IMRPhenomHM SXS:BBH:2158 aligned-spin holdout — green receipt
The prospective protocol
and runner were committed as
4ec1383 before retrieving or decoding any 2158 strain. The case,
conditioning, optimization, and thresholds were not changed after seeing
the data. The frozen two-resolution screen and both native frontend model
screens pass. This is one case-specific physical result, not full-domain
qualification or a repair of the earlier red SXS cases. The
complete machine-readable metrics
retain all source hashes, alignment candidates, optimization seeds, norms,
and mode diagnostics.
Source and execution identity
The metadata-only filter and sort in the protocol selected
SXS:BBH:2158v3.0 from the
official SXS v3.0.0 catalog, SHA-256
132d38520b3562c8f254b3c9aaf8efd4678428f94dbbabfef4436bc50c1c87c6.
All six Lev4/Lev3 N2 strain files came from the official
CaltechDATA record.
Their byte sizes and published MD5 values matched the precommitted manifest
before decoding; their SHA-256 values are in the metrics. The prepared
fixture SHA-256 is
003806ce22416cd1ecf607c33471aaa1fa4ff93e9bf7a24ea2a3c626141bc0be;
the runner SHA-256 is
c1e32275d709d2b5a2020a37fd0ade4f263da6f6eff22478a6e5cae42d3bd431.
Raw and derived strain remain outside Git on len under
/tmp/diffgw-hm-sxs-2158-20260927/data.
On len, from /tmp/diffgw-hm-source-audit-20260927/repo, the preparation
used PYTHONPATH=src and
/tmp/diffgw-hm-sxs-2265-20260927/.venv/bin/python
scripts/run_imrphenomhm_sxs_2158.py --data-dir
/tmp/diffgw-hm-sxs-2158-20260927/data --prepare-only --source-revision
4ec1383. The evaluation used the same runner and directory with
/tmp/diffgw-waveform-library-1315f6e/venv-latest/bin/python and
--evaluate-only --source-revision 4ec1383, with JAX_ENABLE_X64=1 and
JAX_PLATFORMS=cpu. The preparation environment had Python 3.12.3,
NumPy 2.5.3, SciPy 1.18.1, sxs 2026.0.6, and SymPy 1.14.0. The
evaluation environment had Python 3.12.3, NumPy 2.3.5, SciPy 1.16.3,
PyCBC 2.11.0, Torch 2.14.0, JAX/JAXlib 0.11.2, and SymPy 1.14.0.
The 30–1024 Hz PSD SHA-256 was
0d20d2092bfaeb6073b0d656f4a3eb495bc3adbeb350ef279420cb4ffd0d67ea.
The official inertial, spin-weight -2 strain contained every signed-
m mode through ell=5, including m=0. The common interval was
1410.0–6820.165009984545 M, with 21,830 uniform samples. The
Lev4 |h_22| peak at 6672.998518942362 M set time zero. The
mode-basis FFT reconstructed all direct Riemann spectra to relative
maximum error below 3.58e-13; the synthetic time-shift and all direct
inner-product checks passed.
Frozen gates
One Lev3-to-Lev4 two-view plus/cross alignment selected orbital phase
6.225823345700972 rad and arrival shift -0.0009765625 s.
The combined network match was 0.999996513679 and norm ratio
1.000530895011. That same alignment gave:
Fixed (iota,psi_s) |
Lev3/Lev4 match (floor 0.995) | Norm ratio (0.95–1.05) |
|---|---|---|
(0.7,0.0) |
0.999996706940 |
1.000412490033 |
(1.2,0.4) |
0.999996028926 |
1.000877198583 |
With catalog mass ratio 3.0001323450539252, fixed aligned spins
(0.49997718539303,0.5000516952582452), total mass 100 M_sun,
and no intrinsic or amplitude fitting, the independently optimized
Torch and JAX templates gave:
| Frontend | View | Model/Lev4 match (floor 0.980) | Norm ratio (0.85–1.15) | Four-seed match spread |
|---|---|---|---|---|
| Torch | (0.7,0.0) |
0.994088357164 |
1.003313403120 |
5.16e-11 |
| JAX | (0.7,0.0) |
0.994088357164 |
1.003313403120 |
5.16e-11 |
| Torch | (1.2,0.4) |
0.986499006522 |
1.074226637473 |
3.14e-11 |
| JAX | (1.2,0.4) |
0.986499006522 |
1.074226637473 |
3.14e-11 |
All optimizer seeds met the precommitted 1e-5 agreement rule. The full
NR signal minus its six template-mode pairs had weighted residual norm
fractions 0.01647 and 0.04887 in the two views; that is a diagnostic,
not a fitted correction. The weakest model match is 0.98650, above the
frozen floor. This fixed-view result is not directly comparable to London's
polarization/orbital-phase population averages. The existing
2265,
3631,
3619, and
3630 failures
remain unchanged. This green result supports only the stated aligned-spin
case, mass, band, PSD, two views, finite resolutions, and fixed native
six-mode option; it does not establish all-sky or model-family accuracy.
Post-result diagnostic and case closeout
The maintained scripts/diagnose_imrphenomhm_sxs_2158.py reproduces this
audit from the frozen inputs and best-fit Torch alignment. It checks source,
fixture, runner, metrics, and PSD hashes before evaluation. Neither mode nor
frequency-band diagnostics fit a new phase, time, amplitude, or intrinsic
parameter. This report retains the selected post-result findings below;
the full two-view, six-pair diagnostic arrays are excluded from the
distributed repository. Their JSON SHA-256 is
8234de50443d167c6aa3885811200ba547c73a1e4f162449ab9ad75123b19120,
and the diagnostic runner SHA-256 is
b9114f5630756a8e404e54af9903e487f9e15eb20d925f59e50fa3b6ffeb9b0e.
The input metrics SHA-256 is
a1497fbbda34d7cb73d14dfde0113b4dd0c4acc513abfbe2ded4bb4716c79526;
fixture, frozen runner, PSD and evaluation package identities are recorded
above, with LALSuite 7.26.15 added for the binary comparison. These are
post-result diagnostics, not new prospective gates. The frozen runner and
original metrics remain unchanged.
For the weaker (1.2,0.4) view, replacing the native modes with modes from
the public LALSuite 7.26.15 binary, using the previously fixed mode
adapter and the same projection and alignment, gives:
| Candidate / reference | Match | Norm ratio |
|---|---|---|
| Native HM / full NR | 0.986499006522 | 1.074226637473 |
| LAL HM / full NR | 0.986498963403 | 1.074226679116 |
| Native HM / six-pair NR | 0.987735823627 | 1.070556561918 |
| Six-pair NR / full NR | 0.998824938077 | 1.003428193975 |
Across the six complex carriers, native/LAL relative L2 is at most
1.985e-5; relative error on bins above 1% of each LAL mode's peak is at
most 1.252e-5. Torch/JAX relative L2 is below 3.404e-13. The LAL row
uses the independent reference projector, not LAL's public polarization
API; that separate API is covered by the maintained
polarization holdout.
No external implementation source was inspected, and no production
coefficient was fitted to binary or NR output.
This near-identical binary result points away from a native frontend discrepancy as the explanation for this match. Restricting NR to the model's six pairs still leaves substantial disagreement. Missing modes alone therefore do not explain the result. Shared model approximations and comparison conventions remain possible sources; this diagnostic does not uniquely identify a physical cause.
For the full detector spectrum, the exact pointwise identity
splits squared residual into amplitude and phase terms, integrated with
the frozen PSD. Normalizing by full NR power gives total 0.03451584736,
amplitude 0.02308207006 (66.87%), and phase 0.01143377730 (33.13%).
This is squared complex residual, not mismatch 1-match; the norm
offset contributes to it. Amplitude/phase here describe the coherently
summed detector signal, not the source of individual mode errors.
| Frequency bins (Hz) | Share of total squared residual |
|---|---|
| 30–59 | 12.44% |
| 60–99 | 19.47% |
| 100–199 | 43.94% |
| 200–399 | 20.41% |
| 400–1024 | 3.75% |
These disjoint bands partition the residual; their boundaries are
diagnostic choices, not waveform transition frequencies. Projected
mode-pair diagnostics expose a (4,4) model/NR norm ratio of 1.35083
and substantial (3,2) phase disagreement under the shared alignment.
They are not intrinsic amplitude measurements or independently optimized
mode matches. Mode errors interfere coherently, so their powers cannot
be added or interpreted as unique causal shares. In particular, the
shared alignment compromises across modes.
Likewise, the diagnostic's nr_power_fraction_of_full is a per-pair/full-signal
power ratio: destructive interference permits values above one, and these
ratios do not sum to one.
London et al., arXiv:1708.00404v4 describe approximate quadrupole mappings and omitted spherical–spheroidal mixing. These provide relevant model limitations, not proof that a particular approximation causes this case's residual. Improving the physics would require a separately specified, independently sourced model change and new held-out validation, rather than a correction tuned to this result.
Disposition: the 2158 case and its diagnostic are complete. Its frozen gates pass, with no change to thresholds or production waveforms. Whole-HM qualification remains open for the inherited fitted-input provenance and physical coverage listed in the family ledger; the earlier red cases retain their original status.
Reproduce the diagnostic
On len, from /tmp/diffgw-hm-source-audit-20260927/repo:
PYTHONPATH=src JAX_ENABLE_X64=1 JAX_PLATFORMS=cpu \
/tmp/diffgw-waveform-library-1315f6e/venv-latest/bin/python \
-m scripts.diagnose_imrphenomhm_sxs_2158 \
--data-dir /tmp/diffgw-hm-sxs-2158-20260927/data \
--output /tmp/diffgw-hm-sxs-2158-20260927/diagnostic.json
The data directory can instead be regenerated with the frozen preparation
command above. Nine maintained analytical tests in
tests/test_hm_residual_budget.py pass on len, covering known amplitude
and phase offsets, frequency partitioning, coherent cancellation,
zero model support above a mode cutoff, and invalid PSD rejection.