emu-hmf 1.0.0__tar.gz

emu_hmf-1.0.0/CHANGELOG.md

@@ -0,0 +1,18 @@
+# Changelog
+
+## 1.0.0 — 2026-08-29
+
+First public release.
+
+* Recalibrated Tinker08 multiplicity function at two halo definitions,
+  `RockstarM200m` and `RockstarMvir`, each fitted on 2000 CSST cosmologies over
+  twelve redshifts and 456 526 rows inside ν ∈ [0.5, 3].
+  Residual in ln f, on cosmologies held out entirely: 7.00 % → 0.52 % at 200m,
+  10.92 % → 0.54 % at virial.
+* Pure-JAX inference path: `jax.grad`, `jax.jit` and `jax.vmap` all pass through
+  the cosmology. Two runtime dependencies, numpy and JAX.
+* `target.nu_covered(z)` records the peak heights the training set actually
+  spans, which narrows with redshift as growth pushes σ down.
+* Both weight files are float64 and carry their own provenance — mass
+  definition, sample size, held-out split and residual.
+* Released under the BSD 3-Clause licence.

emu_hmf-1.0.0/CITATION.cff

@@ -0,0 +1,72 @@
+cff-version: 1.2.0
+title: "emu_hmf: a differentiable, cosmology-dependent recalibration of the Tinker08 halo multiplicity function"
+message: "If you use this software, please cite it together with the works it is built on (references below)."
+type: software
+authors:
+  - family-names: Comparat
+    given-names: Johan
+    email: jcomparat@lpsc.in2p3.fr
+    orcid: "https://orcid.org/0000-0001-9200-1497"
+repository-code: "https://github.com/JohanComparat/emu_hmf"
+url: "https://emu-hmf.readthedocs.io"
+license: BSD-3-Clause
+version: 1.0.0
+date-released: "2026-08-31"
+keywords:
+  - cosmology
+  - halo mass function
+  - large-scale structure
+  - emulator
+  - automatic differentiation
+references:
+  - type: article
+    title: "Toward a Halo Mass Function for Precision Cosmology: The Limits of Universality"
+    authors:
+      - family-names: Tinker
+        given-names: Jeremy
+      - family-names: Kravtsov
+        given-names: "Andrey V."
+      - family-names: Klypin
+        given-names: Anatoly
+      - family-names: Abazajian
+        given-names: Kevork
+      - family-names: Warren
+        given-names: Michael
+      - family-names: Yepes
+        given-names: Gustavo
+      - family-names: Gottlöber
+        given-names: Stefan
+      - family-names: Holz
+        given-names: "Daniel E."
+    journal: "The Astrophysical Journal"
+    volume: 688
+    start: 709
+    year: 2008
+    doi: 10.1086/591439
+  - type: article
+    title: >-
+      CSST cosmological emulator II: Generalized accurate halo mass function
+      emulation
+    authors:
+      - family-names: Chen
+        given-names: Zhao
+      - family-names: Yu
+        given-names: Yu
+    journal: "Science China Physics, Mechanics, and Astronomy"
+    volume: 68
+    issue: 10
+    start: 109513
+    year: 2025
+    doi: 10.1007/s11433-025-2764-x
+    url: "https://ui.adsabs.harvard.edu/abs/2025SCPMA..6809513C/abstract"
+    notes: >-
+      The halo mass function emulation this package recalibrates against.
+  - type: software
+    title: "CSSTemu: emulators for the CSST simulation suite"
+    authors:
+      - family-names: Chen
+        given-names: Zhao
+      - family-names: Yu
+        given-names: Yu
+    year: 2025
+    repository-code: "https://github.com/czymh/csstemu"

emu_hmf-1.0.0/LICENSE

@@ -0,0 +1,29 @@
+BSD 3-Clause License
+
+Copyright (c) 2026, Johan Comparat
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+
+3. Neither the name of the copyright holder nor the names of its contributors
+   may be used to endorse or promote products derived from this software
+   without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

emu_hmf-1.0.0/MANIFEST.in

@@ -0,0 +1,7 @@
+include LICENSE README.md CHANGELOG.md CITATION.cff
+include environment-gen.yml
+recursive-include emu_hmf/data *.npz
+recursive-include tests *.py
+recursive-include docs *.rst *.py *.txt *.png *.npz Makefile
+prune docs/_build
+global-exclude __pycache__ *.py[cod]

emu_hmf-1.0.0/PKG-INFO

@@ -0,0 +1,201 @@
+Metadata-Version: 2.4
+Name: emu_hmf
+Version: 1.0.0
+Summary: A differentiable, cosmology-dependent recalibration of the Tinker08 halo multiplicity function
+Author-email: Johan Comparat <jcomparat@lpsc.in2p3.fr>
+License-Expression: BSD-3-Clause
+Project-URL: Homepage, https://github.com/JohanComparat/emu_hmf
+Project-URL: Documentation, https://emu-hmf.readthedocs.io
+Project-URL: Source, https://github.com/JohanComparat/emu_hmf
+Project-URL: Issues, https://github.com/JohanComparat/emu_hmf/issues
+Keywords: cosmology,halo mass function,large-scale structure,emulator,jax,automatic differentiation,Tinker08
+Classifier: Development Status :: 5 - Production/Stable
+Classifier: Intended Audience :: Science/Research
+Classifier: Topic :: Scientific/Engineering :: Astronomy
+Classifier: Topic :: Scientific/Engineering :: Physics
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Operating System :: OS Independent
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: numpy>=1.24
+Requires-Dist: jax>=0.4.20
+Provides-Extra: train
+Requires-Dist: optax>=0.1.7; extra == "train"
+Provides-Extra: dev
+Requires-Dist: pytest>=7; extra == "dev"
+Requires-Dist: pytest-cov; extra == "dev"
+Requires-Dist: optax>=0.1.7; extra == "dev"
+Requires-Dist: build; extra == "dev"
+Requires-Dist: tomli>=2.0; python_version < "3.11" and extra == "dev"
+Provides-Extra: docs
+Requires-Dist: sphinx; extra == "docs"
+Requires-Dist: sphinx-rtd-theme; extra == "docs"
+Requires-Dist: myst-parser; extra == "docs"
+Requires-Dist: matplotlib; extra == "docs"
+Dynamic: license-file
+
+# emu_hmf
+
+[![tests](https://github.com/JohanComparat/emu_hmf/actions/workflows/ci.yml/badge.svg)](https://github.com/JohanComparat/emu_hmf/actions/workflows/ci.yml)
+[![docs](https://readthedocs.org/projects/emu-hmf/badge/?version=latest)](https://emu-hmf.readthedocs.io/en/latest/)
+[![PyPI](https://img.shields.io/pypi/v/emu_hmf.svg)](https://pypi.org/project/emu_hmf/)
+[![Python](https://img.shields.io/pypi/pyversions/emu_hmf.svg)](https://pypi.org/project/emu_hmf/)
+[![License](https://img.shields.io/badge/license-BSD--3--Clause-blue.svg)](LICENSE)
+
+A differentiable, cosmology-dependent recalibration of the Tinker et al. (2008)
+halo multiplicity function, trained against the CSST emulator
+([Chen & Yu 2025](https://github.com/czymh/csstemu)) over the box that
+emulator was built on.
+
+Tinker08 is a fit to simulations — and not to the simulations anyone compares
+against now. At a Planck cosmology it is offset by a few per cent at z = 0, and
+the size of that offset is itself a function of cosmology and redshift, which a
+fit whose only inputs are σ(M) and z cannot express. This package learns that
+offset.
+
+What it learns is *not* a mass function. It is a correction to Tinker08's four
+shape parameters (A, a, b, c) as a function of the eight CSST cosmological
+parameters and redshift:
+
+```
+f(σ) = A [ (σ/b)^-a + 1 ] exp(-c/σ²),   with   (A, a, b, c) → (A, a, b, c) · e^g(θ, z)
+```
+
+Keeping Tinker08 as the carrier is the whole design. At `g = 0` the answer *is*
+Tinker08, exactly, so the baseline is a point in the same parameterisation
+rather than a different code. The peak-height dependence stays where the physics
+put it, and the network only has to express what the simulations add. And the
+result is a fit with named parameters, so you can ask what the recalibration did
+to the amplitude as against the tilt.
+
+## Install
+
+```bash
+pip install emu_hmf
+```
+
+Two dependencies, numpy and JAX, and 90 kB of trained weights. No Boltzmann
+solver, no Gaussian-process emulator, no training stack, no conda environment —
+a forecast that wants to *evaluate* a mass function should not have to install
+the machinery that fitted one. `tests/test_public_api.py` asserts that split
+rather than trusting it.
+
+### A dedicated environment
+
+If you want one anyway — for a reproducible box rather than because the package
+needs it — [`environment.yml`](environment.yml) is the minimal one:
+
+```bash
+mamba env create -f environment.yml     # or: conda env create -f environment.yml
+mamba activate emu_hmf
+pip install -e .
+```
+
+It pins the **CPU** build of `jaxlib`: 64 MB against 199 MB for the CUDA one,
+and left unpinned the build depends on whether the machine that solved the
+environment happened to have a driver. The file says how to swap it for a GPU,
+and carries commented blocks for the `[dev]`, `[train]` and `[docs]` extras.
+
+## Use
+
+```python
+import numpy as np
+from emu_hmf.model import HmfCorrection
+
+corr = HmfCorrection()                       # 200m; HmfCorrection(WEIGHTS["vir"]) for virial
+
+theta = np.array([0.049, 0.31, 67.36, 0.9649, 2.1, -1.0, 0.0, 0.06])
+#                 Ω_b    Ω_cb  H0     n_s     10⁹A_s  w    w_a  Σm_ν
+
+f = corr.fsigma(sigma=0.8, theta=theta, z=0.5)          # the multiplicity function
+n = corr.dndlnM(m, sigma, dlnsigma_dlnm, rho_cold, theta, z=0.5)   # the abundance
+```
+
+σ(M) is passed in, not computed: this package has no power spectrum and should
+not acquire one, and the σ(M) the fit was made against is the *cold* field
+against ρ̄_cb. Fitting f(σ) against one variance and evaluating it with another
+is the mismatch that makes a multiplicity function look wrong when the
+convention around it is what moved.
+
+Everything is JAX, so `jax.grad`, `jax.jit` and `jax.vmap` all work through the
+cosmology. That is the reason this exists rather than a table of numbers.
+
+## Two mass definitions, two files
+
+The correction is not the same function at two halo boundaries, so there is no
+single correction with a Δ argument. Both are fitted against a *Rockstar*
+spherical-overdensity mass, so the comparison isolates the boundary rather than
+mixing in a change of halo finder.
+
+| weights | halo definition | Tinker08 unchanged | recalibrated | improvement |
+| --- | --- | --- | --- | --- |
+| `WEIGHTS["200m"]` | SO 200 × mean, Rockstar | 7.00 % | **0.52 %** | 13.1× |
+| `WEIGHTS["vir"]` | SO virial, Rockstar | 10.92 % | **0.54 %** | 19.1× |
+
+rms in ln f, on 200 cosmologies held out *entirely* from training — not held-out
+rows. Each design contributes several hundred rows and at fixed cosmology ln f
+is smooth in σ, so a random row split measures interpolation between neighbouring
+masses of a cosmology the network has already seen. This correction is only ever
+asked for a cosmology it has not seen.
+
+Both files carry `WEIGHTS["200m"]`'s Δ = 200m Tinker08 as the carrier, so the
+virial weights absorb the *boundary change* as well as the recalibration. They
+are not a per-cent correction: at z = 0 they sit some 13 % below the 200m
+carrier. Reading "correction" as "small" at `vir` is a misreading.
+
+## Where it is defined
+
+Outside either bound the package refuses rather than extrapolating.
+
+**The cosmology** must be inside CSST's box, which is copied into `box.py` and
+checked against the emulator's own `param_limits` by `tests/test_box.py`:
+
+| Ω_b | Ω_cb | H₀ | n_s | 10⁹A_s | w | w_a | Σm_ν |
+| --- | --- | --- | --- | --- | --- | --- | --- |
+| 0.04–0.06 | 0.24–0.40 | 60–80 | 0.92–1.00 | 1.7–2.5 | −1.3–−0.7 | −0.5–0.5 | 0–0.3 |
+
+Note Ω_cb: CSST bounds the **cold** density, with massive neutrinos excluded.
+
+**The peak height** must be inside ν = δ_c/σ ∈ [0.5, 3], and the mass inside
+10¹²–10¹⁴ M⊙/h. Those two cuts do not commute with redshift: growth pushes σ
+down, so a fixed mass is a higher peak later, and the low-ν half of the band is
+simply absent above z ≈ 0.25. `target.nu_covered(z)` records what the training
+set actually spans — ν ≥ 1.4 by z = 3 — because a caller who checked only the
+nominal range would be extrapolating with no warning.
+
+The correction is a few per cent at z = 0 and grows with redshift, reaching
+about 12 % rms by z = 3. Quoting the low-redshift figure alone would understate
+it several-fold over most of the range it is defined on.
+
+## Documentation
+
+[emu-hmf.readthedocs.io](https://emu-hmf.readthedocs.io) — concepts, a tutorial
+with figures, the validity domain, and how to reproduce the training set.
+
+## Reproducing
+
+The 2000-cosmology training set (11.7 MB, both mass definitions) is archived
+with a DOI; the fit that turns it into the shipped weights needs only `optax`:
+
+```bash
+pip install emu_hmf[train]
+python -m emu_hmf.fit --shards ./shards --out weights.npz
+```
+
+Regenerating the shards themselves needs CLASS and the CSST emulator; see the
+documentation's *Reproducing the training set* page.
+
+## Citation
+
+If you use this package, please cite Tinker et al. (2008) for the functional
+form, Chen & Yu (2025) for the CSST emulator this is calibrated against, and
+this package for the recalibration. See `CITATION.cff`.
+
+## Licence
+
+BSD 3-Clause. See [LICENSE](LICENSE).

emu_hmf-1.0.0/README.md

@@ -0,0 +1,160 @@
+# emu_hmf
+
+[![tests](https://github.com/JohanComparat/emu_hmf/actions/workflows/ci.yml/badge.svg)](https://github.com/JohanComparat/emu_hmf/actions/workflows/ci.yml)
+[![docs](https://readthedocs.org/projects/emu-hmf/badge/?version=latest)](https://emu-hmf.readthedocs.io/en/latest/)
+[![PyPI](https://img.shields.io/pypi/v/emu_hmf.svg)](https://pypi.org/project/emu_hmf/)
+[![Python](https://img.shields.io/pypi/pyversions/emu_hmf.svg)](https://pypi.org/project/emu_hmf/)
+[![License](https://img.shields.io/badge/license-BSD--3--Clause-blue.svg)](LICENSE)
+
+A differentiable, cosmology-dependent recalibration of the Tinker et al. (2008)
+halo multiplicity function, trained against the CSST emulator
+([Chen & Yu 2025](https://github.com/czymh/csstemu)) over the box that
+emulator was built on.
+
+Tinker08 is a fit to simulations — and not to the simulations anyone compares
+against now. At a Planck cosmology it is offset by a few per cent at z = 0, and
+the size of that offset is itself a function of cosmology and redshift, which a
+fit whose only inputs are σ(M) and z cannot express. This package learns that
+offset.
+
+What it learns is *not* a mass function. It is a correction to Tinker08's four
+shape parameters (A, a, b, c) as a function of the eight CSST cosmological
+parameters and redshift:
+
+```
+f(σ) = A [ (σ/b)^-a + 1 ] exp(-c/σ²),   with   (A, a, b, c) → (A, a, b, c) · e^g(θ, z)
+```
+
+Keeping Tinker08 as the carrier is the whole design. At `g = 0` the answer *is*
+Tinker08, exactly, so the baseline is a point in the same parameterisation
+rather than a different code. The peak-height dependence stays where the physics
+put it, and the network only has to express what the simulations add. And the
+result is a fit with named parameters, so you can ask what the recalibration did
+to the amplitude as against the tilt.
+
+## Install
+
+```bash
+pip install emu_hmf
+```
+
+Two dependencies, numpy and JAX, and 90 kB of trained weights. No Boltzmann
+solver, no Gaussian-process emulator, no training stack, no conda environment —
+a forecast that wants to *evaluate* a mass function should not have to install
+the machinery that fitted one. `tests/test_public_api.py` asserts that split
+rather than trusting it.
+
+### A dedicated environment
+
+If you want one anyway — for a reproducible box rather than because the package
+needs it — [`environment.yml`](environment.yml) is the minimal one:
+
+```bash
+mamba env create -f environment.yml     # or: conda env create -f environment.yml
+mamba activate emu_hmf
+pip install -e .
+```
+
+It pins the **CPU** build of `jaxlib`: 64 MB against 199 MB for the CUDA one,
+and left unpinned the build depends on whether the machine that solved the
+environment happened to have a driver. The file says how to swap it for a GPU,
+and carries commented blocks for the `[dev]`, `[train]` and `[docs]` extras.
+
+## Use
+
+```python
+import numpy as np
+from emu_hmf.model import HmfCorrection
+
+corr = HmfCorrection()                       # 200m; HmfCorrection(WEIGHTS["vir"]) for virial
+
+theta = np.array([0.049, 0.31, 67.36, 0.9649, 2.1, -1.0, 0.0, 0.06])
+#                 Ω_b    Ω_cb  H0     n_s     10⁹A_s  w    w_a  Σm_ν
+
+f = corr.fsigma(sigma=0.8, theta=theta, z=0.5)          # the multiplicity function
+n = corr.dndlnM(m, sigma, dlnsigma_dlnm, rho_cold, theta, z=0.5)   # the abundance
+```
+
+σ(M) is passed in, not computed: this package has no power spectrum and should
+not acquire one, and the σ(M) the fit was made against is the *cold* field
+against ρ̄_cb. Fitting f(σ) against one variance and evaluating it with another
+is the mismatch that makes a multiplicity function look wrong when the
+convention around it is what moved.
+
+Everything is JAX, so `jax.grad`, `jax.jit` and `jax.vmap` all work through the
+cosmology. That is the reason this exists rather than a table of numbers.
+
+## Two mass definitions, two files
+
+The correction is not the same function at two halo boundaries, so there is no
+single correction with a Δ argument. Both are fitted against a *Rockstar*
+spherical-overdensity mass, so the comparison isolates the boundary rather than
+mixing in a change of halo finder.
+
+| weights | halo definition | Tinker08 unchanged | recalibrated | improvement |
+| --- | --- | --- | --- | --- |
+| `WEIGHTS["200m"]` | SO 200 × mean, Rockstar | 7.00 % | **0.52 %** | 13.1× |
+| `WEIGHTS["vir"]` | SO virial, Rockstar | 10.92 % | **0.54 %** | 19.1× |
+
+rms in ln f, on 200 cosmologies held out *entirely* from training — not held-out
+rows. Each design contributes several hundred rows and at fixed cosmology ln f
+is smooth in σ, so a random row split measures interpolation between neighbouring
+masses of a cosmology the network has already seen. This correction is only ever
+asked for a cosmology it has not seen.
+
+Both files carry `WEIGHTS["200m"]`'s Δ = 200m Tinker08 as the carrier, so the
+virial weights absorb the *boundary change* as well as the recalibration. They
+are not a per-cent correction: at z = 0 they sit some 13 % below the 200m
+carrier. Reading "correction" as "small" at `vir` is a misreading.
+
+## Where it is defined
+
+Outside either bound the package refuses rather than extrapolating.
+
+**The cosmology** must be inside CSST's box, which is copied into `box.py` and
+checked against the emulator's own `param_limits` by `tests/test_box.py`:
+
+| Ω_b | Ω_cb | H₀ | n_s | 10⁹A_s | w | w_a | Σm_ν |
+| --- | --- | --- | --- | --- | --- | --- | --- |
+| 0.04–0.06 | 0.24–0.40 | 60–80 | 0.92–1.00 | 1.7–2.5 | −1.3–−0.7 | −0.5–0.5 | 0–0.3 |
+
+Note Ω_cb: CSST bounds the **cold** density, with massive neutrinos excluded.
+
+**The peak height** must be inside ν = δ_c/σ ∈ [0.5, 3], and the mass inside
+10¹²–10¹⁴ M⊙/h. Those two cuts do not commute with redshift: growth pushes σ
+down, so a fixed mass is a higher peak later, and the low-ν half of the band is
+simply absent above z ≈ 0.25. `target.nu_covered(z)` records what the training
+set actually spans — ν ≥ 1.4 by z = 3 — because a caller who checked only the
+nominal range would be extrapolating with no warning.
+
+The correction is a few per cent at z = 0 and grows with redshift, reaching
+about 12 % rms by z = 3. Quoting the low-redshift figure alone would understate
+it several-fold over most of the range it is defined on.
+
+## Documentation
+
+[emu-hmf.readthedocs.io](https://emu-hmf.readthedocs.io) — concepts, a tutorial
+with figures, the validity domain, and how to reproduce the training set.
+
+## Reproducing
+
+The 2000-cosmology training set (11.7 MB, both mass definitions) is archived
+with a DOI; the fit that turns it into the shipped weights needs only `optax`:
+
+```bash
+pip install emu_hmf[train]
+python -m emu_hmf.fit --shards ./shards --out weights.npz
+```
+
+Regenerating the shards themselves needs CLASS and the CSST emulator; see the
+documentation's *Reproducing the training set* page.
+
+## Citation
+
+If you use this package, please cite Tinker et al. (2008) for the functional
+form, Chen & Yu (2025) for the CSST emulator this is calibrated against, and
+this package for the recalibration. See `CITATION.cff`.
+
+## Licence
+
+BSD 3-Clause. See [LICENSE](LICENSE).

emu_hmf-1.0.0/docs/Makefile

@@ -0,0 +1,22 @@
+SPHINXBUILD   ?= sphinx-build
+SOURCEDIR     = .
+BUILDDIR      = _build
+
+.PHONY: help html strict figures clean
+
+help:
+	@$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)"
+
+html:
+	@$(SPHINXBUILD) -b html "$(SOURCEDIR)" "$(BUILDDIR)/html"
+
+strict:
+	@$(SPHINXBUILD) -W -b html "$(SOURCEDIR)" "$(BUILDDIR)/html"
+
+# Regenerate the committed figures.  Not part of `html`: the documentation
+# builder has no matplotlib, and the PNGs are in the repository.
+figures:
+	@python make_figures.py
+
+clean:
+	rm -rf "$(BUILDDIR)"

emu_hmf-1.0.0/docs/api/box.rst

@@ -0,0 +1,19 @@
+emu_hmf.box
+===========
+
+The hypercube the recalibration lives in, taken from the CSST emulator rather
+than chosen, and copied here rather than imported so that evaluating a mass
+function does not require installing a Gaussian-process emulator to find out
+what the bounds are.  ``tests/test_box.py`` asserts the copy against the
+emulator's own ``param_limits``.
+
+* :data:`~emu_hmf.box.PARAMS` — the column order, which is the emulator's own.
+* :data:`~emu_hmf.box.BOX` — the closed bounds.  Note that ``Omegam`` is the
+  **cold** density.
+* :func:`~emu_hmf.box.sample` — a Latin hypercube, deterministic in its seed.
+* :func:`~emu_hmf.box.check` — raise, naming every parameter outside the box.
+* :func:`~emu_hmf.box.inside` — the same, as data rather than an exception.
+
+.. automodule:: emu_hmf.box
+   :members:
+   :show-inheritance:

emu_hmf-1.0.0/docs/api/fit.rst

@@ -0,0 +1,21 @@
+emu_hmf.fit
+===========
+
+Fitting the four Tinker08 parameters as functions of cosmology and redshift.
+Needs ``optax``, which is the ``[train]`` extra — and nothing else, so this is
+the step an outsider can rerun from the archived shards.
+
+* :class:`~emu_hmf.fit.Shards` — one flat training set, with the record of what
+  it was built from.
+* :func:`~emu_hmf.fit.load_shards` — every shard in a directory, cut to the
+  trusted peak heights, refusing to mix mass definitions.
+* :func:`~emu_hmf.fit.fit` — the optimisation.  Splits on whole cosmologies,
+  pins float64, and writes its provenance into the output.
+
+.. code-block:: bash
+
+   python -m emu_hmf.fit --shards ./shards --out weights.npz
+
+.. automodule:: emu_hmf.fit
+   :members:
+   :show-inheritance:

emu_hmf-1.0.0/docs/api/generate.rst

@@ -0,0 +1,22 @@
+emu_hmf.generate
+================
+
+Building the training set: the emulator's abundance, converted into a
+multiplicity function in this package's variance convention.  Needs the
+generation stack — see :doc:`../reproducing`.
+
+* :func:`~emu_hmf.generate.solve_one` — one cosmology: one CLASS solve and one
+  emulator call.
+* :func:`~emu_hmf.generate.shard` — one contiguous slice of the design, written
+  in chunks and resumable.
+* :data:`~emu_hmf.generate.M_GRID`, :data:`~emu_hmf.generate.K_GRID`,
+  :data:`~emu_hmf.generate.CHUNK` — the grids and the write cadence.
+
+.. code-block:: bash
+
+   python -m emu_hmf.generate --shard 0 --n-per-shard 250 --n-total 2000 \
+          --out shards/hmf_000.npz --massdef RockstarM200m
+
+.. automodule:: emu_hmf.generate
+   :members:
+   :show-inheritance:

emu_hmf-1.0.0/docs/api/model.rst

@@ -0,0 +1,24 @@
+emu_hmf.model
+=============
+
+Inference: Tinker08 with cosmology-dependent parameters, in pure JAX.  This is
+the only module a user of the released package needs, and the only one that
+imports nothing beyond numpy and JAX.
+
+* :func:`~emu_hmf.model.tinker08` — the carrier, at
+  :math:`\Delta_{\rm m} = 200`, optionally corrected.  With ``g=None`` it is the
+  published fit, unchanged.
+* :class:`~emu_hmf.model.HmfCorrection` — the recalibrated multiplicity
+  function.  :meth:`~emu_hmf.model.HmfCorrection.fsigma` gives
+  :math:`f(\sigma)`, :meth:`~emu_hmf.model.HmfCorrection.dndlnM` the abundance,
+  and :meth:`~emu_hmf.model.HmfCorrection.g` the four log-corrections
+  themselves.
+* :data:`~emu_hmf.model.WEIGHTS` — one file per halo definition, keyed
+  ``"200m"`` and ``"vir"``.  See :doc:`../massdefs`.
+* :func:`~emu_hmf.model.load_weights` — the arrays and the provenance, cached
+  and read-only.
+* :func:`~emu_hmf.model.normalise` — :math:`(\theta, z)` onto the unit cube.
+
+.. automodule:: emu_hmf.model
+   :members:
+   :show-inheritance:

emu_hmf-1.0.0/docs/api/target.rst

@@ -0,0 +1,32 @@
+emu_hmf.target
+==============
+
+What is being fitted, and the conversions that reach it.  The functions that
+call the CSST emulator import it lazily, so this module is importable without
+it — but calling them is not.
+
+* :data:`~emu_hmf.target.FIDUCIAL` — a Planck-like point in the middle of the
+  box; the cosmology every worked example and quoted number here uses.
+* :data:`~emu_hmf.target.MASSDEFS`, :data:`~emu_hmf.target.DEFAULT_MASSDEF` —
+  the three halo definitions the emulator offers, and what each one *is*.
+* :data:`~emu_hmf.target.M_TRUSTED`, :data:`~emu_hmf.target.NU_TRUSTED`,
+  :data:`~emu_hmf.target.DELTA_C` — the fitted domain.
+* :data:`~emu_hmf.target.NU_COVERED`, :func:`~emu_hmf.target.nu_covered` — what
+  the training set *actually* spans at each redshift, which is narrower.  See
+  :doc:`../validity`.
+* :func:`~emu_hmf.target.to_ggah_cosmology`,
+  :func:`~emu_hmf.target.theta_from_cosmology` — the two directions of the
+  cosmology conversion, and the only places it happens.
+* :func:`~emu_hmf.target.csst_dndlnM`, :func:`~emu_hmf.target.csst_tinker08`,
+  :func:`~emu_hmf.target.sigma_chain` — the offline half; these need the
+  generation stack.
+
+.. note::
+
+   ``ggah_mod`` in the text below is the halo-model code this package was built
+   for, and ``emu_pk`` its linear-spectrum emulator.  Neither is needed to
+   *use* ``emu_hmf``; see :doc:`../halo_model`.
+
+.. automodule:: emu_hmf.target
+   :members:
+   :show-inheritance:

emu_hmf-1.0.0/docs/changelog.rst

@@ -0,0 +1,5 @@
+Changelog
+=========
+
+.. include:: ../CHANGELOG.md
+   :parser: myst_parser.sphinx_