Published-calibration IMRPhenomHM: prospective two-case NR protocol

Freeze this protocol, its two JSON manifests, runner and data-free tests in Git before retrieving or decoding either selected strain. The target is the exact published-calibration option below. Preserve every failure or infeasible result; do not replace a case, alter conditioning, tune coefficients or relax a gate after inspection. Earlier default-calibration receipts remain unchanged.

Target and literature

Use native Torch and JAX IMRPhenomHM with all of:

calibration="published"
hm_21_pn_variant="mishra_eq12b"
hm_32_pn_variant="mishra_eq12d"
hm_43_pn_variant="mishra_eq12g"
hm_mapping_transitions="london_2018"

The independent mathematical specification defines the printed Khan/Husa calibration and independently generated N61 QNM inputs. This is a named source-based variant; it does not claim to reproduce unrounded historical calibration or published population accuracy exactly.

London et al., Results/Fig. 3 motivates the full-NR signal through ell=5, Advanced LIGO design PSD, 100-solar-mass comparison, 30-Hz floor, fixed intrinsic parameters and inclination, and template time/orbital-phase/polarization optimization. SXS 2019 Appendix B motivates one common time/orbital rotation across NR modes in a plus/cross network. SXS Catalog 3 supplies versioned strain and conventions. The two views, taper and thresholds here are project choices, not published guarantees or a reproduction of a paper's population average.

Metadata-only selection

Use official sxscatalogdata v3.0.0 simulations.json, 32,618,880 bytes, SHA-256 132d38520b3562c8f254b3c9aaf8efd4678428f94dbbabfef4436bc50c1c87c6. Only catalog metadata and file manifests were inspected for this selection.

Reuse the exact nonspinning filter/ranking in the 3144 protocol and aligned-spin filter/ranking in the 2158 protocol. Additionally exclude every explicit SXS ID found in tracked repository text: 0001, 0002, 0008, 0030, 0056, 0063, 0166, 0169, 0182, 0185, 1166, 2158, 2265, 2331, 2483, 2484, 2485, 2486, 2489, 2491, 2707, 3144, 3619, 3630, 3631, 9997, 9998, 9999. Metadata-only candidate-list mentions are not prior strain inspection. The remaining nonspinning rank is 2374, 2499, 2164, 2498; the aligned-spin rank is 2163, 2162, 2155, 2502. Select the first in each.

Case	Reference q	Model chi1, chi2	Highest/lower resolution	Reference time M
SXS:BBH:2374v3.0	4.999912483619435	0, 0	Lev3 / Lev2	570.5
SXS:BBH:2163v3.0	3.0000780199258803	0.5999108664616073, 0.6000124147825587	Lev4 / Lev3	1379.5

2374 has eccentricity 3.26e-5 and 22.055961932088238 post-reference orbits; 2163 has eccentricity 0.0001502 and 23.187710232835745 post-reference orbits. Their run dates are 2018-09-21 and 2018-12-04. Their alternative identifiers are BFI:Rerun2013Catalog:0055 and BFI:RerunAlignedCatalog:0291. They are unseen local prospective holdouts, but these rerun identities mean independence from every historical model-development dataset is unproven. No inference of a uniformly validated mass-ratio/spin domain follows from them.

The machine-readable manifests 2374 and 2163 bind exact reference spin vectors, record IDs, six official file identities per case, levels and this protocol's hash. Retrieve only those listed files from CaltechDATA after the freeze. Verify official size/MD5 before decoding and record SHA-256. Raw and conditioned arrays remain outside Git; dataset terms remain separate from the software license. Preserve v3 memory/center-of-mass modifications.

Frozen conditioning and comparison

Apply the complete fixed conditioning, PSD, projection, NR alignment and model optimizer of the 3144 protocol, substituting each manifest's real highest/lower levels, q, spins and reference time, and the exact published options above. The reusable runner must retain real level labels in source records and disclose any internal role aliases.

In particular: full inertial NR modes ell=2..5 including m=0; 100 solar masses; 500 Mpc; f_ref=30 Hz; exact catalog q; common post-reference interval; cubic complex interpolation at 8192 Hz; 500 M rise and 100 M fall; direct Fourier sum at integer 30..1024 Hz; aLIGOZeroDetHighPower PSD; views (iota,psi)=(0.7,0) and (1.2,0.4). The nonspinning model sets both spins to zero; its tiny catalog spins are residuals. The aligned model uses exact reference z-spins and does not fit the negligible transverse components.

For NR resolution, choose one orbital phase and one arrival time jointly across all modes and both plus/cross views, using the 2048-point periodic phase grid and eight-peak refinement. Apply the same optimum to both view screens. For each model/backend/view, optimize only arrival time, orbital phase and polarization with seeds (17,23,41,67), popsize=12, maxiter=120, tol=1e-9 and polish=True. Intrinsics, inclination, distance and all amplitudes remain fixed. Time shifts use k/8192 s for integer k=-819..819. No mode-specific alignment or amplitude correction is allowed.

Acceptance and reporting

Each resolution view requires match >=0.995 and norm ratio in [0.95,1.05] under the same common alignment. Each native backend at each model view requires optimized match >=0.980, norm ratio in [0.85,1.15], and four-seed match spread <=1e-5. These unchanged project thresholds separately constrain shape and amplitude. A failed/infeasible resolution screen prevents a physical qualification claim even when model scores pass. Record cases individually; a two-case pass requires every case, backend, view and resolution gate to pass.

Validate direct/FFT agreement, mode-basis reconstruction, manifest and source hashes, and exact option forwarding before scoring. Record all optimizer seeds, norms, source/fixture/PSD hashes, package versions and full-minus-six-mode NR diagnostics. No production coefficients or mathematical definitions change. A pass is limited to these cases, mass, PSD, views and declared options; a failure is retained with its measured cause, without rescoring as a new pass.

Execution

Use scripts/run_imrphenomhm_published_nr.py with each frozen manifest and an external data directory on len. Prepare with the current SXS environment; evaluate with the current compatible Torch/JAX/PyCBC environment. Record the freeze commit and exact commands in the subsequent receipt. The runner and both manifests must be committed before either preparation command begins.