Api generalisation - #43
Merged
Merged
Conversation
Replace the LikelihoodModel zoo with additive covariance Terms over each constraint's stacked observation vector: cross-block terms couple observations (case A) and Parameter-identity sharing wires one sampled value into several terms (case B). Likelihoods become thin functionals (Gaussian, Student-t, chi-squared) of the precomputed Mahalanobis statistics, with the full-tuple parameter convention validated everywhere. Also: constraint-scoped parameter validation (cross-constraint sharing and duplicate names are hard errors), fail-fast singular-covariance checks naming the offending dataset, frozen block classification with cached dense and per-block Cholesky factors, measurement systematics retained as inert Observation metadata with an opt-in systematic_terms factory (norm-divided units), latent-scale models (ScaledModel, PerObservationScaledModel), predictive utilities including GP discrepancy propagation, and comprehensive unit + regression coverage (178 tests), including real-solver smoke tests and a fix for the never-working dXS/dA-from-dXS/dRuth unit conversion. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Port all notebooks to the Term-based covariance API and extend coverage: new measurement_to_calibration (EXFOR-shaped measurement to calibrated potential, unit contract, guardrails) and robust_likelihoods (Student-t vs Gaussian) notebooks; error-model catalog completed with noise_term and offset_term options; likelihood_scaling/weights equivalence; shared systematics across real cross-section datasets (case A); GP model discrepancy on a differential cross section. Notebooks use the stacked_supports helper and no longer write PDF artifacts on execution. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Rewrite the API reference around the covariance/predictive modules, replace the design spec with an accurate architecture page (docs/design.md), refresh README to the Term API with the behavior change called out, drop the broken environment.yml install path, run notebooks in CI with pytest-xdist under a raised timeout, test 3.11, scope bare pytest to the unit suite, and ignore local tooling artifacts. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
x4i3 downloads its database at import time when missing from site-packages, and each nbmake notebook runs in its own kernel process: under pytest-xdist, concurrent kernels raced the same download/unpack and failed mid-extraction. Warm the database with a single serial import before the parallel run, and cache the data directory across runs (one key shared by all jobs, keyed on the dependency pins) so tests, docs, and the wheel smoke test stop re-downloading it every run. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
jitr 3.0 changed DifferentialWorkspace.xs and the (p,n) Workspace.xs to take potential arrays evaluated on ws.radial_grid() instead of callables plus argument tuples. Move the potential evaluation into a private _xs helper on each reaction model so evaluate and visualizable_model_prediction share it, accept an optional separate Coulomb interaction (calculate_interaction_from_params may return two or three argument tuples), and fail loudly on an unknown elastic quantity instead of leaving the extractor unset. Require jitr>=3.0 from PyPI, and raise the Python floor to 3.12 (the only version jitr 3.0 is built for), trimming the CI matrix to match. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Introduce a single low-level transform type shared by the observation (comparison space), model (parametric latent scale) and covariance (term coordinates) layers that follow. A Transform is a numpy-style callable fn(a, *values) with an optional tuple of Parameters, an optional analytic derivative (central finite difference otherwise), an optional inverse, and composition via `|` with parameters concatenated. Plain callables are accepted anywhere a Transform is and wrapped as parameter-free. Ship the parameter-free `identity`, `log` (-inf where the argument is not positive, no warnings) and `exp`, plus the parametric `scale()` and `per_observation_scaling()` latent normalisations that will replace the ScaledModel classes. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Both reaction observations' from_measurement classmethods copied the same eight fields out of an exfor_tools Distribution and re-listed every constructor option by hand, so each new constructor keyword had to be threaded through twice. Collect the Distribution fields in one measurement_kwargs helper and forward the remaining keywords with **kwargs, so from_measurement accepts everything the constructor does (solver settings, compound_correction, and the observation-level options added in the following commits). Tests get a make_measurement stub factory in place of four hand-built SimpleNamespaces. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Replace the DenseTerm / DiagonalTerm / RankOneTerm / KernelTerm class
hierarchy with a single Term: a numpy-style callable fn(c, *values) of a
TermContext (the term's local view of the stacked x, y and ym on its
support) plus a `kind` saying how the result enters the covariance
("diag" adds v**2 to the diagonal, "mode" adds the outer product v v^T,
"matrix" adds a symmetric block). A plain array instead of fn is a fixed
contribution; `constant=True` lets an x-dependent callable be evaluated
once and cached, and the Constraint's eager singular-covariance check
now passes a real StackContext so such terms can read x.
Terms default to support=None, meaning "the whole constraint", and are
bound when added to a ConstraintCovariance, so the factory helpers take
support as a trailing keyword instead of a leading positional. An
optional `coords` transform (rxmc.transforms) is applied to x before fn
sees it, so a kernel can live in momentum transfer without knowing.
The factory helpers become one-liners over Term: statistical_term,
offset_term, normalization_term, noise_term (now with an optional
parametric basis), noise_fraction_term, model_error_term, the new
systematic_term (a mode with a user basis) and kernel_term (replacing
KernelTerm and discrepancy_term, with an optional parametric amplitude
so that Sigma += a a^T o K). Bases are ordinary callables of the
TermContext: ones, ym, averaging, x_basis(scale), exp_growth(scale),
constant_amplitude and exp_growth_amplitude(scale).
Port Observation.statistical_term / systematic_terms, the predictive
docstrings, the unit tests and five example notebooks to the new
signatures.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Let an Observation compare in a transformed space without the model knowing: Observation(x, y, transform=rxmc.transforms.log) takes raw y, keeps it as y_raw, stores y = log(y_raw) and the delta-method statistical error |t'(y_raw)| sigma, and fails loudly at construction if the transform is not finite at a data point. The Constraint applies the same transform to every prediction when stacking, so the model is written once in physical space and can never be double-transformed; predict(raw=True) returns the untransformed prediction, and a non-finite prediction (a non-positive cross section under log) yields -inf log likelihood and +inf chi-squared instead of a LinAlgError. Reported systematics are propagated to the comparison space the same way: the offset mode is |t'(y_raw)| omega and the normalisation mode becomes a systematic_term whose basis is eta ym_raw |t'(ym_raw)|, evaluated at the prediction. obs.log_jacobian and constraint.log_jacobian expose the constant needed to compare evidences across comparison spaces. Parametric transforms are rejected here; they belong on the model. The reaction observations forward transform= to the base class. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Make hold-out part of a constraint's support instead of a data surgery. Observation(mask=) marks which points are active; masked() and masked_where(predicate) derive views that share the data and any precomputed solver workspaces and only change which points enter the likelihood. Constraint(mask=) additionally switches whole observations off, and Constraint.masked() / complement() build the fit / held-out counterpart over the same Term and Parameter objects, so both views have an identical parameter vector and their log likelihoods add up to the unmasked one. Terms are still authored over the full stack: ConstraintCovariance gains active= and always assembles the full N x N matrix, restricting the Cholesky factorisation and the Mahalanobis distance to the active rows (fully masked blocks are skipped on the block path). The non-finite-transform guard, log_jacobian and the coverage counts are scoped to active points, and masked views keep an `identity` pointing at their root so identity-routed transforms treat them as one dataset. predict_and_covariance, x, y and covariance_matrix(active_only=) give the active-point views that model comparison needs. The reaction observations forward mask= to the base class. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Drop ScaledModel and PerObservationScaledModel in favour of one mechanism: PhysicalModel(params, transform=...) appends the transform's Parameters to the model's and applies it to the output of evaluate, so evaluate stays physical-space only and split_params / apply_transform give subclasses the same hook for their visualisation predictions. rxmc.transforms.scale() is the Kennedy-O'Hagan latent scale rho (it changes the mean, so it is a model transform, not a covariance term) and per_observation_scaling(observations) is one rho_i per dataset, routed by observation identity so masked views scale like their root. The reaction models accept transform= and route their visualisation path through split_params / apply_transform; the normalization inference notebook is ported to Polynomial(transform=...). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Expose the entrance-channel wavenumber of an ElasticDifferentialXSObservation and a momentum_transfer helper q = 2 k sin(theta/2) on its data angles, so a kernel_term can be placed in momentum-transfer space (coords=) for the error-model study forms instead of every caller reaching into the jitr workspace. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ence Provide the model-comparison bookkeeping the error-model study needs without touching a sampler: everything consumes a Constraint plus rows of posterior samples. predictive_draws draws from N(ym(theta), Sigma(theta)) on the active points (or the model-only predictive), coverage_curve / coverage_error / sharpness score calibration and width of those draws, heldout_log_predictive and elpd score a held-out constraint (typically fit.complement()), logz_summary and compare_logz summarise replicate nested-sampling evidences with a conservative tie verdict, log_jacobian supplies the comparison-space constant, and split_samples cuts flat sampler rows into model and per-constraint parameter blocks. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…rm API Rewrite the covariance section of the README and docs/design.md around the single generic Term (kind table, support=None, coords, the factory helpers and bases), add sections on comparison-space and model transforms, point masks and held-out complements, and the sampler-free model-comparison module, and record the error-model recipes the API was generalised for. The API reference gains the transforms and model_comparison modules and the new covariance helpers, and drops the removed term classes, ScaledModel classes and discrepancy_term. The README notes the jitr>=3.0 / Python 3.12 requirement. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
docs/generated/ is pure build output: docs/conf.py sets autosummary_generate = True, so Sphinx rewrites every stub on each build and `make clean` deletes the directory. Keeping them under version control meant every API change churned dozens of generated files (this refactor alone added 24, deleted 7 and modified 12). Ignore the directory and drop the tracked copies. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Constraint.__init__ copied the observations into a list and then kept iterating the original argument, so a generator (or any one-shot iterable) was exhausted on the first line and produced an empty constraint without complaint. Iterate the stored list throughout. Also from the review of the masks commit: empirical_coverage returns nan instead of dividing by zero when no point is active; an integer observation mask of length n holding only 0/1 is rejected as ambiguous (bool mask or index list?) instead of being read as indices; masked() re-runs the non-finite comparison-space guard so a view cannot re-activate a point that is -inf under log (and the guard no longer warns about inf*0 at inactive points); ConstraintCovariance keeps an active index array that merely has size N unless it is exactly arange(N); covariance_matrix documents active_only and its two possible shapes; masked() explains why sharing Terms between the view and its parent is safe; stray blank lines left by hoisted test imports are removed. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The generic-Term rewrite lost the eager shape check the old offset_term / normalization_term had: a magnitude or mask array of the wrong length was run through np.broadcast_to in the basis, which silently spreads a length-1 array over the support. Route every magnitude and basis value through _full(), which broadcasts scalars (0-d arrays included) and rejects any array whose shape is not (n,). A Term binds its support once and caches x-dependent values, so it cannot be shared between constraints of different length without returning stale or misplaced blocks; bind() now raises when asked to re-bind to a different N, and the Term docstring states the one-Term-one-constraint rule (masked views of one constraint share its stack and may share Terms). local_context() checks the parameter count before handing values to a parametric coords transform, so the error comes from the covariance layer rather than the Transform. Drop the discrepancy_term alias: nothing used it, the module docstring and API reference already omitted it, and kernel_term is the one name. Document StackContext in the API reference (it is the context a ConstraintCovariance is evaluated on), extract the kernel hyperparameter-to-Parameter loop into _kernel_params, and state that a kernel_term amplitude sees the transformed coordinate. Tests: length-1 magnitude/mask, 0-d magnitude, re-binding, local_context parameter check; the constant-term test moves out of the kernel class and the noqa'd lambda becomes a def. Not changed: the _AVERAGING alias stays, because removing it means renaming model_error_term's averaging= keyword for no behavioural gain. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
predictive_draws and heldout_log_predictive validated model_samples against its own length, so the row check could never fire and a single 1-D posterior row [a0, a1] was reshaped to (2, 1) - two samples of a one-parameter model - before failing deep in Constraint._split. _rows now treats a 1-D input as one sample (as split_samples already did) and checks the row count only against an explicit n. _psd_factor added an absolute 1e-10 jitter to Sigma on every call, including the successful Cholesky path; in a log comparison space with 1e-8 variances that inflated every draw by ~1 %. Factor Sigma as-is first, retry with a jitter relative to the mean variance only when the plain Cholesky fails, and fall back to the eigen square root last. The model-only predictive no longer assembles the covariance at all: it stacks Constraint.predict and takes the active points. Rename elpd to log_posterior_predictive: it is the log-mean-exp of the joint held-out log likelihood over posterior samples, i.e. the joint log posterior predictive of the block, not the pointwise-summed elpd of Vehtari et al., and the old name invited comparison with ArviZ/LOO numbers that are not comparable. sharpness(levels=) becomes sharpness(percentiles=) so the two interval conventions in the module (probabilities in (0, 1) for coverage, percentiles in 0-100 here) no longer share a name. The default coverage levels are one module constant; scipy.linalg is dropped for numpy.linalg; docstrings gain the missing Returns and parameter descriptions; test imports are hoisted. Not changed: compare_logz still ignores the replicate count (it is informational; now said so in the docstring), and the log_jacobian wrapper stays because the design doc and API reference point to it. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Polynomial.evaluate reported len(self.params) in its parameter-count error, which counts the transform's parameters too; report order + 1. A contextual per_observation_scaling evaluated without a context now says so instead of raising a misleading "observation was not registered" KeyError, and it keeps its registered observations under a private attribute (the test checks the id-routing behaviour instead of reaching in). Transform.is_identity and inverse document that they are an object-identity test and parameter-free only, the composition closures share one argument unpacker, and _safe_log says it expects an ndarray. Observation.systematic_terms checks eagerly that the transform has an inverse when a normalisation systematic is reported (the failure used to surface at the first likelihood evaluation), explains why the offset and normalisation modes are linearised at y_raw and ym_raw respectively, and says what support=None means inside a multi-observation constraint. The elastic observation keeps the kinematics it already unpacks and reads k from them; the momentum_transfer docstring no longer suggests passing its array as a kernel coordinate transform, and the momentum-transfer study form checks the helper against the inline map. The base evaluate docstrings say they receive base parameters only, the ias potential docstrings are wrapped to the line length, and _xs states the two-or-three-tuple contract. The design doc notes that the latent scale parameters now come last and are named log_rho, unlike ScaledModel. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Add the module-level identity, log and exp transforms to the API reference (they are used throughout the README, design doc and tests but had no entry), and document StackContext's fields as constructor parameters so its autosummary page no longer duplicates the attribute descriptions under sphinx -W. Bring the example notebooks back in line with the code: the stored output in measurement_to_calibration that still printed the old RankOneTerm class names now shows Term, the gp_discrepancy prose refers to kernel_term rather than the dropped discrepancy_term alias, and linear_calibration_demo is re-executed so its help() output shows the current statistical_term signature instead of DenseTerm. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
docs/groundup_design.md sketches a declare-then-compile architecture (hashable identity Parameters, Dataset/Block split, bind-time predictors, stateless Terms with block-reference supports, a single ParameterIndex, and a structured diag + modes + block covariance with a Woodbury path) and compares it with the current api_generalisation design, ending with an incremental adoption path. docs/bugs_found.md enumerates the bugs and driver inconsistencies found while reading the branch for that comparison, with file:line references. Both pages are added to the Sphinx toctree.
Confirmed bugs - IsobaricAnalogPNObservation forwards wavelengths_beyond_range and zeros_per_node to set_up_solver instead of silently dropping them. - BatchedAdaptiveMetropolisSampler: docstrings now match the code (the proposal adapts after every batch, burn-in included) and the public proposal attribute is refreshed alongside args. - Parameter is hashable (value hash consistent with its value __eq__), coerces bounds to a float 2-tuple, and has a repr. - One pint UnitRegistry, DEFAULT_LMAX, XS_UNIT, RUTHERFORD_UNIT and MB_PER_B live in observation_from_measurement (now exported from rxmc); both reaction observations import them and both reaction models divide by MB_PER_B instead of a bare 1000. Driver parity between CalibrationConfig and Walker - Walker.log_posterior and the Gibbs conditional evaluate the prior first and return -inf without touching the likelihood. - Walker gains likelihood_scaling with the config's semantics. - A list of scipy marginals is wrapped in IndependentPrior by a shared priors.as_prior in both ParameterConfig and Sampler; the four list branches in ParameterConfig are gone. x0 for a list prior is now seeded. - ParameterConfig._infer_dim calls mean() when it is a method. Fragile spots - Reaction models reject an observation of the wrong class up front. - Walker.walk runs one _run_batch sweep for burn-in and active batches. - priors.clip_unit_cube is applied in every prior_transform. Tests added for each item (295 total, from 268); docs/bugs_found.md records the resolution of every item, with 11 deferred to groundup_design.md.
…s page docs/groundup_design.md now guides a from-scratch rewrite on a branch of this repository rather than an incremental migration: immutable declarations compiled by one Problem, structured (Woodbury) covariance, external samplers only, a capability map covering every notebook and test behaviour, a harvest table, the gaps to fill, milestones that slot the recipe tests in as capabilities arrive, and the 0.x to 1.0 release path. docs/recipes.md is the companion: 42 user stories with the new spelling, expected behaviour and citations, each of which becomes a test in the rewrite, plus the list of what the API deliberately does not express.
Closed
There was a problem hiding this comment.
🔵 Needs a closer look
It introduces a broad, cross-cutting API refactor across core modeling, sampling, reaction integration, docs, and CI, so it warrants final human review despite limited localized issues found.
Pull request overview
This PR generalizes and simplifies the rxmc API by consolidating uncertainty modeling around composable covariance Terms, introducing a shared low-level Transform type, and aligning the built-in Walker with CalibrationConfig semantics (tempering, prior short-circuiting, consistent prior handling). It also updates the reaction-model integration for jitr>=3.0, adds new predictive/model-comparison utilities, and refreshes docs/CI to match the new architecture.
Changes:
- Replace/retire the prior “likelihood model zoo” patterns in favor of stacked-covariance composition (
rxmc.covariance.Term) and shared transforms (rxmc.transforms.Transform). - Align sampling drivers and prior handling (list-of-marginals ->
IndependentPrior, unit-cube clipping, likelihood tempering inWalker). - Add new utilities and regression coverage (predictive bands, model comparison utilities, covariance-refactor regression pins), plus CI/docs updates.
File summaries
| File | Description |
|---|---|
| test/test_transforms.py | New unit tests for rxmc.transforms behavior (composition, derivatives, contextual scaling). |
| test/test_sampler.py | Update sampler/walker tests for constraint-scoped params, tempering, prior wrapping, and burn messaging. |
| test/test_regression.py | Regression pins for the covariance-refactor behavior change and block-diagonal equivalence. |
| test/test_reaction_models.py | Smoke tests for jitr-backed reaction models and observation type checks. |
| test/test_priors.py | Add coverage for unit-cube clipping and as_prior wrapping behavior. |
| test/test_predictive.py | New tests for GP posterior predictive and predictive-band utilities. |
| test/test_params.py | New tests for value-based Parameter equality/hashing and bounds coercion. |
| test/test_model_comparison.py | New tests for sampler-agnostic predictive/model-comparison utilities. |
| test/test_evidence.py | Update Evidence tests for auto-detected parametric constraints and parameter validation. |
| test/test_config.py | Update config tests for new covariance-term patterns, list-prior wrapping, and tempering parity. |
| test/helpers.py | Shared test helpers for MVN reference computations / stack-context building. |
| test/conftest.py | Ensure helpers are importable under different pytest import modes. |
| src/rxmc/walker.py | Add likelihood tempering, prior-first short-circuiting, and refactor batch execution/messaging. |
| src/rxmc/transforms.py | Introduce Transform abstraction (composition, derivative/inverse, contextual transforms). |
| src/rxmc/priors.py | Add clip_unit_cube and as_prior; make prior transforms boundary-safe. |
| src/rxmc/predictive.py | New predictive-uncertainty helpers (GP conditioning + total predictive bands). |
| src/rxmc/physical_model.py | Add optional model-side transforms with parameter plumbing and splitting helpers. |
| src/rxmc/params.py | Make Parameter hashable/value-equal; coerce bounds; improve repr. |
| src/rxmc/param_sampling.py | Normalize list-priors via as_prior; clarify/adapt burn-in adaptation semantics. |
| src/rxmc/observation_from_measurement.py | Centralize shared unit registry/constants + measurement kwarg helpers and angle-grid checks. |
| src/rxmc/model_comparison.py | New sampler-agnostic posterior predictive and model comparison utilities. |
| src/rxmc/ias_pn_observation.py | Convert to an Observation subclass; normalize measurement/unit handling via shared helpers. |
| src/rxmc/ias_pn_model.py | Update to jitr>=3.0 array-potential API; add observation type checks; support model transforms. |
| src/rxmc/evidence.py | Simplify Evidence API to one constraints list; auto-detect parametric constraints; validate constraint-scoped params. |
| src/rxmc/elastic_diffxs_observation.py | Convert to an Observation subclass; unify units/normalization; add momentum-transfer helper. |
| src/rxmc/elastic_diffxs_model.py | Update to jitr>=3.0 array-potential API; add observation type checks; optional Coulomb term; model transforms. |
| src/rxmc/correlated_discrepancy_likelihood_model.py | Remove old GP discrepancy likelihood model (superseded by covariance kernel_term + predictive tools). |
| src/rxmc/config.py | Wrap list priors via as_prior, clip unit-cube consistently, fix dimension inference, add parametric prediction alignment helper. |
| src/rxmc/init.py | Export new/renamed modules (covariance, predictive, transforms, model_comparison, measurement helpers). |
| requirements.txt | Bump minimum jitr version to match new solver interface. |
| README.md | Update package narrative and quickstart to covariance-term composition; document behavior change and jitr>=3.0. |
| pyproject.toml | Raise minimum Python to 3.12; adjust ruff/isort first-party settings; narrow default pytest scope. |
| examples/sampling_algos.ipynb | Update example to new constraint/term-based noise modeling and constraint-scoped params. |
| examples/calibration_config_emcee_dynesty.ipynb | Update examples for new likelihood API and output formatting. |
| examples/30s_optical_potential_calibration.ipynb | Update examples for new likelihood API. |
| environment.yml | Remove conda environment file (installation guidance updated elsewhere). |
| docs/installation.rst | Remove conda/mamba guidance; keep venv workflow. |
| docs/index.rst | Update docs landing to covariance-term model; expand to include new design docs pages. |
| docs/generated/rxmc.walker.Walker.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.proposal.ProposalDistribution.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.proposal.NormalProposalDistribution.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.proposal.LogspaceNormalProposalDistribution.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.proposal.HalfNormalProposalDistribution.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.priors.TruncatedNormalPrior.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.priors.IndependentPrior.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.physical_model.Polynomial.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.physical_model.PhysicalModel.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.params.Parameter.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.param_sampling.Sampler.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.param_sampling.MetropolisHastingsSampler.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.param_sampling.BatchedAdaptiveMetropolisSampler.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.param_sampling.AdaptiveMetropolisSampler.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.observation.Observation.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.observation.FixedCovarianceObservation.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.metropolis_hastings.metropolis_hastings.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.likelihood_model.UnknownNormalizationModel.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.likelihood_model.UnknownNormalizationErrorModel.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.likelihood_model.UnknownNoiseFractionErrorModel.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.likelihood_model.UnknownNoiseErrorModel.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.likelihood_model.UnknownModelError.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.likelihood_model.StudentTLikelihoodModel.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.likelihood_model.ParametricLikelihoodModel.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.likelihood_model.LikelihoodModel.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.likelihood_model.FixedCovarianceLikelihood.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.likelihood_model.Chi2LikelihoodModel.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.ias_pn_observation.IsobaricAnalogPNObservation.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.ias_pn_model.IsobaricAnalogPNXSModel.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.evidence.Evidence.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.elastic_diffxs_observation.ElasticDifferentialXSObservation.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.elastic_diffxs_model.ElasticDifferentialXSModel.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.correlated_discrepancy_likelihood_model.SklearnKernelGPDiscrepancyModel.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.constraint.Constraint.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.config.ParameterConfig.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.config.CalibrationConfig.rst | Remove generated API stub file from repo. |
| docs/generated/rxmc.adaptive_metropolis.adaptive_metropolis.rst | Remove generated API stub file from repo. |
| docs/examples.rst | Expand examples index for new notebooks/advanced topics. |
| docs/design.md | New design document describing the stacked-covariance architecture. |
| docs/conf.py | Exclude jupyter_execute from docs build artifacts. |
| docs/bugs_found.md | New “bugs found” write-up documenting issues found/fixed during review. |
| docs/api.rst | Update API reference to new transforms/covariance/predictive/model-comparison layout. |
| .gitignore | Ignore generated docs outputs and other artifacts (docs/generated, uv.lock, etc.). |
| .github/workflows/ci.yml | CI updates: Python 3.12-only, EXFOR DB caching/warmup, parallel notebook execution, docs build prep. |
Review details
Suppressed comments (1)
src/rxmc/observation_from_measurement.py:82
check_angle_gridallows an angle exactly equal topi(it only rejects> np.pi), but the error message says the allowed range is[0,pi). Since callers passangles_vis=np.linspace(..., 180, ...)(i.e.,pi), this message is misleading; either the check should be half-open or the message should be inclusive.
- Files reviewed: 54/101 changed files
- Comments generated: 0
- Review effort level: Lite
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
Remove the BUQEYE truncation covariance, model averaging and mixing, the realised-discrepancy posterior predictive check and power-scaling sensitivity recipes for brevity; renumber 30-42 to 28-38 and fix every cross-reference in the design document's capability map, milestones and notebook mapping.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Supersedes #42, simplifies API and implementation