Skip to content

IMRPhenomHM SXS:BBH:3144 prospective qualification — green receipt

The precommitted one-case qualification passes. Both available SXS resolutions pass the two-view plus/cross screen under one shared orbital phase and arrival shift. Native Torch and JAX IMRPhenomHM then pass the separate match and norm-ratio gates at both fixed source views. The protocol was committed as 3b3d50d before any 3144 strain was retrieved or decoded. Its conditioning, optimizer, modes, and thresholds were unchanged after inspection. The prior 2265, 3631, 3619, and 3630 red results and their gates remain unchanged. The complete numeric metrics include every alignment candidate, optimizer seed, norm, mode, and diagnostic.

Source identity and conditioning

The metadata-only selection used the official SXS v3.0.0 catalog JSON (32,618,880 bytes, SHA-256 132d38520b3562c8f254b3c9aaf8efd4678428f94dbbabfef4436bc50c1c87c6; 4,170 entries), with the exact filter and sort in the protocol. The first eligible case after excluding the four earlier cases was SXS:BBH:3144v3.0; the remaining eligible IDs in order were 2374, 2499, 2164, and 2498. Its reference mass ratio is 4.499621846563142, with negligible but nonzero NR spin vectors. All six files came from the official SXS:BBH:3144v3.0 record. Each published byte size and MD5 in the protocol was verified before decoding. The local SHA-256 values, in protocol order, are:

Official file SHA-256
Lev3:Strain_N2.h5 59ce19bbed32e257c0bcb307ada7adf5b75c297ffc8d3ae4cfe6a7762eb4440a
Lev3:Strain_N2.json d7a74026b07f0dff57be6f9a437e52f7610847773342fcb5fd8531c4da3da95f
Lev3:metadata.json 16360fe120d177b1daefba42387d666a552aa03332e0eebd2a4c9b2140ba303c
Lev2:Strain_N2.h5 2c1dac19a419aceda36af366516a9807b6f1e2880da77f10ecb227d6359b9787
Lev2:Strain_N2.json f0e31fcfca7ab124f8792aca295dfa5b50375979d6119d64adc2edee71b99d1d
Lev2:metadata.json 9aa7fdac4d35f6b52066e8bdc3aa23a267ab36c9014ab57958ba34a660d9106d

The public SXS loader reported inertial, spin-weight -2, h strain with m and radius scaled out, and every signed-m mode for 2≤ell≤5, including m=0. The sidecars record extrapolation, center-of-mass translation and boost, memory addition, and truncation; no modification was reversed. Memory integration starts at 1580.5 M for Lev3 and 692 M for Lev2. The common interval was 1580.5–10092.011834583338 M; the 8192 Hz grid contained 34,344 samples and its Lev3 |h_22| peak set time zero at 9933.490303373206 M. The nominal time conversion was 0.0004925490947641267 s/M; the loader's scaled-out strain was multiplied by 9.570831683175683e-21 for 100 M_sun at 500 Mpc. No amplitude scale was fitted. The external fixture SHA-256 is cf127c18a2211f0994ba0a03d406c360b8230ef27609920b5c14d11f49745cf2.

The direct Riemann plus/cross spectra and independently computed mode Fourier sums agree in every resolution/view/channel to relative maximum errors 3.46e-13–5.54e-13, below the frozen 1e-10 tolerance. A synthetic two-channel shift of +100/8192 s was recovered exactly, checking the inverse-FFT sign. Direct inner products at all reported optima agree with FFT matches within 3.33e-16. The separately generated PyCBC design PSD has comparison-bin SHA-256 0d20d2092bfaeb6073b0d656f4a3eb495bc3adbeb350ef279420cb4ffd0d67ea.

Frozen gates

The one Lev2-to-Lev3 alignment was phi_NR=5.743179471018169 rad, tau_NR=-0.014892578125 s, selected from the 2048-point grid and the eight highest of 14 periodic local phase maxima. Its combined two-view network match was 0.999939629340 with Lev2/Lev3 norm ratio 0.999601745405. This same phase and shift gave:

Fixed (iota,psi_s) Lev2/Lev3 match, floor 0.995 Norm ratio, bounds 0.95–1.05
(0.7,0.0) 0.999944177938 0.999708145957
(1.2,0.4) 0.999926931964 0.999303375647

Each native model comparison held mass, spins, distance, inclination, source polarization, and the six mode carriers fixed. It optimized only template orbital phase, polarization, and the shared discrete arrival shift. Each row below passed match ≥0.980 and norm ratio in [0.85,1.15]; all four seeds 17,23,41,67 passed the 1e-5 optimizer-agreement check.

Backend; view Match Norm ratio Orbital phase (rad) Polarization (rad) Shift (s) Seed spread
Torch; (0.7,0.0) 0.992168374557 0.995166487244 5.296249409445 1.022772935771 0.0062255859375 7.92e-12
JAX; (0.7,0.0) 0.992168374557 0.995166487214 5.296249407960 1.022772935466 0.0062255859375 7.92e-12
Torch; (1.2,0.4) 0.982689246447 1.128589480708 5.253069413264 1.287188873549 0.0062255859375 2.70e-11
JAX; (1.2,0.4) 0.982689246447 1.128589480708 5.253069413264 1.287188873549 0.0062255859375 2.70e-11

The full NR comparison used 32 modes, all 2≤ell≤5 and -ell≤m≤ell. The model used positive-m carriers (2,2), (2,1), (3,3), (3,2), (4,4), (4,3) and their symmetry partners. At the two views, the norm of the NR modes outside those pairs was respectively 0.026614 and 0.074496 of the full NR network norm. The six-pair NR subset, without new alignment, matched full NR at 0.999646 and 0.997231. Individual mode norms are coherent contributions and do not add in quadrature to the full norm; the full list is in the metrics JSON. For example, the (5,0) and (4,0) network norm fractions were 1.34e-5 and 8.83e-5. These diagnostics neither remove modes nor alter the primary gates.

This is an operational pass for one nearly nonspinning q≈4.5 case, 100 M_sun, the two stated views, 30–1024 Hz, this PSD and conditioning. The London et al. comparison motivates optimizing arrival time, template polarization, and orbital phase at fixed intrinsic parameters; its typical >0.99 population statement is not a result of this one-case test. The SXS 2019 Appendix B motivates a common time/azimuth shift and plus/cross network, while this fixed taper and band are protocol choices. The resolution screen does not establish continuum NR accuracy, and the pass does not imply all-sky, other-mass, spin, or independent-implementation performance.

Reproduction and checks

Host len; cwd /tmp/diffgw-hm-sxs-3144-20260927/repo; all raw and derived strain in /tmp/diffgw-hm-sxs-3144-20260927/data, outside Git. The final runner SHA-256 was 89856b9b94af8cf546620d2f07928016a35f8ddafa4277e9c479b0638d18b2c0. Evaluation requires this exact preparation-script hash as well as the raw-source and fixture hashes. The complete tracked metrics JSON SHA-256 is cfd71dc820ab3ff6210de1591de1ffb35baf4cc63ecfe8e309e6eea5fe751dee. The exact final commands from that cwd were:

/tmp/diffgw-hm-sxs-2265-20260927/.venv/bin/python \
  -m scripts.run_imrphenomhm_sxs_3144 \
  --data-dir /tmp/diffgw-hm-sxs-3144-20260927/data \
  --prepare-only --source-revision 3b3d50d
env PYTHONPATH=.:src JAX_PLATFORMS=cpu JAX_ENABLE_X64=1 \
  OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 \
  /tmp/diffgw-waveform-library-1315f6e/venv-latest/bin/python \
  -m scripts.run_imrphenomhm_sxs_3144 \
  --data-dir /tmp/diffgw-hm-sxs-3144-20260927/data \
  --evaluate-only --source-revision 3b3d50d

The preparation environment was Python 3.12.3, NumPy 2.5.3, SciPy 1.18.1, SXS 2026.0.6, and SymPy 1.14.0. The evaluation environment was 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. These were the latest available compatible versions on len for their respective stages. Preparation launcher PID 3155412 completed; log /tmp/diffgw-hm-sxs-3144-20260927/prepare-final.log. Evaluation launcher PID 3170169 completed with qualified: true; log /tmp/diffgw-hm-sxs-3144-20260927/evaluate-final.log. No job remains active, hence no next check or stop command applies; on a rerun, stop its active PID with kill <PID> on len. Data-free tests covering two-channel shift, direct network norm/inner product, signed-m time-domain azimuth rotation, and preparation-script hash enforcement passed 4/4. Ruff and python -m mkdocs build --strict passed on len; the latter log is /tmp/diffgw-hm-sxs-3144-20260927/mkdocs-strict.log.