bayesmith 0.1.0__tar.gz

bayesmith-0.1.0/.gitignore

@@ -0,0 +1,14 @@
+__pycache__/
+*.py[cod]
+*.egg-info/
+build/
+dist/
+.venv/
+venv/
+.pytest_cache/
+.hypothesis/
+.coverage
+htmlcov/
+.mypy_cache/
+.ruff_cache/
+.DS_Store

bayesmith-0.1.0/CHANGELOG.md

@@ -0,0 +1,32 @@
+# Changelog
+
+## 0.1.0 — 2026-08-26
+
+First release. Published so that downstream packages can declare a dependency
+on it by name rather than by path; until now the version was `0.0.0` and the
+package was not on any index.
+
+### What it does
+
+- **Graph core.** Deterministic and probabilistic nodes, plates, and the joint
+  log-density assembled from them.
+- **NumPyro bridge.** Any graph is runnable through NUTS, which is also the
+  oracle every exact path is verified against.
+- **Structural dispatch** with the linear-Gaussian exact solves: conjugate,
+  Wiener, GCR and GLS, selected per subgraph, with declarations such as
+  `linear_in` checked at three scales rather than trusted.
+- **Exact enumeration of discrete latents** (`bayesmith.exact.discrete`),
+  reading the `Discrete(n)` support declaration. Not yet dispatcher-selected —
+  see the README's Status section.
+- **Streaming evidence** as square-root information factors, combined exactly
+  across epochs.
+- **Graph diagnostics**: identifiability, prior sensitivity, linearity.
+- **Per-parameter convergence diagnostics** on chain paths: split r-hat and ESS
+  per coordinate, gated ESS-first because a fixed r-hat threshold is not a
+  well-posed test — see `r_hat_ceiling`'s docstring for the measurements.
+
+### Known limits
+
+- Forward-backward for chain-structured discrete latents is not implemented.
+- Discrete enumeration is not yet chosen by the dispatcher.
+- The API may move; this is an alpha release.

bayesmith-0.1.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Zheng Zhang
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

bayesmith-0.1.0/PKG-INFO

@@ -0,0 +1,87 @@
+Metadata-Version: 2.5
+Name: bayesmith
+Version: 0.1.0
+Summary: A graph of operators is a Bayesian model; its structure chooses the inference.
+Project-URL: Repository, https://github.com/zzhang0123/bayesmith
+Author: Zheng Zhang
+License: MIT
+License-File: LICENSE
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Mathematics
+Requires-Python: >=3.11
+Requires-Dist: equinox>=0.13
+Requires-Dist: jax>=0.5
+Requires-Dist: numpy
+Requires-Dist: numpyro>=0.15
+Description-Content-Type: text/markdown
+
+# bayesmith
+
+A Bayesian model is a graph of operators. Deterministic operators propagate
+dependence; probabilistic operators contribute a conditional density. Together
+they *are* the joint distribution.
+
+bayesmith makes that graph **explicit and inspectable**, and then uses its
+structure to choose how the model is fitted — an exact solve where the structure
+permits one, NUTS where it does not.
+
+```
+block 0  {x}          Wiener exact        (linear_in checked, 3 scales)
+block 1  {z}          enumerate 4 states
+block 2  {sigma, nu}  NUTS (numpyro)      no exact structure found
+```
+
+The model tells you how it will be fitted, before it is fitted.
+
+## What bayesmith is not
+
+It is **not another probabilistic programming language**. Distributions, MCMC
+kernels, variational inference and transforms all come from
+[NumPyro](https://github.com/pyro-ppl/numpyro). bayesmith is the dispatch layer
+above them, and every line in it must answer *"why can NumPyro not do this?"*
+
+What it owns, because a trace-based PPL structurally cannot:
+
+- **Structural exact inference** — conjugate / Wiener / GCR / GLS solves, and
+  exact enumeration of discrete latents, selected per subgraph.
+- **Streaming evidence** — square-root information factors combined exactly
+  across datasets and observing epochs.
+- **Diagnostics on the graph** — identifiability, prior sensitivity, and
+  linearity checking of the declarations the dispatcher relies on.
+
+Declarations such as `linear_in` are *claims about the model*, not hints, so
+they are **checked rather than trusted**: a node declared linear is probed at
+three scales before any exact solve is allowed to use it.
+
+## Status
+
+**0.1.0, the first release.** Published so other packages can depend on it by
+name. Alpha in the classifier's sense: the API may still move.
+
+Implemented and tested, 1163 tests: the graph core with plates and joint
+log-density; the NumPyro bridge, so any graph is runnable through NUTS;
+structural dispatch with the linear-Gaussian exact solves; exact enumeration of
+discrete latents; streaming evidence as square-root information factors; and
+graph diagnostics for identifiability, prior sensitivity and linearity.
+
+**Two things the page above describes that 0.1.0 does not do yet.** Stated here
+because a front page is a claim, and finding out afterwards is worse than
+reading it now:
+
+- **Enumeration is not dispatcher-selected.** `bayesmith.exact.discrete`
+  computes the exact marginal and the posterior marginals over declared
+  discrete latents, and reads the `Discrete(n)` support declaration to do it —
+  but `classify` does not yet route a discrete subgraph to it. The
+  `block 1  {z}  enumerate 4 states` line above is therefore a design sketch
+  rather than a transcript; call the module directly.
+- **Forward-backward is not implemented**, so a chain of `T` discrete latents
+  costs `n ** T` by enumeration rather than `T * n**2`. Enumeration refuses
+  past a budget rather than hanging, and names the count it would have visited.
+
+## License
+
+MIT

bayesmith-0.1.0/README.md

@@ -0,0 +1,66 @@
+# bayesmith
+
+A Bayesian model is a graph of operators. Deterministic operators propagate
+dependence; probabilistic operators contribute a conditional density. Together
+they *are* the joint distribution.
+
+bayesmith makes that graph **explicit and inspectable**, and then uses its
+structure to choose how the model is fitted — an exact solve where the structure
+permits one, NUTS where it does not.
+
+```
+block 0  {x}          Wiener exact        (linear_in checked, 3 scales)
+block 1  {z}          enumerate 4 states
+block 2  {sigma, nu}  NUTS (numpyro)      no exact structure found
+```
+
+The model tells you how it will be fitted, before it is fitted.
+
+## What bayesmith is not
+
+It is **not another probabilistic programming language**. Distributions, MCMC
+kernels, variational inference and transforms all come from
+[NumPyro](https://github.com/pyro-ppl/numpyro). bayesmith is the dispatch layer
+above them, and every line in it must answer *"why can NumPyro not do this?"*
+
+What it owns, because a trace-based PPL structurally cannot:
+
+- **Structural exact inference** — conjugate / Wiener / GCR / GLS solves, and
+  exact enumeration of discrete latents, selected per subgraph.
+- **Streaming evidence** — square-root information factors combined exactly
+  across datasets and observing epochs.
+- **Diagnostics on the graph** — identifiability, prior sensitivity, and
+  linearity checking of the declarations the dispatcher relies on.
+
+Declarations such as `linear_in` are *claims about the model*, not hints, so
+they are **checked rather than trusted**: a node declared linear is probed at
+three scales before any exact solve is allowed to use it.
+
+## Status
+
+**0.1.0, the first release.** Published so other packages can depend on it by
+name. Alpha in the classifier's sense: the API may still move.
+
+Implemented and tested, 1163 tests: the graph core with plates and joint
+log-density; the NumPyro bridge, so any graph is runnable through NUTS;
+structural dispatch with the linear-Gaussian exact solves; exact enumeration of
+discrete latents; streaming evidence as square-root information factors; and
+graph diagnostics for identifiability, prior sensitivity and linearity.
+
+**Two things the page above describes that 0.1.0 does not do yet.** Stated here
+because a front page is a claim, and finding out afterwards is worse than
+reading it now:
+
+- **Enumeration is not dispatcher-selected.** `bayesmith.exact.discrete`
+  computes the exact marginal and the posterior marginals over declared
+  discrete latents, and reads the `Discrete(n)` support declaration to do it —
+  but `classify` does not yet route a discrete subgraph to it. The
+  `block 1  {z}  enumerate 4 states` line above is therefore a design sketch
+  rather than a transcript; call the module directly.
+- **Forward-backward is not implemented**, so a chain of `T` discrete latents
+  costs `n ** T` by enumeration rather than `T * n**2`. Enumeration refuses
+  past a budget rather than hanging, and names the count it would have visited.
+
+## License
+
+MIT

bayesmith-0.1.0/docs/migration/README.md

@@ -0,0 +1,100 @@
+# Cross-check records — and what §六 is still waiting for
+
+Migration spec §二 requires one page per module and gates **all** of §六
+(rheplicant's wind-down) on step 6: *"全部通过后，才动 rheplicant 侧对应模
+块"*. This directory is where those pages live. It did not exist until
+2026-08-25, which is why §六 had never started.
+
+**Do not read the table below as authority.** It is prose, and prose has no
+test. `tests/test_migration_records.py` is the authority: it
+derives the module list from the spec's own §四 tables, the page list from
+this directory, and the test list from `tests/crosscheck/`, and fails when
+they disagree. If the table and that test ever disagree, the test is right.
+
+| §四 row | module | cross-check test | page |
+|---|---|---|---|
+| 4.1 | `linear.py` → `exact/block,linearity,solve` | `test_linear.py` | ✅ |
+| 4.1* | `conditioning.py` → `exact/conditioning` | `test_conditioning.py` | ✅ |
+| 4.1 | `gls.py` → `exact/gls` | `test_noise_logdet.py` (B1 half) | ✅ |
+| 4.1 | `uncertainty.py` (Fisher) → `exact/fisher` | `test_noise_logdet.py` | ✅ |
+| 4.1 | `likelihood.py`/`noise.py` → `exact/gaussian` | `test_gaussian.py` | ✅ `noise.md` |
+| 4.2 | `parameters.py` → node declarations | `test_parameters.py` | ✅ |
+| 4.2 | `noise.py` → probabilistic nodes | `test_gaussian.py` | ✅ `noise.md` |
+| 4.2 | `plan.py`+`engines.py` → dispatch | `test_dispatch.py` | ✅ `plan.md` |
+| 4.2 | `identifiability.py` → `diagnose/` | `test_diagnose_identifiability.py` | ✅ |
+| 4.2 | `sensitivity.py` → `diagnose/` | `test_diagnose_sensitivity.py` | ✅ |
+| 4.2 | `priors.py` → `diagnose/` | `test_diagnose_jeffreys.py` | ✅ |
+| 4.2 | `numpyro_bridge.py` → `bridge/` | — | — |
+| 4.3* | `sqrtinfo` (rewritten; kernel preserved per B11) | `test_sqrtinfo_agrees.py` | ✅ |
+
+`*` — has a page but **no source row of its own** in §四, and the test
+records why. `conditioning.py` appears only in the `linear.py` row's
+DESTINATION cell (upstream moved it to `rheplicant.core` so `radio` could
+use it without importing `inference`); `sqrtinfo` belongs to the evidence
+layer, which §四 4.3 marks 不迁移 — rewritten under iron law 2, with its
+numerical kernel required to be preserved exactly and a bitwise
+cross-check enforcing it.
+
+## Where the gate actually stands
+
+**Open, since 2026-08-25.** Twelve pages exist and **every §四 module has
+one**. `tests/test_migration_records.py::test_the_gate_on_section_six_is_open_and_every_module_has_a_page`
+now asserts that in the other direction: while the list was non-empty it
+blocked §六, and now that §六's steps are being taken against these pages,
+a page may not go missing.
+
+| §四 row | closed |
+|---|---|
+| `linear.py` → `exact/{block,linearity,solve}` | `linear.md` |
+| `likelihood.py`/`noise.py` → `exact/gaussian` | `noise.md` |
+| `gls.py`, `uncertainty.py` (Fisher) | `gls.md`, `uncertainty.md` |
+| `parameters.py` → node declarations | `parameters.md` |
+| `noise.py` → probabilistic nodes | `noise.md` |
+| `plan.py`+`engines.py` → dispatch | `plan.md` |
+| `identifiability.py`, `sensitivity.py`, `priors.py` → `diagnose/` | three pages |
+| `numpyro_bridge.py` → `bridge/` | `numpyro_bridge.md` |
+| *(out of ledger)* `conditioning.py`, `sqrtinfo` | two pages, with their reasons |
+
+**What that does and does not authorise.** §六 step 1 still governs:
+nothing in `src/rheplicant/inference/` moves except the two exceptions
+already in e-RHINO's Track A Batch 1 (B1's `plan.py` docstring and B4's
+one-line fix), plus docstring pointers to bayesmith. Read §六's five steps
+before starting, and note step 3's possible dividend — the two pytest
+sessions may be able to merge once the evidence layer is out.
+
+## What a page must contain
+
+The five headings §二 names, in its order. The P5 pages are the worked
+examples.
+
+**One thing NOT to write: a case count, unless the module is finished.**
+Some pages here carry one and some do not, and the asymmetry is a
+measurement rather than an oversight. The P5 pages' counts (17, 17, 10)
+were written in session 3 and are still exact, because nothing touched
+those modules afterwards. `plan.md`'s said 6 and was 7 within hours,
+because its author kept working on the module after writing the page. So
+the rule is about timing, not taste: a count is a safe thing to record only
+once the thing counted has stopped moving, and the four pages written
+during active work say `pytest --collect-only -q` instead. Nothing in this
+repository reads a count out of a page, so none of them can go red.
+
+The five headings:
+
+1. **Fixtures** — reused from rheplicant where possible (they are pinned
+   measurements, not fresh guesses), at least one healthy and one that must
+   be refused.
+2. **Numerical agreement** — deterministic exits to float64 roundoff,
+   sampled exits within MC error. Say which tolerance and why; where
+   exactness is not mathematically available (a null direction is a ray, a
+   null space is basis-dependent), say what is compared instead.
+3. **Refusal agreement** — every loud rheplicant refusal has a same-shape
+   refusal here, with the exception-class mapping recorded.
+4. **Independent oracle** (iron law 4) — analytic truth, a hand-written
+   NumPyro model, scipy, or mutation. *Two implementations agreeing is not
+   evidence.*
+5. **Intended differences** — with the reason and the equivalence argument.
+   A difference that is deliberate and unrecorded is indistinguishable from
+   a bug six months later.
+
+Iron law 5 adds one hard item: any §三 defect belonging to the module is
+fixed **before** step 2, or written as a signed, sized expected difference.

bayesmith-0.1.0/docs/migration/conditioning.md

@@ -0,0 +1,74 @@
+# Cross-check: `conditioning`
+
+`rheplicant.core.conditioning` (moved there from `inference/` upstream when
+`radio` needed it and could not import `inference`) →
+`bayesmith.exact.conditioning`. Test: `tests/crosscheck/test_conditioning.py`
+(9 cases). Page written 2026-08-25 from that test's assertions, re-run on
+that date; the cross-check itself predates this page.
+
+## 1. Fixtures
+
+A symmetric operator with an EXACTLY known spectrum, built as
+`Q diag(λ) Qᵀ` from a QR of a fixed-seed normal matrix and handed to both
+packages as the same callable — so a difference is the algorithm and never
+the setup. Spectra: `[1,2,3]`, `[1e-6,1,5,5.5]`, six-fold degenerate
+`[1]*6`, and a 20-point geometric spectrum with κ = 1e4.
+
+For `tree_norm`, three pytrees including the **overflow** case both module
+docstrings argue about: `{a: [1e20, 1e20], b: [0]}`. Squaring first turns
+1e20 into `inf`, and the scaling that avoids it is the entire content of
+the function — a port that dropped it agrees on the ordinary row and
+disagrees here.
+
+## 2. Numerical agreement
+
+| quantity | agreement |
+|---|---|
+| `tree_norm`, all three trees | **exact float equality** |
+| `largest_eigenvalue` (same operator, template, key, 12 iterations) | **exact float equality** |
+
+Equality rather than a tolerance, because it is the same arithmetic in the
+same order on the same library; a tolerance would let a genuinely different
+iteration pass, which is what this file exists to notice.
+
+## 3. Refusal agreement
+
+Neither module refuses anything — these are numerical estimators, not
+gates. The refusal that *consumes* them (`condition_bound`'s κ ceiling)
+belongs to `solve.py` and is checked in `tests/exact/`.
+
+## 4. Independent oracle
+
+`jnp.linalg.eigvalsh` of the materialised matrix. Both packages'
+`largest_eigenvalue` must approach the true λ_max **from below** (within
+1e-5 above, and within 1e-3 relative) — "agreeing on a wrong number is
+still agreement", so the shared claim is checked against truth as well as
+against each other.
+
+## 5. Intended differences — one, and it is the reason the harness exists
+
+**`extreme_eigenvalues` was deliberately NOT ported.** rheplicant estimates
+λ_min by running a second power iteration on `λ_max·I − M`, which on a
+graded spectrum cannot separate the eigenvalues crowded against λ_max — and
+the error is **one-sided and toward danger**: λ_min comes back too LARGE,
+so κ too SMALL, so a convergence guard built on it is silent exactly when
+it should fire. Measured upstream at κ = 1e4: λ_min over-large by 33.9×,
+reported κ = 2.947e+02 against a true 1.000e+04.
+
+bayesmith bounds λ_min below by the prior's own curvature instead
+(`exact/solve.py::condition_bound`), giving an UPPER bound on κ — the
+direction a safety guard needs.
+
+Both halves are live tests, not prose:
+
+- `test_bayesmith_does_not_carry_extreme_eigenvalues` goes red if someone
+  ports it after all, and points at the docstring that rejected it.
+- `test_rheplicant_still_carries_it_and_still_leans_the_unsafe_way`
+  asserts the bias as an **inequality against the truth** (λ_min > 10×
+  true), so it survives an iteration-count retune. When rheplicant fixes
+  this, that test goes red and should be deleted together with the
+  paragraph it guards.
+
+This is a **Track A item for e-RHINO, not this migration** — recorded here
+because it is the first time the harness's reason for existing paid out: a
+manual comparison holds only on the day it is written.

bayesmith-0.1.0/docs/migration/gls.md

@@ -0,0 +1,81 @@
+# Cross-check: `gls` — and the B1 log-determinant ledger
+
+`rheplicant.inference.gls` (`iterative_gls`) → `bayesmith.exact.gls`
+(§四 4.1). Test: `tests/crosscheck/test_noise_logdet.py` (17 cases, shared
+with the Fisher row — see [uncertainty.md](uncertainty.md)) plus
+`tests/exact/test_gls.py::test_the_fixed_point_is_the_unbiased_estimator_
+not_the_gls_biased_one` on this side. Page written 2026-08-25 from those
+tests' assertions, re-run on that date; the cross-checks predate this page.
+
+**This page is unusual: its subject is a difference, not an agreement.**
+§三 B1 is a defect in rheplicant, and iron law 5 forbids aligning a new
+implementation to a defective one. So the acceptance is a *signed, sized*
+expected difference on one side and a proof of correctness on the other.
+
+## 1. Fixtures
+
+A constant-mean model under `RadiometerNoise`, so both estimators have a
+closed form and nothing samples: `f ∈ {0.05, 0.2, 0.5}`, `n = 2000` for the
+closed-form rows and `n = 200 000` for the asymptotic one (measured: at
+f=0.5 the ratio lands 0.024 % from `1+f²` at n=200 000, against 2.5 % at
+n=2000). A `HomoscedasticNoise` row is the anti-vacuity control.
+
+## 2. Numerical agreement — of each estimator with its own closed form
+
+| claim | agreement |
+|---|---|
+| `include_logdet=False` maximiser = `Σd²/Σd` | rel **1e-8** |
+| `include_logdet=True` maximiser = positive root of `n f² μ² + μ Σd − Σd² = 0` | rel **1e-8** |
+| ratio of the two = `1 + f²` | rel **2e-3**, at three f |
+| the full density's estimate = the truth | rel **5e-3** |
+
+The quadratic's root is derived in the test rather than taken from
+rheplicant's docstring, because the asymptotic claim rests on it: it is
+what makes the full density **exactly unbiased** rather than differently
+biased, and a test asserting only "closer to the truth" would pass on an
+estimator that was merely less wrong.
+
+The maximiser is found by **scipy on a Python closure** over rheplicant's
+likelihood — no JAX gradient, no algebra of ours. A closed form checked
+against its own rearrangement would be checking arithmetic, not the
+estimator.
+
+## 3. Refusal agreement
+
+None to compare: neither package refuses the mixture. That is the defect —
+B1's point is that rheplicant's `nuts` route (whose `dist.Normal` carries
+its own `−log σ`) and its `plan.sample` gradient block (which does not)
+target different estimators on one model, **with no guard between them**,
+while the same package's evidence layer raises full-versus-GLS to a
+refusal.
+
+## 4. Independent oracles
+
+- The two closed forms above (algebra, not either implementation).
+- scipy's optimiser, sharing nothing with either package's gradients.
+- **Anti-vacuity**: under `HomoscedasticNoise` the log-determinant is an
+  additive constant and the two estimators must **coincide** — so the gap
+  is attributed to the prediction-dependence and not to some other
+  difference between the two likelihood spellings.
+- The ratio is checked across an f that varies the answer by a factor of
+  25 between its ends, so a constant fudge cannot satisfy it; and the SIGN
+  is asserted separately, because it is the half that says which engine is
+  the optimistic one (`gls > full > 0`).
+
+## 5. Intended differences — the whole point of the row
+
+**bayesmith's `iterative_gls` does NOT carry this bias, and must not be
+made to.** It is frozen-sigma IRLS: each inner solve holds σ fixed and
+recomputes it afterwards, so its fixed point satisfies `w = mean(u)`,
+`u = d/x` — the same side as the full density. Measured: the fixed point is
+44–128× closer to `mean(u)` than to `Σu²/Σu`, at `κ ∈ {0.05, 0.2, 0.5, 1.0}`.
+
+The spec's first draft asked for the opposite — that bayesmith's frozen-σ
+path differ from a live-σ path by `(1+f²)` — and building that test would
+have pulled a correct estimator toward a bias it does not have. **The
+initial statement of the acceptance was wrong, and measurement is what
+found it**; the correction is recorded in §三 B1 itself, marked
+`[实测确认]`.
+
+Consequence for the pending `plan`/`engines` row (§四 4.2): B1 must land
+first, or that comparison will fix the GLS-type target as the reference.

bayesmith-0.1.0/docs/migration/identifiability.md

@@ -0,0 +1,86 @@
+# Cross-check: `identifiability`
+
+`rheplicant.inference.identifiability` → `bayesmith.diagnose.identifiability`
+(migration spec §八 step 5, acceptance §四 4.2). Executed 2026-08-25; every
+number below was measured in this repo's venv on that date.
+
+## 1. Fixtures
+
+Per §二 step 1, rheplicant's own pinned fixtures, re-expressed on the graph
+(`tests/diagnose/models.py`):
+
+- **Healthy + degenerate pair**: `data[t,f] = gain[t] * (T_ant[t,f] +
+  tone[f])`, 8×8, antenna temperature free-per-cell (72 parameters, the
+  under-determined case) or through a (3,3) polynomial basis (17). Same
+  coefficient matrix, deliberately asymmetric.
+- **Must-refuse fixtures**: ambient float32; a graph pinning its own
+  arithmetic to float32; complex/integer selected latents; empty/unknown/
+  repeated names; a graph with no observed node; a Poisson observation
+  (no `loc`).
+- Auxiliary: mixed-scale (1e10 unit gap), zero-column (a dead latent),
+  sort-trap (declaration order ≠ sorted order).
+
+## 2. Numerical agreement
+
+`tests/crosscheck/test_diagnose_identifiability.py`, same arrays handed to
+both packages (never the same construction recipe — §0.1's PRNG trap):
+
+| quantity | agreement |
+|---|---|
+| (n_par, rank, nullity), all four rows | equal, cell for cell: (72,64,8)×2, (17,17,0), (17,16,1) |
+| singular values (basis/off) | rel 1e-10 |
+| `weakest_identified` | rel 1e-9, both 4.822138e-05 |
+| null direction, nullity 1 | elementwise abs 1e-9, sign-fixed on the largest entry |
+| null space, nullity 8 | as projectors `VᵀV`, abs 1e-9 — rows are basis-dependent, the subspace is not |
+
+## 3. Refusal agreement
+
+Every loud rheplicant refusal has a same-shape refusal here (empty/unknown/
+repeated names, complex → the R-linear explanation, integer, out-of-range
+direction index, inconsistent report). Exception classes map
+`ParameterSpaceError → GraphError`, `StateValidationError → StructureError`
+— bayesmith's own taxonomy; the shared-identity concern of the design doc's
+appendix applies to rheplicant's importers, not to new code here.
+
+**One refusal changed mechanism, deliberately.** rheplicant forces
+process-global x64 *inside* the diagnostic and refuses only a model that
+pins its output to float32. This package's rule is that `src/` never touches
+`jax.config`, so there are TWO refusals: ambient float32 (before any
+tracing — a graph built in x64 and called outside it otherwise dies inside
+`jax.linearize` with a bare dtype inconsistency) and result float32 (the
+graph itself casts). Both are pinned, and the float32 counterfactual is
+computed by hand in the suite to show what they protect.
+
+## 4. Independent oracles (iron law 4)
+
+- "Moving along the reported direction does not move the model": the
+  end-to-end statement, evaluated through the graph itself, against a
+  random direction of the same size (< 1e-3 of the random movement), on
+  both the over- and under-determined fixtures.
+- The free model's `weakest_identified` = 1/√2 exactly — fixed by the
+  fixture's geometry, not by either implementation.
+- The no-normalisation counterfactual: the mixed-scale fixture's raw
+  spectrum ratio is measured (1.000000e-10) and shown to sit below the
+  default tolerance.
+- 5 of the 17 hand-applied mutations in the port's mutation pass targeted
+  this module; all killed.
+
+## 5. Intended differences
+
+1. **The x64 mechanism** (above). Consequence: `DEFAULT_RANK_RTOL = 1e-8`
+   was **re-measured, not ported**, per the spec's recorded trap. Under
+   this regime: null direction 7.479266e-17 (upstream arithmetic 6.6e-17 —
+   the spectrum moved with evaluation order, the verdict did not), weakest
+   identified 4.822138e-05 (identical to every printed digit), float32
+   null direction 3.116759e-08 (upstream pinned 3.1168e-8). The window and
+   the constant survive; the justification now cites this side's numbers.
+2. **Signature**: `(space, pipeline, state_template)` → `graph`. `at`
+   defaults to prior centres (`prior_environment`, the dispatch layer's own
+   anchoring rule) rather than declared inits — the same point, one
+   spelling.
+3. **The Jacobian's row layout** is `dense_operator`'s (observed nodes in
+   sorted name order, flattened); rheplicant's is a single prediction
+   array. Irrelevant to the rank and the null space; recorded because the
+   `jacobian` field is public.
+4. rheplicant's raw-`Bind`-space and x64-subprocess tests do not port
+   (no such concepts here); the refusal-mechanism family replaces them.