. The runtime warning proposed in that PR was not adopted.
The p_outlier regression and random-slope tutorials execute again (#1366, closes #1363). Both had failed at model construction since the 0.5.0 safe-prior guards. tutorial_p_outlier_regression now gives v the gen_logit link through link_settings="log_logit", since the guard refuses an unbounded identity-scale default for a group-only intercept on a bounded parameter; random_slope_safe_priors now shows HSSM rejecting the legacy free-mean group priors under noncentered=True instead of building that graph, and keeps every other comparison.
A JAX version of the analytical DDM likelihood (#1421, towards #1420). hssm.likelihoods.logp_ddm_jax ports logp_ddm (Navarro & Fuss, 2009) to JAX with the same parameterization, response coding, series switch and lower bound. Tests pin it to logp_ddm (per-trial difference at most 1e-10 in float64, including the floor and response times below t) and keep its float32 gradients finite. The DDM's analytical entry declares it as its jax_loglik, which is how RLSSM uses the analytical DDM (next item).
RLSSM can use the analytical DDM (#1422, towards #1420). hssm.rl.RLSSM(data, model="2AB_RW_DDM", decision_process_loglik_kind="analytical") fits the ssm-simulators Rescorla–Wagner preset with the exact DDM likelihood (logp_ddm_jax, #1421) instead of its LAN, and does not load the network. get_rlssm_model_config, RLSSMConfig.from_ssms_model and register_rlssm_model take the same argument, and the LAN stays the default. A likelihood entry declares its JAX implementation under the new optional jax_loglik key, as the DDM's analytical entry now does; plain hssm.HSSM still evaluates loglik. Behaviour change: decision_process_loglik_kind used to be a label only. "analytical" now switches the likelihood, a decision process without a JAX implementation of the requested kind (e.g. angle with "analytical") raises, RLSSM(model_config=..., decision_process_loglik_kind=...) raises when the two disagree, and any kind other than "analytical", "approx_differentiable" or "blackbox" is rejected (register_rlssm_model used to store any string).
Documentation: using the analytical DDM in an RLSSM (#1423, towards #1420). A new how-to fits a simulated two-armed bandit with decision_process_loglik_kind="analytical", compares the fit with the LAN, and checks logp_ddm_jax against logp_ddm. See Use the analytical DDM in an RLSSM.
The HSSM-side two-armed-bandit learner handles -1/1 responses (#1419, closes #1418). hssm.rl.likelihoods.two_armed_bandit.compute_v_subject_wise and its copy in hssm.likelihoods.rldm_optimized used the response as the arm index, so with -1/1 responses JAX wrapped -1 to arm 1 and every trial updated arm 1, silently giving wrong drifts. The response now maps to arm response > 0. Data coded 0/1 and the ssms.rl presets (2AB_RW_DDM, ...) were not affected; refit any custom RLSSM that used this learner with -1/1 responses.
lba2, lba3 and lba4 start sampling at a valid point (closes #1426). Their default start was A = 2.0, b = 1.2, where the likelihood's b > A check fails, so with p_outlier=None or 0 the PyMC sampler stopped with "Bad initial energy" and numpyro chains froze in place (r̂ above 3 on simulated lba4 data); under the default lapse the LBA term contributed no gradient at the start. Each config now declares A's default prior, unchanged at HalfNormal(sigma=2), with initval=0.5. Priors you supply for A replace this default as before.
Opt-in MLflow tracking of fits: hssm.track() (new module hssm.tracking, optional extra hssm[tracking]). Wrapping a workflow in with hssm.track(experiment=..., run_name=...): records HSSM(...), sample()/vi() and save_model() as one MLflow run following the HSSM-ecosystem run schema (phase=infer): model, likelihood kind, network file and its HuggingFace commit (and the missing-data network's, under missing_data_network_file/missing_data_hf_revision), sampler settings, n_trials/n_subjects, a spec_sha256 of the model specification and a data_sha256 of the data; metrics sampling_seconds, divergences, r_hat_max, ess_bulk_min, ess_tail_min; artifacts model_spec.json, summary.csv, traces.nc, data.parquet — the fitted data, covariates included (and model.pkl with log_artifacts="all"). Record your own variables with params= or run.log_param(...) so runs can be compared on the dimensions you are varying. dataset_name= registers the observed data as an MLflow dataset, filling the run's Datasets panel with its schema and row count. Variational fits add method, niter, backend, vi_seconds, elbo_final/elbo_max and vi_traces.nc. With nothing configured, runs go to a local mlflow.db. The lineage_id that ssm-simulators and LANfactory attach to training data and published networks is read from the repository's manifest.json, so one query lists every fit built on a dataset. Off by default; nothing changes for users who do not call track(). Failures while recording a fit are logged, never raised; an unreachable tracking server raises when the block opens, before the fit starts. See the "Track fits with MLflow" how-to.
Rebuild a tracked fit: hssm.load_run(run_id). Reconstructs an HSSM model from a run's model_spec.json and data.parquet and restores its traces.nc / vi_traces.nc. Hugging Face networks, including a missing-data network, are fetched at the revision the run recorded, even when the working directory holds a file of the same name. Runs now carry model_class, spec_restorable and rebuild_artifacts tags (the last set once data.parquet and model_spec.json are stored, so hssm.list_runs needs no extra request per run); spec_restorable=false marks runs whose spec holds arguments recorded as text (bmb.Prior or Link objects, custom functions), which load_run refuses with a clear error.
List tracked fits: hssm.list_runs(). Returns a DataFrame of the runs recorded by hssm.track, newest first, with run_id, run_name, experiment, start_time, status, user, model, loglik_kind, dataset_name (from a new dataset_name tag, so inputs you log yourself are not mistaken for it), n_trials and restorable (whether hssm.load_run can rebuild the run). Runs logged by other tools on the same server are left out.
RLSSM can use the LBA (towards #1420). lba2, lba3 and lba4 declare a JAX likelihood (jax_loglik): hssm.likelihoods.analytical.logp_lba_jax, the lba4 JAX code generalised to any number of accumulators. An ssms.rl.ModelConfig with decision_process="lba3" and the RescorlaWagnerRaceDrifts learner then fits through RLSSMConfig.from_ssms_model(..., decision_process_loglik_kind="analytical"); ssm-simulators has no Rescorla–Wagner + LBA preset yet (see Use the analytical DDM in an RLSSM). RLSSM now starts each non-regression parameter at the config's params_default (for ssm-simulators presets, the simulator's defaults) instead of its prior's midpoint, unless you specify its prior or HSSM's shared defaults set its start (a, t, v, p_outlier). The midpoint start put A and b both at 0.5, where the LBA's b > A check fails; for the existing presets it moves rl_alpha, scaler, theta and the softmax beta to the ssm-simulators defaults. The PyTensor lba2 and lba3 likelihoods computed their normal CDF with a float32 sqrt(2) even under float64, an error of up to about 3e-7 in log p; they now agree with logp_lba_jax to 1e-10, which the jax_loglik parity test checks. New tests check the model's log-likelihood against a NumPy learner for 2, 3 and 4 choices, and parameter recovery for 3 choices.
plot_quantile_probability with ellipses no longer fails on other legend entries or cuts off ellipses (closes #1430, closes #1431). With predictive_style="ellipse" or "both", the ellipse colours were read from every legend entry in the axes, so drawing into axes that already held, for example, the bands of plot_predictive (which plt.gca() returns right after a plot_predictive call in a script or marimo) failed with AttributeError: 'FillBetweenPolyCollection' object has no attribute 'get_color'; only lines are read now. The ellipses were also added without rescaling the view, which stayed fitted to the observed quantiles and cut off ellipses reaching beyond them; the view now includes them, and axis limits you set before the call are kept.
The lapse mixture no longer drops the lapse mass on missing-RT and omission rows (closes #1322). The lapse mixture evaluated the lapse RT density at the -999.0 placeholder of every missing-RT and deadline row, so the default Uniform(0, 20) lapse contributed -inf (no lapse mass at all) and other lapse priors contributed a meaningless finite value, biasing p_outlier and the SSM parameters whenever such rows were present. The lapse term is now marginalised like the SSM term: observed rows keep the lapse density at rt, missing-RT rows (choice observed, no deadline) use 1 / n_choices, and omission rows use the lapse survival probability at the deadline, 1 - CDF_lapse(deadline). Lapse priors without a PyMC logcdf now raise a clear error on deadline data. The missing 1 / n_choices factor on observed rows is tracked separately in #1323. Refit any model with a lapse (p_outlier, 0.05 by default) whose data has missing-RT or deadline rows.
Documentation: compile_logp() returns the log-posterior; distribution_utils API page fixed (closes #1436). HSSM.compile_logp() compiles the model's log-prior plus log-likelihood, plus any Bambi potentials (the unnormalized log-posterior), as a function of natural-scale parameter values, not the log-likelihood alone; the tutorial, renamed "Compile a log-posterior function", and the method's docstring now say so. It also gives pip install zeus-mcmc for zeus: hssm[notebook] installs no zeus, since notebook is a dependency group, not an extra. The distribution_utils API page now lists make_hssm_rv in place of the nonexistent make_ssm_rv, and adds make_distribution_for_supported_model.

0.5.0¶

This version includes the following changes:

plot_model_cartoon displays posterior uncertainty as graded bands by default (#1124), adopting the same uncertainty="band" | "samples" | "both" | None vocabulary as plot_predictive (#1122): pointwise quantile bands on the RT histograms, fan-chart ribbons on the decision bounds, a graded drift-quantile fan, graded non-decision-time spans (replacing the per-draw vertical-line smear), and a starting-point whisker. uncertainty=None reproduces the previous mean-only look; the default call now runs per-draw simulations (seconds). New parameters: hdi, alpha_mean, alpha_uncertainty, hist_height, legend; plot_predictive_mean/plot_predictive_samples are deprecated with a FutureWarning. The previously documented-but-ignored bins, step, colors, linestyles, and linewidths parameters now work. With uncertainty display on, the drawn mean is the pointwise predictive mean across draws (it cannot exit its own bands) rather than the plug-in simulation at the posterior-mean parameters. The figure uses a three-way color encoding — model geometry stays neutral black (color_model), predictions and data use the predictive palette. Note for regression/hierarchical models: the drawn geometry of each posterior draw derives from that draw's trial-mean parameter vector (see the correctness entry below). See the new "Model cartoon plot gallery" tutorial.
plot_model_cartoon correctness follow-up (#1125). Each posterior draw's geometry now derives from one coherent θ vector — the across-trials mean by default, or one trial via the new obs= parameter (which then conditions every simulated layer on that trial; the observed histogram stays pooled, with a warning). New random_state= (int or np.random.Generator) makes the whole figure reproducible: which posterior draws are displayed (previously chosen by the unseeded global RNG before any seed applied), every simulator seed, and the trajectories — facets consume successive segments of one stream. Geometry and trajectory simulations now receive a max_t horizon derived from the x-limits instead of the simulator's 20 s default (previously ~75% of every boundary polyline lay outside the axes); the noisy RT simulations deliberately keep the long horizon so defective densities are not silently re-normalized by censoring. Trajectories are noisy realizations of the reduced reference θ, so their crossing markers land on the drawn boundary by construction. n_trajectories, xlims, and ylims are promoted to documented parameters (None keeps each renderer's defaults). Bug fix: the >2-choice renderer simulated each draw's RT histogram with n_samples=1 (silently ignoring n_reps) and one seed shared across draws, so its uncertainty bands summarized degenerate, noise-correlated histograms. Deliberate value changes for a fixed random_state: a new seeding protocol, trial-mean instead of trial-0 geometry for regression models (intercept-only models unaffected), and reference-θ trajectories. Documented caveat: geometry is nonlinear in θ, so the trial-mean curve is not the mean of per-trial curves and need not sit mid-band — the reduction is a display convention.
plot_model_cartoon accepts hist_height="auto" (#1127): fits the tallest RT-histogram curve to 90% of the vertical headroom between the histogram baseline (ribbon-aware — expanding bounds are cleared too) and the upper y-limit, so histograms can never overrun the axes; resolves per facet. Its mirror image, ylims="auto", keeps the raw density scale and grows the frame around the content instead (never below the default limits; the two spellings are mutually exclusive). The defaults remain unchanged. The >2-choice renderer now also draws per-choice drift-uncertainty cones in band mode (graded fills in each accumulator's color; previously slope uncertainty appeared only as per-draw spaghetti in the samples display).
Safe priors now zero-center matching group-specific effects (#1224). HSSM matches each Formulae group expression to an exact common-term counterpart and assigns the generated group deviation mu=0. This removes disconnected group-mean variables under non-centering and the redundant common/group location under centering. Explicit priors are not rewritten, including Bambi's common and group_specific wildcards; parameterization diagnostics now describe the graph Bambi actually builds. Item 12 below adds compatibility validation and defines population-location ownership for unmatched group-only effects.
hssm.Link now constructs Bambi built-in and custom links (#1231). Standard links delegate correctly to Bambi, while HSSM's bounded gen_logit behavior remains unchanged.
Safe common-intercept priors now recognize every identity-link spelling consistently (#1232). An omitted link, "identity", bambi.Link("identity"), and hssm.Link("identity") all retain response-scale priors, including HDDM-derived priors; transformed links use a coefficient-scale Normal(mu=0, sigma=0.25). Item 12 below extends semantic identity handling to generated unmatched group-only intercepts and defines their transformed-link predictor-scale behavior. See Link functions and safe priors for the underlying model-design logic.
Invalid model-level prior and link presets now fail fast (#1233). prior_settings accepts only "safe" or None, and link_settings accepts only "log_logit" or None; wrong-case strings, booleans, mappings, and other unsupported values now raise a clear ValueError instead of silently changing prior, link, initial-value, or display behavior. The API documentation now clarifies that both presets act on regression parameters: prior_settings=None delegates missing regression-term priors to Bambi but leaves HSSM's simple-parameter defaults unchanged. The Poisson-race tutorial has also been migrated to marimo and no longer presents prior_settings as a prior dictionary.
hssm.load_data now returns a pd.DataFrame unconditionally (#1146). Its dataset argument is required, and the return type is no longer pd.DataFrame | str, which forced type-checkers (and users) to narrow away a str branch that existed only to print the dataset listing. Use the new hssm.list_data() to get the names of the built-in datasets as a tuple[str, ...] (mirroring hssm.list_models()). Breaking: hssm.load_data() with no argument now raises TypeError instead of returning a listing string.
New built-in model: gamma_drift (#1248) -- a drift-diffusion model whose drift is a constant v plus a gamma-shaped bump with signed peak amplitude c, its time course set by shape and scale. This is the conflict-task (DMC-style) family: a negative c pulls evidence toward the wrong boundary early in the trial, then decays back to baseline. Parameters v, a, z, t, shape, scale, c.
New built-in model: gamma_drift_angle (#1296) -- gamma_drift with a linearly collapsing (angle) decision bound. Parameters v, a, z, t, theta, shape, scale, c in ssms registry order (theta at index 4 -- the order is the ONNX input contract), bounds set to the network training box, and a registry cross-check test asserting both against ssm-simulators. The LAN artifact ships at the root of franklab/HSSM; its parameter-recovery report (published alongside) documents identifiability limits of shape/scale when the collapsing bound truncates the drift bump -- see the model card on the Hub.
New built-in model: angle_extended (#1298) -- the angle model with drift bounds widened to (-6, 6) for designs producing strong evidence. Bounds and parameter order asserted against the ssm-simulators registry; the LAN passed a 240-fit parameter-recovery sweep with zero coverage or bias failures (report published beside the artifact). Raises the ssm-simulators floor to >=0.14.0, the first release containing the model.
Safe priors now preserve a unique unmatched group-only population location and reject ambiguous or unrepresentable specifications (#1225). When HSSM generates a prior for a group term with no exact common Formulae counterpart, the group distribution owns the population location. HSSM therefore preserves its location-bearing hierarchy and centers that generated term even when the model or component requested non-centering, because Bambi's current non-centered construction omits mu. For previously generated unmatched Normal terms under default non-centering, this intentionally changes the likelihood by reconnecting the location that Bambi had discarded. Exact common/group matches remain zero-mean deviations and honor the requested parameterization (#1224). Repeated unmatched expressions now fail when safe generation has no unique owner. Explicit priors are never rewritten, but HSSM raises before Bambi when a group prior cannot be represented faithfully and warns about buildable centered specifications with multiple free owners of one location. Identity-linked group-only intercepts retain their response-scale hierarchy for every identity spelling (#1232); transformed links use a hierarchy on the linear-predictor scale. For generic models, safe generation now fails closed when such an identity-linked intercept has finite response bounds, rather than inventing an unqualified unbounded hierarchy (#1269). Explicit priors remain authoritative, and exact/black-box HDDM likelihoods retain their calibrated natural-support hierarchies. Follow-ups cover upstream location-aware non-centering (#1268, Bambi #1003) and broader numeric regression-term semantics (#1271).

Breaking changes that require migration:¶

Safe priors now refuse to invent a hierarchy for a bounded parameter (#1269). With the default prior_settings="safe", a group-only intercept — a formula with no matching common intercept, such as v ~ 0 + (1|subject) or v ~ -1 + (1|subject) — on an identity-linked parameter that carries any finite bound now raises ValueError instead of silently generating an unbounded hierarchy. HSSM's generic group hierarchy is unbounded, so applying it to a bounded parameter through an identity link would place a response-scale default outside the parameter's support.

This affects every approx_differentiable (LAN) model, including existing ones such as angle, weibull, levy and ornstein — not only models new in this release. Exact and blackbox ddm, ddm_sdv and full_ddm keep the previous warning. The ordinary hierarchical form v ~ 1 + (1|subject) is unaffected.

To migrate, pick one: - keep a common intercept: v ~ 1 + (1|subject); - use a support-respecting link, e.g. link_settings="log_logit"; - supply an explicit group prior for the term (or a group_specific wildcard) — note this currently also requires noncentered=False for that parameter; - set prior_settings=None and specify the complete prior policy yourself.

Explicit group-specific priors that bambi cannot represent faithfully now raise (#1225). Previously a logged warning, these are now an aggregated ValueError raised before sampling. Under the default non-centered parameterization, an explicit group prior supplying a mu hyperprior cannot be represented, because bambi builds offset * sigma and drops mu. Pass noncentered=False for that parameter to keep the prior as written.
ddm_sdv's approx_differentiable sv bound widened from (0.0, 1.0) to (0.0, 2.5) (#1230), matching the range the shipped network was actually trained on. The bound feeds default priors and initial values, so fits of existing ddm_sdv models may differ from 0.4.0.
hssm.load_data() now requires its dataset argument (#1146) and always returns a pd.DataFrame. Use the new hssm.list_data() to list the built-in datasets.

Dependency changes:¶

ssm-simulators floor raised to >=0.14.0 — the first release containing angle_extended. 0.14.0 also fixes omission_p / choice_p_no_omission, which were computed against the wrong column and read as identically zero on deadline models.

0.4.0¶

This version contains major breaking updates for HSSM. Please read the release notes below to migrate to HSSM 0.4.0.

Major new features:¶

A new RLSSM class has been added to support reinforcement learning sequential sampling models.
Attentional Drift Diffusion Model (aDDM): new hssm.aDDM model (with aDDMConfig) built on a vendored, differentiable JAX first-passage-time likelihood. Supports per-trial fixation covariates (r1, r2, flag, sacc_array, d, sigma), trial-wise regression / hierarchical priors on the core parameters (eta, kappa, a, b, x0), a sampled non-decision time t, and posterior-predictive checks conditioned on the observed fixations (via the aDDM simulator in ssm-simulators). See the "Attentional DDM" tutorial and scripts/addm_parameter_recovery.py for an end-to-end recovery check.
Analytical four-choice LBA (LBA4), with a JAX likelihood enabling JAX/NumPyro-compatible sampling and gradients for the analytical LBA4 model.
Per-parameter centered vs. non-centered parameterization. Pass noncentered to HSSM(...) as a dict keyed by parameter name (e.g. noncentered={"v": False, "a": True}), or set a per-prior noncentered field on a group term's prior (dict or hssm.Prior) to override the model-level choice. Requires Bambi >= 0.19. See the "Per-parameter centered vs. non-centered parameterization" tutorial.

Breaking changes that require migration:¶

Dependencies have been streamlined to support PyMC 6.0+, pytensor 3.0+, ArviZ 1.0+, and Bambi 0.19+.
We added support for Python 3.14.
Consistent with PyMC 6.0+ and ArviZ 1,0+ expectations, the model.sample() by default uses numba as the compute backend.
model.sample() now returns an xarray.DataTree object instead of the arviz.InferenceData object. Other functions that expect arviz.InferenceData objects have been updated to accept xarray.DataTree objects.
model.summary() and model.plot_trace() methods are now removed. Use az.summary() and az.plot_trace_dist() instead.
HSSM can now be installed directly from PyPI via pip or uv. Conda support is no longer provided.

Dependency changes:¶

bambi floor raised to >=0.19.0 (per-parameter non-centered support); ssm-simulators floor raised to >=0.13.1 (aDDM engine + fixation continuation).

Bug fixes:¶

Restore JAX-NUTS jitter control: HSSM again disables the built-in initial-value jitter of the numpyro/blackjax samplers (PyMC 6 removed the public switch that made this possible), so sampling starts from HSSM's own controlled initval_jitter instead of an extra uniform jitter (#999).
Python 3.14: posterior/prior predictive sampling fixed. On 3.14, the dynamically created SSM random-variable class carried PEP 649 annotation metadata that numba's vendored cloudpickle could not serialize (TypeError: cannot pickle '_abc._abc_data' object), breaking sample_posterior_predictive and related plotting under the numba backend. The class attributes are now un-annotated plain assignments, and the previous 3.14 xfail markers on the predictive/plotting tests have been removed.
model.vi(..., backend="jax") now works (#1056). PyMC 6 added a backend option to pm.fit(), which HSSM.vi() forwards. Requesting backend="jax" failed with NotImplementedError: No JAX conversion for the given Op: LANLogpVJPOp, because the operator that computes LAN likelihood gradients had no JAX translation. It now has one, and its gradients are tested to match the existing execution path exactly. Two remaining PyMC bugs (which break pm.fit(backend="jax") for any model, HSSM or not; tracked in #1056) are worked around by scoped, self-disabling compatibility shims in hssm._vi_compat, applied only when backend="jax" is requested — they detect a fixed PyMC and become no-ops, and will be removed once the pinned PyMC includes the upstream fixes. Known limitation: models with a single free parameter dimension may still fail to compile; use backend="c" there.
Fixed incorrect gradients for choice-only models with all-scalar parameters and outlier modeling. For likelihoods that take parameters only (CPN/OPN networks) with no regression parameters, the gradient computation ignored any transformation applied downstream of the likelihood — for example the p_outlier mixture. Gradient-based fitting (VI, or NUTS on the default backend) of such models silently used incorrect gradients. The gradient is now exact, and a regression test asserts the chain rule holds on every backend.
Internal cleanup of the LAN likelihood operators. The Op classes are now defined once at module level instead of once per model, so building many models in one long-running process (e.g. a parameter-recovery loop in a notebook) no longer grows PyTensor's dispatch registry — which previously kept every model's network weights in memory for the lifetime of the process. The gradient hook was also migrated from PyTensor's deprecated grad to the new pullback API, silencing a FutureWarning that appeared on every model fit.
Documentation build fix for the marimo aDDM tutorial. The docs build now includes the marimo dependency needed to parse the aDDM marimo tutorial, and skips CI execution for its button-gated sampling cells.

0.3.1¶

This version includes the following changes:

sbi NRE ONNX integration: train an sbi Neural Ratio Estimator (NRE), export it to ONNX via lanfactory.onnx.transform_sbi_to_onnx, and load it into HSSM's loglik_kind="approx_differentiable" pipeline for MCMC inference (keystone tutorial sbi_nre_integration.ipynb).
ONNX loader hardening: onnx2jax now rejects ONNX graphs with dynamic/symbolic input dimensions (enforcing the single-trial + jax.vmap contract), and raises a clear error when JAX x64 is disabled and a graph carries int64 constants that would be truncated (guarding against silent likelihood corruption from flow-export sentinels). Applies across all ONNX exporters (LANfactory, BayesFlow, sbi).

0.3.0¶

This version includes the following changes:

Support for choice-only models: the HSSM class, data validator, distributions, and model configs now handle models without reaction times, including a softmax likelihood family.
Racing Diffusion Model (RDM3): analytical likelihood, model configuration, and tests for a 3-choice RDM, with safe negative-RT handling and JAX backend support.
Poisson Race model: initial implementation of the Poisson race model.
Hidden Markov Model (HMM-SSM) example notebook.
BayesFlow LRE integration through HSSM for likelihood ratio estimation.
RLSSM config system: new RLSSMConfig and BaseModelConfig classes with comprehensive validation, plus a generalized RL likelihood builder supporting multiple computed parameters.
New tutorials: choice-only modeling, first HMM-SSM example, and updated existing tutorials.
DataValidator refactored to DataValidatorMixin for improved extensibility.
Bug fixes: default prior assignment, dimensionality errors with Bambi 0.17.0, negative-RT checks on missing data, flaky tests.
Infrastructure: Python 3.13 support, restructured CI test workflows with coverage reporting, updated model.sample() API to match Bambi conventions.

0.2.12¶

This version includes the following changes:

Compatibility with Python 3.13 and Bambi 0.17.1+.
A new make_distribution_for_supported_model convenience function for creating pm.Distribution.
Bug fix for default prior assignment.
Initial private implementations of RLSSM features.
Added disaggregation of quantile computation by grouping of choice.