IMRPhenomHM SXS:BBH:3630 projected-match holdout: frozen protocol

Commit this document before retrieving or decoding either 3630 strain. This is one new, independent physical holdout. Do not change its sources, conditioning, optimizers, or gates after seeing the waveform. The earlier 3619 projected holdout remains red under its frozen, no-orbital-retuning resolution screen, even though a clearly post-result orbital-alignment diagnostic gave high NR agreement. The 2265 and 3631 six-mode gates also stay red.

Pre-data case selection and source identity

The official sxs.load("dataframe", tag="3.0.0") metadata were filtered for nondeprecated runs beginning after 2018-05-01, 5<=q<=8, both reference spin magnitudes <0.001, and reference eccentricity <0.005. Only 3619 and 3630 met these criteria; 3619 has already been inspected. The remaining SXS:BBH:3630v3.0 has q 6.000520906220441, spin magnitudes 5.9110497766760964e-05 and 2.4576284716767174e-05, reference eccentricity 0.000434, 21.24783108430796 orbits, initial orbital frequency 0.019315320434/M, and Lev3 reference time 778.0 M. The run began 2022-05-11. Selection used metadata only, with no 3630 strain outcome. The initial (2,2) frequency at the frozen 100 M_sun mass is about 12.5 Hz, below the 30 Hz metric floor. The 2022 run date is later than the 2018 model paper, but is not proof that no later model development used it.

The official SXS:BBH:3630v3.0 DOI resolves to CaltechDATA record 28pv5-4wj69. The untouched Lev3/Lev2 N2 strain, sidecar, and metadata files below are the only source files. Retrieve them from the record's /api/records/28pv5-4wj69/files/<key>/content endpoints. Check byte size and the published MD5 before decoding and record SHA-256. Keep raw and derived arrays outside Git.

key	bytes	official MD5
Lev3:Strain_N2.h5	3040680	f415773285072330601c0c39d4d3f0b5
Lev3:Strain_N2.json	2227	67aeb7928db5fdab25bb008465bdbeee
Lev3:metadata.json	3364	6629f45f5a1fd26c0e2b284c18ae57ea
Lev2:Strain_N2.h5	3276474	e29f4cafff41290630179327824ac7ba
Lev2:Strain_N2.json	2227	c136be72e617904f2f1895321406cc5d
Lev2:metadata.json	3367	a4846f6ab9c7a265476eed5245e40b4b

Frozen NR and model signals

Use the official sxs decoder, requested version 2026.0.6, and record exact dependency versions. Require decoded inertial-frame spin--2 strain h=h_+-ih_×, mass and radius scaled out, and every (ell,m) for 2<=ell<=5, -ell<=m<=ell. Use all these NR modes, including m=0; the native IMRPhenomHM template uses its six published positive-m modes with symmetry partners. Check sidecars and decoded conventions, record CoM/memory modifications, and never undo them. Reject missing modes, nonfinite data, non-inertial frame, or metadata inconsistent with q≈6 and negligible spins.

Set detector-frame total mass 100 M_sun, native component masses 600/7 and 100/7 M_sun, zero model spins, distance 500 Mpc, f_ref=30 Hz, and native phic=tc=phi0=0. 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 NR strain by GM/(c^2 D). For both resolutions use the common interval from the largest of Lev3 reference time 778.0 M and file starts to the smallest file end. Cubic-spline interpolate each complex mode's real and imaginary parts to one 8192 Hz uniform grid. Set time zero at Lev3 maximum |h_22|; apply a half-cosine rise over the first 500 M and fall over the last 100 M to every mode. Reject an interval too short for both ramps. Fourier transform with exp(-2πift) at precisely integer f=30,...,1024 Hz; use the direct Riemann sum for full NR projections. For the resolution-screen orbital basis, a zero-padded FFT on the same samples is permitted only if its sum at zero orbital phase reconstructs the direct full NR spectra to relative maximum error <=1e-10 for plus and cross at each inclination. No fitted mode-dependent factor or cut is allowed.

Project NR with spin-weight -2 spherical harmonics at source inclinations and polarizations (iota,psi_s)=(0.7,0.0),(1.2,0.4) radians and zero source azimuth. For the model, use the independent six-mode scalar projection in diffgw_reference.imrphenomhm_projection. At template orbital phase φ, multiply each native positive-frequency carrier (ell,m) by exp(-imφ) before projection. The one-detector model/NR spectra are cos(2ψ) h_+ + sin(2ψ) h_×, with the frozen source ψ_s for NR and optimized template ψ. The NR resolution comparison instead uses both unrotated plus/cross channels at each inclination, as specified below.

Frozen metrics, optimization, and pass conditions

Generate PyCBC's documented aLIGOZeroDetHighPower PSD with length 1026, spacing 1 Hz, and low cutoff 30 Hz; require finite positive samples at every comparison bin and record the PSD SHA-256. This is a validation-time dependency, not copied LAL source or array. For one detector define (a|b)=4 Re Σ_(f=30)^1024 a*(f)b(f)/S_n(f) and the normalized overlap. For the Lev2–Lev3 screen use the two-detector network sum of this inner product over plus and cross, with both channel norms summed before normalization. This follows the SXS Catalog 2019 Appendix B alignment: optimize only a single common orbital azimuth applied to all Lev2 modes and one arrival shift. Do not optimize a relative polarization rotation for this NR screen. The paper's separate Planck taper, sphere sampling, and adaptive band are not claimed here; the conditioning and band above remain fixed for this one-case test.

For both model and NR screen, arrival shifts are k/8192 s for integer k=-819,...,819, evaluated at each trial angle by a zero-padded 8192-point real inverse FFT of the weighted cross spectrum. Verify the FFT sign and scale on a synthetic exact shift and by direct inner-product evaluation at reported optima. On the model side, optimize template φ∈[0,2π) and ψ∈[0,π) with SciPy differential_evolution, seeds (17,23,41,67), popsize=12, maxiter=120, tol=1e-9, polish=True, one worker. Take the best of four; all final matches must be within 1e-5, else optimization is infeasible. On the NR resolution side, evaluate φ_j=2πj/2048 for j=0,...,2047, select the eight highest distinct periodic local maxima, then refine each within ±2π/2048 using bounded scalar minimization with xatol=1e-10, maxiter=500. Take the best of the 2048 grid points and eight refinements; record all candidate scores, phase, arrival shift, and norm ratio. The same phase rotates all 2<=ell<=5 Lev2 modes by exp(-imφ) in the time domain before projecting plus/cross, with no polarization change.

At both angles the Lev2–Lev3 network match must be >=0.995. For each native Torch and JAX frontend, at both angles require optimized one-detector model/Lev3 match >=0.980 and PSD-weighted template/Lev3 norm ratio in [0.85,1.15]. Both backends must pass independently; their parity alone is not physical qualification. A pass is limited to this case, mass, band, PSD, angles, modes, and method and cannot override prior red holdouts. Report red or infeasible gates without relaxation or production coefficient edits.