Repository navigation
Differentiability gradient-audit + hardening arc (registry = release gradient-gate) - #8
Merged
Merged
Conversation
Validated design from the Phase-1 brainstorm (HITL with Anna): - shared core + thin pytest/script layers so the gate and the website table cite the same measured AD-vs-FD numbers (single source of truth) - explicit expect taxonomy (consistent / known_zero / known_blocked); known limitations PINNED, not skipped - per-class tolerance (1e-5 closed-form, 1e-3 ODE) + relative central-FD step - per-channel physical IC reductions (mean r / mean v / mean m) - full surface in 4 tiers; T4 consolidates all scattered grad tests into the registry (the release gradient-gate; fixes audit T6) - confirmed-unfixed hazards -> xfail(strict) + hazard-id (self-cleaning ratchet) - preliminary probes: H4/H6 verify-benign; alpha=1-exactly is a known branch point; live suspect surface narrows to H2 (psi=0 masks) + shared r_t kink Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…se 2) Strict RED->GREEN task-by-task plan, 4 tiers + checkpoints: - Tier 0: TDD the engine (toy NaN/silent-zero/wrong-by-2x must be caught) + reductions + parametrized gate + audit_gradients.py - Tier 1: headline samplers + packaged summaries + king r_t known_zero pin + alpha=1 known_blocked split + H4/H6 boundary probes; mutation-check teeth - Tier 1.5: living website validation doc, build to 0 warnings - Tier 2: MultiComponentCluster A/B (H2 psi=0 + r_t-kink probes), binaries, rotation - Tier 3: vendored frozen-edge binner -> binned sigma/beta/N Fisher path - Tier 4: inventory + migrate/dedupe all scattered grad tests into the registry (release gradient-gate; fixes audit T6); plots; completion doc; ONE final PR Plan-time refinement: segregation is a consistent (non-zero) pin, not known_zero (proves stop_gradient didn't block the real gradient); only king r_t is known_zero. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…by-2x + expect taxonomy
… type-guard expect; drop dead import
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…zard_id; drop dead import (review)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
King positions (W0, r_c), King velocity DF (W0): all clean (FD-consistent, tol=1e-3); W0=12 edge clean via explicit ODE domain (xi_max=700) so the traced-W0 pinned-r_t guard does not trip. KingProfile.r_t added as expect=known_zero -- at theta0=7.0 r_t is grid-snapped so AD~0 AND FD~0 (known-limitation). MEASURED finding: at W0=8.0 the r_t VALUE moves with W0 (FD=62.5, fixed grid) while AD stays 0 -- a genuine silent-zero on the argmax-protected path, flagged for Anna's triage. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ichie/LIMEPY/multicomponent via unclamped-psi interpolation (audit finding; forward r_t now interpolated crossing, Anna-ratified; no compat shim) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…rentiable) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…fferentiability Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… alpha=1 known_blocked split Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…prox summaries (segregation stop_gradient is a consistent non-zero pin)
…scale — restores FD-consistent gradient at default beta (audit finding; forward value unchanged, Anna-ratified)
…e (Tier 1.5) One row per (entry point x param x direction) with REAL measured AD/FD ratios from validation/data/grad_audit_results.json (38 cases, 35 clean / 3 known-limitation). Documents the two audit-found hazards fixed gradient-only (king r_t silent-zero across 5 models; segregation stop_gradient ~27% bias), the alpha=1 known_blocked branch points, the expect-class taxonomy, and a changelog. Wired into myst.yml validation TOC + index intro. Builds 0 warnings. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ases (Task 2.1) Add the params->IC grad-audit cases for the lowered-isothermal multimass equilibrium: from_imf(...).sample_cluster reduced by mean_radius, audited in the three traceable equilibrium drivers W0, g, delta (n_comp=3, n_stars=400). Measured-first frozen gates (AD vs central-FD, tol=1e-3): W0 @5.0: AD=4.709932e-1 FD=4.710074e-1 |ratio-1|=3.0e-5 g @1.0: AD=1.459179e-1 FD=1.459341e-1 |ratio-1|=1.1e-4 delta @0.5: AD=8.716899e-1 FD=8.716973e-1 |ratio-1|=8.5e-6 H2 probe (ψ=0 boundary masks @ multicomponent.py:266/272/286 + shared r_t kink) CLEARED BENIGN via a W0=3[H2-boundary] edge (most extended, r_t=4.26, |ratio-1|=2.6e-4): (1) the shared tidal radius r_t IS differentiable in W0 through from_imf (ratio 1.000000 at W0=5,7) — the Tier-1 unclamped-ψ fix (9d3365a) still flows here; (2) no sampled star ever reaches ψ(r)<=0 (the boundary shell carries ~zero mass; 0/400 hit the max(ψ,0) clamp), so the jnp.where(psi>0,...) masks are structurally inert for the sampled stars. Categorical-assignment discreteness (OUT of scope) side-stepped by case design, not by weakening tol: at h=1e-4 the fixed-key categorical draw is STABLE — 0/400 assignment flips for all of W0, g, delta — so the FD is the pure smooth position-physics derivative, uncontaminated. Gate teeth verified RED (tol=1e-8 -> hazard at ratio 1.000258) then GREEN (tol=1e-3). Full grad-audit gate: 42 passed (was 38), 0 hazards. JSON regenerated. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…rements (review fixes) Code-review precision fixes to the H2-probe comments in the Engine A sample_cluster cases (comments only; no code change): - Fix wrong file:line citations. The ψ=0 / boundary masks were cited as multicomponent.py:266/272/286: :266 is blank, :272 IS the genuine jnp.where(psi_grid>0,...) xi-grid support mask (kept), :286 is the jnp.where(r_grid<=r_t,...) RADIAL truncation mask (relabeled accurately, not a ψ=0 mask). The max(ψ,0) clamp was mis-cited at multicomponent.py:66 (a docstring line) — the real clamps are sampling.py:64-66 (per-star W_i on the sampling path) and limepy.py:275 (model-build grid clamp). - Replace the W0=5-only "benign" evidence with the W0=3 stress-test numbers (most extended cluster, r_t=4.26, boundary most stretched): measured min ψ(r_sampled)=0.153 (max sampled radius 3.66 < r_t), 0/400 stars hit the max(ψ,0) clamp, 0/400 categorical assignment flips at h=1e-4. The boundary is approached more closely at W0=3 but never reached; the jnp.where masks carry zero live gradient at the edge (masked-out density past r_t is ~0). H2 stays benign; gate green (42 passed). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…s cases (Task 2.2) Add three params->IC grad-audit Cases through the density-defined shared-Psi Eddington/OM build (Poisson quadrature + Eddington inversion, NO ODE) + per-star draw, on the proven-realizable Plummer-halo + EFF-core headline mix reused from test_engine_b_physics._headline_model: - halo Plummer r_h -> mean_radius: AD=6.767522e-1 vs FD=6.767957e-1 (|r-1|=6.4e-5) - core EFF gamma -> mean_radius: AD=-1.204971e-1 vs FD=-1.205328e-1 (|r-1|=3.0e-4) - halo OM r_a_j[0] -> mean_speed: AD=1.768007e-3 vs FD=1.768007e-3 (|r-1|=2.7e-9) All FD probes (theta +- h) stay realizable (f_min_j > 0 measured at every point; the EFF-core ~1.185e-4 margin is small but strictly positive and flat), so the central FD never trips the concrete negative-DF ValueError. 0/400 categorical assignment flips at h for all three params, so FD is discreteness-free. tol=1e-3 (set by the widest band, gamma at 3.0e-4). Reduced resolution n_r=3000, n_e=500 (AD==FD is a graph property; both sides share grids). Full grad-audit gate: 45 passed, 0 hazards. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ed-guard sentinel (review note)
…ary_components / MoeCompanions (Task 2.3) Add three params->IC binary-layer cases to the grad-audit registry, all classified consistent/clean by measurement (no weakening, no cherry-picking): - KeplerElements.to_state (param e): elements -> Cartesian position via the fixed-iteration Newton Kepler solve + trig rotation. mean_radius, tol=1e-5. AD/FD ratio 1.0000000 at e=0.5 AND at the e=0.999 near-parabolic EDGE (the max(1-e²,1e-12) guard is inactive there; 50 Newton iters converge). - resolve_binary_components (param a): binary->spatial connector, N=50 all binaries (sanitization inactive). mean_radius over (2N,3), tol=1e-5, ratio 1.0000000. - MoeCompanions.sample (param m1-scale): a sampler MIXING a smooth grid-CDF (P,q,e) draw with a DISCRETE is_binary Heaviside mask + m2 gate. Audited via <e> (mask-independent P->e coupling). Classified clean by measurement: 0/400 is_binary flips at s=1±h (the discrete selection injects no FD noise), <e> ratio 1.0000000 (|AD|=1.02e-4 >> eps); even the gated <m2=m1·q> is ratio 1.0000000 since 0 flips keeps the where() locally smooth. The grid-CDF reparameterization makes the coupling gradient flow; the mask does not block it. Registry: 45 -> 48 baseline Cases; JSON 49 rows (incl Kepler e=0.999 edge). Full grad-audit gate: 49 passed, 0 hazards. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ocity-channel + Moe is_circular branch, add mixed-is_binary resolve case (Task 2.3 review) Issue 1: the resolve_binary_components comment claimed AD/FD=5.872715e-2, but the committed _resolve_binary_a closure (PRNGKey(11)) measures 8.437447e-2 (ratio 1.0000000008, matches JSON). The 5.87e-2 was a stale draft value from an earlier PRNGKey(0) variant; corrected the comment to the true numbers and noted the seed provenance. Case is valid/clean. Issue 2: re-measured the Kepler to_state VELOCITY channel at e=0.999. With the engine's central FD it is machine-precision clean (AD=-7.484823e-2 vs FD=-7.484823e-2, ratio 1.0); it only softens under a coarse one-sided FD (forward h=1e-3 -> ~0.99993) from the d/de[sqrt(1-e^2)] curvature. Documented that the velocity gradient is NOT blocked, only FD-noisier, hence position is the audited (cleaner) channel. Issue 3: flip-counted the MoeEccentricity is_circular branch (etap1<=1e-6). For the closure masses, 10/400 stars are already deeply circular (min etap1=-698) but 0/400 flip at s=1+-h, so the discrete jnp.where(is_circular,0,e) injects zero FD noise -- consistent with <e> ratio 1.0. Documented. Issue 4: added resolve_binary_components[mixed] case (25 binaries + 25 singles) exercising the single-star sanitization path (a_safe/e_safe/m2_safe, assembly.py:85-87). Confirmed jacfwd is NaN-free with d/da=0 on every single slot and a live gradient on every binary slot -- the sanitization does not leak NaN into a Fisher matrix. AD=3.019178e-2 vs FD=3.019178e-2 (ratio 1.0). Regenerated grad_audit_results.json (50 cases, 0 hazards); full gate 50 passed. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… (Task 2.4)
Add the three rotation kinematic-overlay grad-audit cases (params->IC),
completing Tier 2 of the differentiability gradient-audit arc:
- apply_solid_body_rotation in omega: EXACTLY-linear overlay, AD matches the
closed form mean(|axis x r|)=1.1226200071 to <1e-12; AD/FD ratio 1.0 (6.9e-13).
- apply_differential_rotation in v_peak: linear in v_peak, machine-exact (5.8e-13).
- apply_differential_rotation in R_peak: genuinely NONLINEAR (R_peak in both the
(R/R_peak) prefactor and exp(1-R/R_peak)); FD-consistent to 5.5e-8. tol=1e-5.
All three use zero base velocities to isolate the pure rotation overlay over a
fixed Plummer position sample + fixed z-axis. Regenerated
validation/data/grad_audit_results.json (53 cases, 0 hazards).
Website doc (50-validation/differentiability-audit.md): refresh for ALL of Tier 2
(Engine A, Engine B, binaries, rotation) -- updated the case count 38 -> 53, added
a "multi-component clusters, binaries & rotation" status list-table with real
measured ratios from the JSON, and a Tier-2 findings narrative (H2 psi=0 boundary
benign; discrete is_binary/is_circular branches don't block the q/e gradient;
single-star sanitization NaN-safe under jacfwd; Engine A/B r_t differentiable).
Tier 2 added 15 cases, all clean, 0 new hazards -> changelog unchanged. MyST build
clean (0 warnings).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…nlinear teeth (review nit)
…her-path cases + mutation check (Task 3.1; N(r) framing pending review) Tier 3 of the differentiability gradient-audit arc: audit the exact Jacobian J = d(binned kinematic summary)/d(theta) that the Fisher matrix F = JᵀJ is built from, for the velocity-dispersion sigma_1d(r) and Binney anisotropy beta(r) channels. Part A: vendor the five frozen-edge binners (_bin_index, _grouped_bin_sums, binned_sigma1d, binned_sigma_beta + SigmaBetaResult, binned_number_density) VERBATIM from scripts/_demo_inference.py @ 2090dcd into tests/validation/grad_audit/binners.py so the audit has a stable binner independent of the (non-package) demo script. Part B: two params->summary Fisher cases through build_spatial_ic + the frozen-edge binner. Both are machine-precision FD-consistent (ratio 1.0000000) at h_rel=1e-5: - binned_sigma1d[Plummer] d(sig_hat)/d(r_h): AD=-3.345969e+00, ratio 1.0 - binned_sigma_beta[Plummer+OM] d(beta_hat)/d(r_a): AD=-1.000759e+00, ratio 1.0 The FD step is tightened to h_rel=1e-5 (not tol weakened): at the default h_rel=1e-4 a single boundary star (#821, r=7.99977) straddles the r=8 outer edge and corrupts the COARSE central FD to ratio 0.86 — a coarse-FD edge-straddle, NOT an autodiff bug (per-bin ratios all 1.0; ratio stable for h<=1e-5 and across N=2000/4000/8000). At h_rel=1e-5: 0/2000 bin crossings at +-h. Part B2: test_binned_sigma_mutation_has_teeth — wrapping (pos,vel) in stop_gradient before the binner collapses AD to 0.000000e+00 while the live FD stays -3.35e+00, proving the case differentiates the real params->summary channel (not a constant). N(r): probed BOTH framings (data-side count vs model-side analytic Plummer p_k) but did NOT add a Case — the framing decision is pending review (CHECKPOINT 3). Gate: 55 audit case-rows + 1 mutation test, 0 hazards (was 53). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…leaves bin membership invariant, default h is clean (review fix)
…n (N(r) model-side Fisher channel) + pinned data-count non-diff case + website (Task 3.1 N(r)) Add the public PlummerProfile.enclosed_mass_fraction CDF M(<r)/M = r^3/(r^2+a^2)^(3/2), the differentiable per-shell occupancy p_k(theta) that a Poisson number-density likelihood differentiates -- the model-side N(r) Fisher channel. The frozen binned data count remains correctly non-differentiable (AD=0 by design), pinned as a known_blocked limitation. - src: new enclosed_mass_fraction method (pure jnp, differentiable in r and r_h) - unit tests: M(<r_h)=0.5, M(<a)=1/(2sqrt2), M(<0)=0, M(<inf)=1, monotone, vectorized, JIT, grad in r (>0) and r_h (flows through a) - grad-audit registry: N(r) model case (consistent, ratio 1.0000000, tol=1e-5) + pinned data-count case (known_blocked, AD=0 finite, FD=-5e3 bin-crossing) - website: binned-kinematic Fisher path table (sigma/beta + N(r) split), case count 53->57, changelog Tier-3 addition row - regenerated grad_audit_results.json (57 rows, 0 hazards; gate 58 passed) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…nsolidation (Task 4.1; no deletions — awaiting CHECKPOINT 4a sign-off)
…e+7 new cases; keep borderline; fast-per-file/full-per-batch)
…min, Kepler a/M0/from_state, Michie density) + fix H6 boundary probe + dedupe (Task 4.2a) Grad-audit gate 65 passed. All measured-first; H6 boundary now a standalone case running the real u=1e-12 probe (eps=1e-12 below the live |AD|), duplicate edge removed. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…it (i=0.3,Omega=0.4,omega=0.5) — runs the general inversion path, value unchanged (review note)
…rad-audit registry (audit T6 consolidation; registry is SoT) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ad-audit registry (audit T6 consolidation; registry is SoT) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… AD-FD assertion into grad-audit registry (audit T6 consolidation; registry is SoT) test_profile_gradients.py: removed plummer r_h, eff gamma, king r_c, king W0 sampler FD tests; KEPT test_eff_grad_a with a TODO (EFFProfile.sample_positions(a) is not directly owned by the registry — safety interlock against coverage loss). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…-FD assertion into grad-audit registry (audit T6 consolidation; registry is SoT) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…sertion into grad-audit registry (audit T6 consolidation; registry is SoT) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…D-FD assertion into grad-audit registry (audit T6 consolidation; registry is SoT) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…on into grad-audit registry (audit T6 consolidation; registry is SoT) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…-FD assertion into grad-audit registry (audit T6 consolidation; registry is SoT) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…D assertions into grad-audit registry (audit T6 consolidation; registry is SoT) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ion into grad-audit registry (audit T6 consolidation; registry is SoT) Split test_jit_and_grad_safe -> test_jit_safe: dropped the d|sep|/da grad smoke (owned by the registry resolve_binary_components Case), kept the JIT finite assertion. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…FFProfile.sample_positions(a) cases, migrate test_eff_grad_a, fix Michie W0 pointer Review found the W0 density channel (closed-form density() path, distinct from the sampler) and EFF positions-in-a were uncovered. Both added (measured-first, clean), closing the migrate interlock's one skip. Registry now 66 cases.
…t smoke test(s) — registry FD-covers (audit T6; registry is SoT) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… test(s) — registry FD-covers (audit T6; registry is SoT) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…oke test(s) — registry FD-covers (audit T6; registry is SoT) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…e test(s) — registry FD-covers (audit T6; registry is SoT) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…e test(s) — registry FD-covers (audit T6; registry is SoT) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ke test(s) — registry FD-covers (audit T6; registry is SoT) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…st(s) — registry FD-covers (audit T6; registry is SoT) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…moke test(s) — registry FD-covers (audit T6; registry is SoT) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…t(s) — registry FD-covers (audit T6; registry is SoT) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…yGradients + clarify sample/ppf reparam pointer (4.2c review)
…site 66-case reconcile + consolidation narrative, STATUS (Task 4.3) - scripts/audit_gradients.py: add --plots flag rendering grad_audit_ratio.png (per-case |AD/FD-1| residual, log axis, tol bands, known-limitation annotated) + grad_audit_summary.png (counts by direction × status), straight from the single-source JSON. Guarded matplotlib (viz extra) import. - website differentiability-audit.md reconciled 57→66 cases: 9 Tier-4 rows added with real measured ratios (King/EFF DF scale params, PowerLawIMF.ppf(m_min), Kepler to_state(a,M0)/from_state, Michie density observable, H6 standalone), new "Tier 4 — registry as the release gradient-gate" section (24 migrated / 15 deleted T6 smoke / 2 gaps surfaced / H6 fix), both figures embedded, changelog row. myst build 161 pages, 0 warnings. - STATUS: differentiability gradient-audit arc complete (Tiers 0–4); registry is the release gradient-gate (66 cases, 62 clean / 4 known-limitation, 0 hazards, 2 fixed); awaits final PR / CHECKPOINT-4b. audit_gradients.py exits 0 (66 cases, 0 hazards); test_grad_audit 67 passed. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ion (Michie density = 1 dual-param Case) (4.3 review nit)
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: fe8b8a69b2
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| f"the anisotropy is too strong (no finite tidal radius). Increase ra / xi_max." | ||
| ) | ||
| return xi_grid, psi_grid, rho_j_grid | ||
| return xi_grid, psi_grid, psi_raw, rho_j_grid |
There was a problem hiding this comment.
When this solver now returns four values, the DF-table validation script still unpacks three: scripts/validate_df_tables.py uses _solve(...) as xi_q, psi_q, _ = ... at lines 106-107 and directly calls _, psi, _ = solve_multicomponent_limepy(...) at line 146. Running that validation path after this API change raises ValueError: too many values to unpack, so the table-accuracy/gradient validation script can no longer run until those unpack sites are updated or a compatibility wrapper is kept.
Useful? React with 👍 / 👎.
drannarosen
added a commit
that referenced
this pull request
Jun 14, 2026
…CHECKPOINT 4b closed) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
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
1 participant
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.
Why
progenax's value proposition is that a differentiable forward model turns cluster inference into information geometry: every IC is$F = J^\mathsf{T}J$ is computable. That has a sharp failure mode — a silently zeroed or wrong column of $J$ makes a parameter look infinitely well-constrained (a confidently wrong uncertainty) with no error raised. A
jax.grad-able, so the Fisher matrixNaNcrashes loudly; a zeroed gradient does not. Gradient integrity is Fisher integrity.This arc builds a systematic AD-vs-finite-difference audit over every public sampling/physics entry point, fixes the hazards it finds, and makes the audit registry the release gradient-gate.
What's in here
A 66-case AD-vs-FD harness (
tests/validation/grad_audit/) over both protected directions (params → ICandparams → binned summary), with a computed (never hand-set) status: 62 clean, 4 pinned known-limitations, 0 hazards. Numbers are regenerated byscripts/audit_gradients.py→validation/data/grad_audit_results.json; the living tracker isdocs/website/50-validation/differentiability-audit.md.Two real hazards found and fixed — both in sites previously believed correct-by-design, both gradient-only (forward physics bit-identical):
r_twas a silent zero (jnp.maximum(\psi,0)clamp zeroed the crossing-node gradient before the shared_find_tidal_radius. Fixed by feeding unclampedstop_gradientomitted ~27% of the gradient at the default softmin temperature. Removed; now FD-consistent.The Fisher path itself audited — binned$\sigma(r)$ / $\beta(r)$ channels are clean with a mutation check proving teeth ($N(r)$ Fisher channel, while the binned data count is pinned as correctly non-differentiable (frozen data is out of scope; the gradient lives in the model).
stop_gradient(pos,vel)collapses AD to 0). A new publicPlummerProfile.enclosed_mass_fractionsupplies the model-sideThe registry is now the release gradient-gate — ~24 duplicate AD-vs-FD assertions migrated in, ~15 finite-only
isfinite(grad)smoke tests deleted (a silent-zero passes them), coverage preserved by a safety interlock + per-batch full gates (which surfaced 2 real gaps that were then filled). Scattered tests carry pointer comments to the registry.Verification
scripts/audit_gradients.py: exit 0, 0 hazardsPlummerProfile.enclosed_mass_fraction: 12 unit tests; King c(W₀) vs Table II unchanged (max|Δlog₁₀c|=0.0023)Completion doc:
.claude-work/TASK_differentiability-audit_COMPLETE.md.🤖 Generated with Claude Code