All-waveform parameter-estimation campaign

This is the 30 September 2026 historical grid study. All 58 declared frontend/route outcomes were recorded: 36 converged baselines, 20 unresolved results and two component-only blockers. Recording every outcome does not mean all experiments passed.

For later reference and sampler evidence, see the current PE results. The historical outcomes below remain unchanged. This study tests bounded same-model numerical recovery; it supplies no full-dimensional, ensemble calibration or independent physical-accuracy claim. The protocol explains the checks.

Coverage and claim

The frozen executed scripts/qualification/pe_cases.py inventory covers 19 canonical dispatcher entries (15 frequency-domain and four time-domain), three additional distinct dispatcher routes, and six specialized waveform/API scopes: 28 scopes in total. A separate TEOB Hamiltonian component row is blocked because it supplies no public waveform; substituting another EOB model would change the claim. All 29 rows cover both Torch and JAX: 56 executed waveform/API experiments and two explicitly blocked component experiments.

Case IDs	Route and two free axes
TaylorF2, TaylorF2_HM, IMRPhenomD, IMRPhenomHM, IMRPhenomXAS, IMRPhenomXHM	FD dispatcher; component masses
TaylorF2_Eccentric	FD dispatcher; primary mass and reference eccentricity
IMRPhenomD_NRTidal, IMRPhenomD_NRTidalv2, IMRPhenomNSBH	FD dispatcher; component masses, fixed tidal parameters
IMRPhenomP, IMRPhenomXP, IMRPhenomXPHM	FD dispatcher; component masses, fixed precessing spins
IMRPhenomPv2, IMRPhenomXPHMSpinTaylor, IMRPhenomXPNR	Additional distinct FD dispatcher routes; component masses
IMRPhenomT, IMRPhenomTHM, IMRPhenomTP, IMRPhenomTPHM	TD dispatcher with fixed FFT adapter; component masses; TP/TPHM select nnlo
KerrRingdown, KerrRingdownDirect	FD dispatcher and separate direct class; final mass and spin
SEOBNRv4_ROM	Cached native public ROM class with dispatcher-equivalent default options; component masses; required checksum-validated external ROM asset
ParameterizedPostEinstein	Direct FD class; primary mass and phase deformation beta_ppe, fixed exponent
EMRIKludge, NewtonianEccentricEMRI	Native source polarizations with fixed FFT adapter; p0 and e0
ContinuousWave	Native detector strains with fixed FFT adapter; f0 and f1
MemoryWaveform	Native memory from IMRPhenomTHM physical (2,2) modes with fixed FFT adapter; carrier component masses
TEOBResumSHamiltonian	Component-only row; blocked, no public waveform

The executed inventory freezes numerical prior bounds, injection truth, options and fixed parameters. Alias spellings are not additional cases. The memory adapter supplies native h22 rather than fitting an arbitrary amplitude; its carrier accuracy remains a separate qualification question. One selected option tuple cannot qualify every version or a whole family.

Initial self-injections can establish numerical posterior recovery and pipeline consistency in their declared scope. They do not establish independent physical accuracy, population coverage or full-dimensional PE. Existing mathematical, binary and physical failures retain their original status. Independent physical injections and calibrated prior-drawn ensembles require separate campaigns.

Each frontend generates its own injection and fixed SNR scale. Comparing their self-injection outcomes therefore does not establish common-data frontend parity. The maintained regression suite separately checks IMRPhenomD likelihoods at fixed candidate points with shared injection data and PSDs; this is a representative check, not a campaign-wide posterior-parity result.

Frozen initial experiment

The initial campaign uses protocol v2, CPU FP64, zero noise, target SNR 20 and uniform bounded native priors in the two listed parameters, with every other parameter fixed. It declares 31, 61, 121, 241 and 481 nodes per axis before full-campaign execution and stops at the first adjacent-grid comparison that passes the numerical gates. It retains the original prior when a boundary or informativeness diagnostic is flagged. CDF and mean drift are recorded diagnostics, not additional pass gates. Protocol v2 adds joint-density total variation at most 0.05 to detect transverse ridge underresolution that the v1 marginal checks missed. The v1 pilot remains a separately retained preflight, not a completed v2 campaign result. The subsequent v2 IMRPhenomD pilot did not meet the joint-density gate by 121 nodes per axis. The schedule was therefore extended to 241 and 481 before the full campaign, retaining every prior, waveform option and numerical threshold. The pilot remains diagnostic evidence of underresolution rather than a final campaign outcome.

Ground-detector cases use native H1/L1 projection and analytical aLIGO PSDs. EMRI cases instead use synthetic plus/cross channels and flat PSD weighting; these do not model a LISA/TDI instrument. Continuous-wave cases already return native detector strains. TD adapters use the same fixed window, sampling grid, Fourier normalization and frequency mask for every candidate and injection.

Reproducibility

The compact receipts identify the original source, case options, priors, precision, numerical adapters and retained data. Compilation and batching choices do not change the frozen likelihood or acceptance gates. Historical self-injections remain separate in each frontend; they are not a common-data parity experiment.

Executed source

Record	SHA256
Executed Python source digest, as recorded by the runner	fb6d1e0f72733e9f4336dfb5e8f35a7b74185f3aafa30ef63488896480090aed
Retained original source-final.tar.gz archive	202d2338a4584092d29963453720b3c76ea2f0a416772c9a6ea7eb9937016c5f

The source digest binds the executed src Python files, runner and inventory; the archive hash identifies the original snapshot that produced these results. The reproduction manifest binds the source archives, runtime versions, verified ROM asset, numerical configuration and retained outputs. Per-row elapsed times describe individual final attempts, rather than total campaign wall time.

Numerical outcomes

The final compact receipt contains 36 completed_grid_baseline, 20 unresolved_convergence and two blocked outcomes across all 58 intended frontend/route combinations. Each frontend has 18 completed baselines, ten unresolved results and one component blocker. Every final row has a recorded outcome; unresolved integration is distinct from a failed execution. The receipt is 352,115 bytes, SHA256 b871b973bd54368ebfce6c5aa95e497328839039dae8eb86551a263f082ebe80. The original receipts and retained arrays are hash-bound; previous failed execution proofs and attempts remain separate evidence.

Each frontend cell gives outcome / largest completed grid / final joint TV. C means completed_grid_baseline, U means unresolved_convergence, B means blocked. Grid size is nodes per axis; TV is rounded to five decimals. Prior flags are reported separately below.

Case ID	Torch	JAX
TaylorF2	C / 481 / 0.03547	C / 481 / 0.03547
TaylorF2_HM	C / 481 / 0.03629	C / 481 / 0.03629
IMRPhenomD	C / 481 / 0.02723	C / 481 / 0.02723
IMRPhenomHM	C / 481 / 0.02737	C / 481 / 0.02737
IMRPhenomXAS	C / 481 / 0.02876	C / 481 / 0.02876
IMRPhenomXHM	C / 481 / 0.02627	C / 481 / 0.02627
TaylorF2_Eccentric	C / 241 / 0.03537	C / 241 / 0.03537
IMRPhenomD_NRTidal	U / 481 / 0.46619	U / 481 / 0.46619
IMRPhenomD_NRTidalv2	U / 481 / 0.46617	U / 481 / 0.46617
IMRPhenomNSBH	U / 481 / 0.65849	U / 481 / 0.65849
IMRPhenomP	C / 481 / 0.02716	C / 481 / 0.02716
IMRPhenomXP	C / 481 / 0.02858	C / 481 / 0.02858
IMRPhenomXPHM	C / 481 / 0.02604	C / 481 / 0.02604
IMRPhenomPv2	C / 241 / 0.00564	C / 241 / 0.00564
IMRPhenomXPHMSpinTaylor	U / 481 / 0.02815	U / 481 / 0.02815
IMRPhenomXPNR	C / 481 / 0.02836	C / 481 / 0.02836
IMRPhenomT	U / 481 / 0.18524	U / 481 / 0.18524
IMRPhenomTHM	U / 481 / 0.18649	U / 481 / 0.18649
IMRPhenomTP	U / 481 / 0.18579	U / 481 / 0.18579
IMRPhenomTPHM	U / 481 / 0.18694	U / 481 / 0.18694
KerrRingdown	C / 121 / 0.01718	C / 121 / 0.01718
KerrRingdownDirect	C / 121 / 0.01503	C / 121 / 0.01503
SEOBNRv4_ROM	C / 481 / 0.02589	C / 481 / 0.02589
ParameterizedPostEinstein	C / 481 / 0.00953	C / 481 / 0.00953
EMRIKludge	U / 481 / 0.03533	U / 481 / 0.03533
NewtonianEccentricEMRI	C / 121 / 0.01688	C / 121 / 0.01688
ContinuousWave	U / 481 / 0.20131	U / 481 / 0.20131
MemoryWaveform	C / 61 / 0.00072	C / 61 / 0.00072
TEOBResumSHamiltonian	B / — / —	B / — / —

The unresolved results all reach the declared maximum of 481 nodes per axis:

Routes	Remaining failed gate, in both frontends unless stated
T/THM/TP/TPHM	Joint TV, about 0.185–0.187
D_NRTidal/D_NRTidalv2/NSBH	Joint TV, about 0.466/0.466/0.658
XPHMSpinTaylor	mass2 90% width drift, about 0.0646, despite joint TV passing
EMRIKludge	p0 quantile drift about 0.0617 of the final 90% width and relative width drift about 0.1224, despite joint TV passing
ContinuousWave	Joint TV, about 0.201

Every available final marginal 90% interval spans at least four fine-grid intervals, including the unresolved results. EMRIKludge's p0 spans about 6.52 and XPHMSpinTaylor's mass2 about 8.50. Passing marginal resolution or TV alone therefore cannot replace the complete frozen gate set. Near-identical rounded frontend metrics remain separate self-injection results, not common-data parity.

The largest completed grid is distinct from the requested maximum or an interrupted attempt. A failed receipt may retain an earlier converged grid; the execution outcome still remains failed. The compact results record retains the final comparison's full failure list, measured SNR, execution strategy, fallback/error and hash-linked original receipt. Counts and tables must refer to the same frozen source and protocol, without mixing pilot receipts or duplicating partition outputs.

A case unresolved at 481 nodes per axis remains unresolved under this frozen schedule. Waveform/likelihood failures and persistence/reporting failures are distinct; an artifact error does not silently become a successful campaign outcome.

Prior and interval diagnostics

The final receipt flags 34 frontend/route results as prior-support limited (36 axes), and ten axes as prior dominated; no joint posterior is prior dominated under the frozen convention. These flags do not change the numerical outcomes above. Both frontends have the following flags.

Routes	Flagged axes
D/HM/XAS/XP/XPNR, T/THM/TP/TPHM	mass1: prior-support limited
TaylorF2_Eccentric	e0: prior-support limited and prior dominated
D_NRTidal/D_NRTidalv2, SEOBNRv4_ROM	mass1: prior-support limited and prior dominated
KerrRingdown/KerrRingdownDirect	final_spin: prior-support limited
ParameterizedPostEinstein	beta_ppe: prior-support limited and prior dominated
MemoryWaveform	Both component masses: prior-support limited

All 112 measured marginal 90% intervals contain their injected truth, including the unresolved grids. This descriptive inclusion is neither a numerical pass nor ensemble coverage. In particular, converged eccentric and ppE experiments can retain a prior-dominated free dimension. Among the numerically completed results, all 72 marginal intervals contain the truth; neither count establishes ensemble coverage.

For each free axis, the compact record retains its name, injection truth, frozen bounds, mean, standard deviation, joint-mode coordinate, 5%/50%/95% quantiles, truth CDF, low/high boundary mass, KL divergence and prior flags. Also retain joint KL and its prior-dominance flag. Report which dimensions remain prior limited or poorly identified, including among numerically completed experiments. Interval inclusion and agreement of separate self-injections supply no physical-accuracy or ensemble-calibration promotion.

Figures and artifacts

The following five figures are selected from the retained 169-figure bundle. They use SNR 20, zero noise and the frozen two-parameter priors; all other parameters remain fixed. They supply no physical-accuracy or ensemble-calibration claim.

Both frontends' recovered medians and marginal 90% intervals, normalized by each axis's frozen prior width. Hollow markers retain the 20 unresolved outcomes; the two component blockers supply no posterior and are excluded from this plot. The free-axis names vary by route. Overlapping frontend self-injections do not establish common-data posterior parity or physical accuracy.

Torch IMRPhenomD component-mass posterior on the final 481-by-481 grid: completed_grid_baseline. The correlated ridge and marginal densities belong to the declared bounded prior, with spins and all other parameters fixed.

Torch IMRPhenomD refinement through 481 nodes per axis. The final adjacent-grid comparison meets all protocol-v2 gates; the figure displays only log-normalizer and joint-TV diagnostics. The earlier near-stable normalizer alone did not establish joint convergence.

Torch IMRPhenomD_NRTidal component-mass posterior on the final 481-by-481 grid: unresolved_convergence, with fixed tidal parameters. Joint TV remains about 0.466, above 0.05, despite stable normalization and marginal summaries. The high TV shows that the joint density along this thin mass ridge remains underresolved. This self-injection experiment does not measure physical-model bias.

Torch EMRIKludge p0/e0 posterior on the final 481-by-481 grid: unresolved_convergence, using synthetic plus/cross channels and flat PSDs rather than a LISA/TDI response. Joint TV is about 0.035 and passes, but the p0 quantile drift is about 0.062 of the final 90% width and relative width drift is about 0.122; both exceed 0.05. The apparent smoothness of this plot therefore does not establish numerical convergence. Other source parameters remain fixed.

A P-P plot belongs only to an executed, appropriately generated calibration ensemble.

The compact receipt links original receipts and retained grid/signal arrays with their actual locations, SHA256 and byte sizes. Compact receipts and these five figures accompany this report. The executed scripts and configurations are preserved in the hash-bound source/reproduction archives; bulk arrays and the full figure collection are retained externally. The original source archive and actual executor configurations define reproduction of the executed experiment.

Retained bundles are identified as parameter-estimation-results.tar.gz, parameter-estimation-figures.tar.gz and parameter-estimation-reproduction.tar.gz. The execution manifest binds their hashes and reproduction settings. The 169-figure collection is retained externally; this page includes five selected figures. These identifiers do not imply a separately published bulk download. Reproduction must use the frozen source archives and original configuration, rather than assuming that a later checkout produced the historical results.