IMRPhenomHM SXS:BBH:3144: prospective physical-validation protocol

Frozen before strain retrieval or model scoring. The only case-selection input inspected was the official SXS v3.0.0 catalog metadata described below. Do not revise the case, waveform conditioning, optimization, or gates after decoding its strain. Record failures and infeasible comparisons as such. This is a new test; the existing 2265, 3631, 3619, and 3630 gates and receipts retain their original status. The selection, conditioning, and gates below were fixed before examining any 3144 strain or score.

Literature target and limits

London et al. (2018) compared IMRPhenomHM against full numerical-relativity signals through ell=5 using the Advanced LIGO design noise curve. Their physical comparison used a 30 Hz floor and, for the figure discussed here, 100 M_sun; intrinsic parameters and source inclination were fixed while template arrival time, polarization, and initial orbital phase were optimized. They reported typically greater than 0.99 match for nonspinning mass ratios at most eight in their 19-waveform set, with signal-strength-weighted angular summaries. The present two-view, one-case gate is deliberately narrower and does not reproduce that population or angular average. It tests an independently selected later SXS simulation with a comparable PSD, frequency floor, mass, and extrinsic optimization.

SXS Catalog 2019 Appendix B motivates comparing h=h_+ - i h_× with one common arrival-time and orbital-azimuth adjustment across NR modes, using plus and cross as a two-detector network. The present fixed taper, fixed band, and two views are specified below; they do not claim to reproduce that paper's Planck taper, adaptive frequency band, or sky sampling. The SXS Catalog 3 paper describes v3 waveform conventions and its separate sphere-averaged resolution analyses. The resolution screen below is an operational two-view network test, not a claim of catalog-style convergence or a NR error bound.

Prospective metadata-only case selection and source identity

Use the official sxscatalogdata release v3.0.0/simulations.json, associated with the SXS v3.0.0 release. The locally cached JSON is 32,618,880 bytes and has SHA-256 132d38520b3562c8f254b3c9aaf8efd4678428f94dbbabfef4436bc50c1c87c6. Its 4,170 catalog entries contain metadata and file manifests, not strain arrays. Preserve this hash and the exact filtering rule in the receipt.

For each key beginning SXS:BBH:, retain entries with no deprecated keyword, reference_mass_ratio in 3 <= q <= 8, Euclidean magnitudes of both reference_dimensionless_spin1 and reference_dimensionless_spin2 vectors below 0.001, reference_eccentricity < 0.001, date_run_earliest >= 2018-01-18T00:00:00, number_of_orbits >= 20, a finite number_of_orbits_from_reference_time, and at least two available Lev*:Strain_N2.h5 files in the manifest. Explicitly remove the four previously examined IDs 2265, 3631, 3619, and 3630. Sort the remaining entries by descending number_of_orbits_from_reference_time, then by ascending numeric SXS ID, and take the first. Longer post-relaxation data improve the chance of a complete 30 Hz comparison at 100 M_sun; the date cut avoids simulations run before the cited London et al. revision, while q >= 3 emphasizes systems where higher modes matter. This selects SXS:BBH:3144v3.0, with 29.592976667962976 post-reference orbits. The remaining eligible IDs in sort order are 2374, 2499, 2164, 2498. The chronological choice does not prove independence from every later model-development activity.

The official SXS:BBH:3144v3.0 DOI resolves to CaltechDATA record 1g62w-w0b49. The catalog gives reference q=4.499621846563142, eccentricity 0.0002219, reference spin vectors (4.021819846540122e-7, 1.7207127249235393e-7, 2.9715550036843704e-5) and (8.852292173961585e-8, 2.673455377441355e-8, 1.6474344378490387e-5), reference time 1580.5 M, 33.26561261546802 total orbits, and run start 2019-11-30T19:56:54. The initial orbital frequency 0.014212721029/M implies an approximate initial (2,2) frequency of 9.2 Hz at 100 M_sun, below the metric floor. The manifest lists only Lev2 and Lev3 strain; therefore both available resolutions are mandatory for the resolution screen.

Retrieve only the following six files from the record's /api/records/1g62w-w0b49/files/<URL-encoded-key>/content endpoints in the execution phase. Verify published byte size and MD5 before decoding, record local SHA-256, and keep raw and derived strain outside Git. No such files had been retrieved when this protocol was written.

Catalog key	Bytes	Published MD5
Lev3:Strain_N2.h5	3948147	78a1bed032fa8ae31a167d4fa30203fb
Lev3:Strain_N2.json	2229	73edba3df5fbcef49acc0f6a8c5c5bc9
Lev3:metadata.json	3373	fbdddb96126b46b4a1881ba84ced2e8d
Lev2:Strain_N2.h5	4149375	0324ae1236564a44f4a998364b829d4f
Lev2:Strain_N2.json	2231	e9551b14ff83ecd735dce7524091cc1b
Lev2:metadata.json	3374	21415531730d1b93b5fd7da69f87b584

Fixed signals and conditioning

Decode the official inertial-frame, extrapolated N2, spin-weight -2 strain with the public sxs loader (request 2026.0.6; record the exact installed package and dependency versions). Require every NR mode 2 <= ell <= 5, -ell <= m <= ell, including m=0. Reject missing modes, nonfinite arrays, an inconsistent frame, or sidecar/metadata contradictions. Record v3 center-of-mass and memory modifications from the sidecars; do not undo them. The native IMRPhenomHM template contains its six published positive-m modes (2,2), (3,3), (4,4), (2,1), (3,2), (4,3) and symmetry partners. Report that missing NR modes, especially m=0 and the remaining ell<=5 modes, are a possible contribution to model mismatch; do not remove them from the primary comparison.

Fix detector-frame total mass to 100 M_sun, component masses to 100q/(1+q) and 100/(1+q) M_sun with the catalog reference q above, model spins to zero, luminosity distance to 500 Mpc, f_ref=30 Hz, and native phic=tc=phi0=0. The tiny nonzero NR reference spins are treated as numerical residuals, not fitted model spins. Use nominal GM_sun=1.3271244e20 m^3 s^-2, exact c=299792458 m/s, and 1 pc=3.0856775814913673e16 m; scale NR time by GM/c^3 and strain by GM/(c^2 D). Record the conversion and any overall strain normalization provided by the loader. Reject a convention ambiguity rather than fitting an amplitude scale.

Use the common time interval starting at the maximum of the Lev2/Lev3 reference times and file starts and ending at the minimum file end. Cubic spline each complex mode's real and imaginary parts onto one uniform 8192 Hz time grid. Set time zero at the Lev3 maximum of |h_22|. Apply the same half-cosine rise across the first 500 M and fall across the last 100 M to every mode of both resolutions; reject an interval unable to support both ramps. Form the h=h_+-ih_× projection with spin-weight -2 harmonics and Fourier convention exp(-2πift), using a direct Riemann sum at integer f=30,...,1024 Hz for the reported spectra. Fix source azimuth to zero and the two source views to (iota,psi_s)=(0.7,0.0) and (1.2,0.4) radians. For the physical model/NR comparison, form each scalar detector spectrum as cos(2ψ)h_+ + sin(2ψ)h_×, with ψ_s fixed for NR and template ψ optimized. Apply template orbital phase φ to each positive-frequency carrier as exp(-imφ) before projection. Do not optimize masses, spins, inclinations, distance, or any mode-specific phase or amplitude.

PSD, common NR alignment, and model optimization

Generate PyCBC's documented aLIGOZeroDetHighPower PSD with length 1026, delta_f=1 Hz, and low-frequency cutoff 30 Hz. Require finite positive samples at all comparison bins and record PSD SHA-256 and library version. For any one real detector channel define (a|b)=4 Re Σ[f=30..1024] a*(f)b(f)/S_n(f) and match (a|b)/sqrt((a|a)(b|b)). A plus/cross network uses the sum of channel inner products before normalization; a two-view network additionally sums both views. Norm ratio always means sqrt((template|template) / (reference|reference)) with the same channel set and PSD.

One global Lev2-to-Lev3 alignment is required for every NR resolution check. Apply exp(-imφ_NR) to all Lev2 (ell,m) modes and one common arrival shift τ_NR. Choose this pair by maximizing the normalized plus/cross two-view network match of full ell=2..5 Lev2 against Lev3, with both view inclinations fixed and their common, unrotated plus and cross channels. The polarization basis cancels from a plus/cross network, so neither source ψ_s nor any new polarization rotation is optimized here. Search φ_j=2πj/2048, j=0,...,2047; at each trial find the best arrival shift τ=k/8192 s, integer k=-819,...,819, via a zero-padded 8192-point real inverse FFT of the weighted cross spectrum. Take the eight highest distinct periodic local phase maxima and refine each within ±2π/2048 by bounded scalar minimization with xatol=1e-10, maxiter=500. Retain the best grid/refined result. Verify FFT sign and normalization on a synthetic exact shift and with direct inner products at the optimum; an FFT-derived spectrum must reconstruct the direct spectrum at zero phase within relative maximum error 1e-10 per plus/cross channel and view. Record candidate scores and optimum. Freeze that same (φ_NR,τ_NR) for the combined and each individual view's Lev2/Lev3 match and norm ratio and for any per-mode diagnostic. There is no per-view, per-polarization, or per-mode retuning.

At each fixed source view, optimize the native Torch and JAX model frontends independently against Lev3 in the one-detector channel over template arrival shift k/8192 s in the same range, orbital φ∈[0,2π), and polarization ψ∈[0,π). Use SciPy differential_evolution for the two angles with seeds (17,23,41,67), popsize=12, maxiter=120, tol=1e-9, polish=True, one worker. Take the best of four and require their final matches to agree within 1e-5; otherwise label optimization infeasible. At each trial angle maximize over the discrete arrival shifts using the same FFT rule. Check reported optima by direct inner products. Apply no global amplitude fit. Record all optimized parameters, signal norms, full mode list, model/backend versions, and numeric conditioning diagnostics in the result receipt.

Precommitted gates and interpretation

Both resolution views must have match >=0.995 and Lev2/Lev3 network norm ratio in [0.95,1.05] under the one common alignment above; report the combined two-view values as diagnostics. A 0.5% overlap deficit and 5% weighted-amplitude interval are predeclared limits for using these two finite-resolution waveforms as a reference, not a published estimate of their continuum error. If this screen fails or is infeasible, no physical model qualification is awarded, even if model scores are high.

For each Torch and JAX native frontend, at both views require optimized model/Lev3 one-detector match >=0.980 and weighted norm ratio in [0.85,1.15]. London et al.'s reported typical >0.99 nonspinning matches motivate a high but deliberately weaker one-case 0.98 floor; the 15% amplitude interval is a conservative, explicit operational allowance for a higher-mode model, not a theorem derived from their mode-error plots. These thresholds are fixed before any selected strain is decoded. The older gates remain unchanged and cannot be rescued by this result. No failed gate may be relaxed after inspection.

A pass supports only this q≈4.5, nearly nonspinning SXS case at 100 M_sun, these two views, band, PSD, conditioning, and six-mode template against the full available NR signal. It does not establish all-sky performance, the 19-case result of London et al., spin or mass extrapolation, continuum NR accuracy, or agreement with another model implementation. Report the omitted-mode contribution, weak-mode diagnostics, and any convention or optimization limitations as diagnostics without redefining the gates.