Compiled noisy reference pilot

2026-10-01 outcome: seven accepted noisy references; one unresolved mode gate. All eight trials are terminal; no sampler or calibration qualification.

The compiled-pilot-v1 experiment tests numerical posterior construction for the eight fixed-truth datasets from the original noisy experiment. It uses compiled JAX and finer grids under the original scientific gates. The original timeouts remain separate retained outcomes. No sampler or prior-drawn ensemble was executed in this pilot, so it cannot establish calibration.

The compact receipt binds the frozen protocol, terminal outcomes, independent verification and externally retained full report and figures.

Frozen design

The two free parameters are native component masses; all other waveform parameters remain fixed. The models, fixed truths, independent uniform native priors, noise seeds and per-model amplitude scales match the original noisy experiment. Masses are in solar masses.

Model	Fixed truth (mass1, mass2)	Native priors	Four fixed-trial noise seeds
PhenomD	(36, 29)	[33, 39] × [26, 32]	610101–610104
NRTidal v1	(1.6, 1.3)	[1.59, 1.61] × [1.29, 1.31]	620101–620104

H1/L1 data use matched Gaussian noise and the same effective PSD in generation and analysis. The fixed per-model waveform amplitude gives noiseless network SNR 20 at the fixed truth. Data-only chart discovery, the full rectangular native prior, chart Jacobian and matched-noise audit settings are unchanged.

Execution uses CPU FP64 and chunks of 128. Torch and the OMP/MKL/OpenBLAS thread settings are one; each trial has a separate four-core affinity. JAX retains its own runtime thread pools. Native Torch and compiled native JAX evaluate paired grids with 31, 61, 121, 241, 481, 961 and 1921 nodes per chart coordinate. JAX JIT is required; eager fallback fails the trial. Each trial has a 3600-second reference budget and a 3900-second external trial limit, with no automatic extensions or replacement trials.

The original scientific gates are unchanged. Acceptance requires two consecutive passing refinements in both frontends, frontend parity and the discovery peak check. A failed refinement resets the passing count. The mode gate concerns coordinates of the grid's joint native-likelihood maximum. Independent Gauss-Legendre integration uses fresh native Torch likelihood evaluations with declared orders 64, 128, 256 and 512; both grid-versus-Gauss agreement and order-to-order convergence must pass. Gauss agreement does not replace the grid quantile or truth-CDF checks.

Both matched-noise audits passed. These checks confirm the declared noise-generation convention; an accepted posterior reference additionally requires every numerical gate below.

Trial results

Seven trials pass every required reference gate. NRTidal fixed index 2 completes the entire ladder but fails the mode gate, so the aggregate outcome is failed_noisy_reference_pilot. All eight independent Gauss audits pass at 64 → 128; integration agreement does not replace the failed grid gate.

Model	Fixed index	Final grid side	Consecutive passes, each frontend	Gauss orders	Trial seconds	Outcome
PhenomD	0	481	2	64 → 128	141.61	Accepted
PhenomD	1	481	2	64 → 128	142.90	Accepted
PhenomD	2	481	2	64 → 128	139.68	Accepted
PhenomD	3	961	2	64 → 128	363.11	Accepted
NRTidal v1	0	961	2	64 → 128	563.80	Accepted
NRTidal v1	1	961	2	64 → 128	570.48	Accepted
NRTidal v1	2	1921	0	64 → 128	1975.34	Unresolved joint mode
NRTidal v1	3	961	2	64 → 128	562.21	Accepted

Trial seconds are canonical total elapsed times, including discovery and Gauss audits. The eight jobs ran concurrently: controller wall time was 1979.30 seconds (32.99 minutes), including process startup, versus 4459.12 seconds summed over the parallel trials. No trial reached its deadline.

Fresh verification of all 46 retained paired grids confirms identical chart coordinates, finite FP64 likelihoods and normalized posteriors. The maximum absolute Torch/JAX log-likelihood difference is 3.5350e-9, passing atol=1e-7, rtol=1e-8; maximum frontend posterior TV is 1.0404e-11 against 1e-5. Every JAX reference used JIT without fallback. Final Gauss-versus-grid TV ranges from 8.4286e-6 to 1.5702e-4 against 0.002; order-to-order TV-estimator drift ranges from 7.7305e-7 to 1.6829e-5 against 0.00025. The receipt retains the other numerical gates, failed refinements and native posterior summaries.

The TV curves show integration refinement. Red crosses mark every failed comparison; M identifies a mode failure. Low TV alone does not establish reference acceptance.

These are descriptive native marginal 5%, 50% and 95% quantiles. The dashed orange NRTidal index-2 intervals come from its unaccepted final grid. Its mass-1 marginal remains prior dominated (KL=0.000637 nats), with a 90% width of 0.017917 solar masses, about 89.6% of the declared prior span. Numerical convergence and parameter identifiability are separate questions.

An accepted fixed-truth reference supports numerical construction of the declared two-mass posterior for its particular same-model Gaussian-noise dataset. Four noise repeats per model do not form a prior-predictive calibration ensemble. Their truth-CDF percentiles are descriptive; no calibration screen, coverage qualification or sampler result follows from this pilot. Physical waveform accuracy, real-detector noise robustness and full-dimensional PE remain outside its scope.

Joint-mode grid alignment

NRTidal fixed index 2 fails the native mass-1 joint-mode gate at 241 → 481, 481 → 961 and 961 → 1921. The coordinate displacement halves from 2.0833e-4 through 1.0417e-4 to 5.2083e-5 solar masses, but the coarse native interval also halves: all three shifts are 2.5 intervals against a limit of one. Each failure resets the consecutive passing count to zero.

The grid maximum alternates around the data-only discovery maximum while its log-likelihood deficit shrinks from 1.2015e-5 through 3.4470e-6 to 6.6893e-7 to 2.8878e-7. At 961 → 1921 the joint-density TV is 2.7550e-5, native mean drift is at most 1.78e-8 of the 90% interval width, native quantile drift is at most 7.64e-9 of that width, and truth-CDF drift is 1.8128e-9. Stable posterior integration does not imply a stable grid argmax.

Both frontends select the same modes, and retained arrays confirm that all eight datasets exactly match the original experiment. The observed pattern supports lattice alignment along a narrow curved ridge with weak variation along its length as the leading explanation. On the final grid, the best and second-best retained nodes differ by 3.8015e-7 in log likelihood, far above the 1.1369e-13 FP64 spacing near that peak and the measured frontend differences. This is not an exact floating-point tie. It does not support a frontend disagreement or noise redraw. This diagnosis does not waive the frozen mode gate or turn the unresolved reference into an acceptance.

NRTidal fixed index 2 therefore remains an unresolved reference in this pilot. No sampler or calibration result is inferred from its otherwise stable posterior summaries.

Reproducibility

Execution used len; the frozen source archive and protocol below identify the exact executed experiment.

Frozen artifact	SHA256
Source archive source-compiled-pilot-v1.tar.gz	d988d0301ba5e6461ca07359971480d248fcb5cef4bd10e35f4f9a0b9a82ceee
Protocol noisy_protocol.json	fdf2330044415bad81a27581bf82d01bd6d1f56acaafd9480e1c75bc0d207d25

The artifact bundle diffgw-pe-compiled-pilot-20261001 retains bulk arrays, full receipts and the executed source snapshot. The linked compact receipt binds the retained files by hash. The bundle is retained evidence, not a public downloadable archive.