Research claims, with their evidence attached.

Increasing resolving power from 100,000 to 150,000 improves the controlled bound by 37.1%, 11.1%, and 2.48% for intrinsic Gaussian widths of 1.0, 2.5, and 5.0 km/s. The broad-line null below 10% is not rejected.

With AR(1) correlation ρ=0.8, nominal 95% iid coverage falls to 54.9% and 56.1% for deep and shallow injections; event-clustered intervals reach 93.2% and 92.1%.

Across 7,000 strong-feature injections, pixelwise HDR lowers centroid RMSE from 0.012954 to 0.002086 pixel, an 83.9% reduction; faint-feature RMSE remains unchanged.

Repository evidence records are represented by one validated schema and a deterministic public registry.

The component-attribution gate fails: six parameters leave two residual degrees of freedom, the standardized condition number is 175.31, maximum VIF is 4,881.4, and three exposure coefficients change sign under leave-one-epoch-out refits.

Thirteen targets support both actions. Stellar radius has the larger own-parameter score for 10, but the indirect route wins for none at |rho| <= 0.90; Kepler-296 f requires |rho| = 0.995225 to break even.

The reviewed 7,500-row chronological prefix contains 586 nominal-close approaches, of which 396 remain close at the three-sigma maximum distance; 190 nominal-close rows cross the distance boundary. Only 235 rows have measured diameters, and 61 nominal-close rows are albedo-sensitive around the 140 m screen.

Seven upstream records now route each research question and generated result to an exact source revision, versioned release, live evidence surface, and project-specific boundary; discovery-only links are explicitly excluded from that verified set.

Across 16 feature/filter comparisons, four are in the half-power core, one is an edge placement, and eleven are outside useful sampled response. CH₄ at 3.30 μm in F277W retains about 0.026% of peak response, a 3,805× source-photon-limited relative-time proxy.

Only 39 of 250 files pass all six inference-readiness gates. The audit finds 34,153 rows in repeated-epoch groups, 181 files with reference/offset confounding, and 44 files with extreme full-to-robust RV-span ratios; non-ready bundled files are now blocked from automatic scanning and fitting.

Every retained sample matches the released FITS source within the declared six-decimal rounding, and all three distribution gates pass. The temporal-fidelity null is rejected: 20.89%, 9.19%, and 64.46% of full-rate gradient-RMS, command-RMS, and mean-flux power lies above the reduced product's 5 Hz Nyquist frequency.

The blind search detects 864 of 900 balanced synthetic injections (96.0%, Wilson 95% CI 94.5–97.1%) but recovers the predeclared parameters in 710 (78.9%, 76.1–81.4%). At F146=24 and tE=0.02 d, detection is 41/75 while parameter recovery is 22/75; 0/1,000 simple constant-flux nulls trigger, constraining that scoped null rate below 0.383% at 95% confidence.

The historical window reproduces at 4.07 ± 23.16 ppm (0.176 sigma), while 30 predeclared window choices span -0.861 to +2.258 sigma and none reaches |3 sigma|. The 27-bin MIRI flat fit gives p = 0.000968, but deleting the 5.375 micrometre bin raises p to 0.07832, so the result fails the declared every-one-bin-deletion robustness rule.

All 113 native source-data channels were reconstructed. The component-removal residual contrast remains positive across all 30 declared designs per geometry: 63.17 to 85.74 in transmission and 217.65 to 245.99 in emission. The result is posterior morphology sensitivity, not an independent methane detection test.

Across 288 predeclared synthetic spectra, the calibrated legacy proxy is at least twofold too optimistic in 39 scenarios and reaches a maximum exact-to-proxy uncertainty ratio of 2.924. From R=100,000 to 250,000 it claims a 60% gain for every line width, while the exact median gain falls to 5.86% for 10 km/s intrinsic lines and rises to 66.89% for 1 km/s lines.

All 60 declared preprocessing designs retain a positive supplied-model contrast, spanning delta chi-square 559.6 to 780.4, and every single-bin deletion remains above 684.8. Yet deleting a 0.5-micrometre block around the 4.3-micrometre feature reduces the contrast to 3.19 and the outside-4.1-to-4.6-micrometre refit gives 6.04, demonstrating preprocessing stability but strong spectral localization.

All 180 predeclared first-harmonic designs place the maximum before eclipse, spanning -9.71 to -7.66 degrees with a median of -8.82 degrees; leave-one-trusted-wavelength results span -9.72 to -8.95 degrees. Four phase observations identify the three-parameter first harmonic with one residual degree of freedom, while the five-parameter two-harmonic design has rank four and is underidentified.

All 74 valid predeclared estimator designs retain a positive helium dip, spanning 0.511% to 1.160% with a median of 0.841%. The 16 default-design leave-one-point-out estimates span 0.747% to 1.003%; formal signal-to-noise values of 3.64 to 14.06 are conditional on independent reported errors and are not an independent detection significance.

On the current six verified MAST FLT/FLC bytes, the raw median suppression is 0.843 across 38 accepted measurements and the 34-coordinate-cluster median is 0.883 with a 95% cluster-bootstrap interval of 0.495 to 1.094. Single-exposure medians span 0.565 to 1.194 and leave-one-exposure-out medians span 0.612 to 1.025. All six archive hashes changed after reprocessing; the positive direction persists, but the magnitude is not stable enough for a population-level calibration claim.

The corrected default design retains 70 pixel measurements across two usable exposures. Across all 18 declared selection designs, the pooled median late-to-early rate change remains negative, spanning -2.762% to -1.074%; however, one exposure's design range crosses zero and only two exposures are usable, so the result is descriptive rather than detector-wide.

The corrected 20,000-trial model gives mean geometric throughput 0.9490 conditional on command success, with a Monte Carlo 95% interval of 0.9482 to 0.9499. A separate conservative 4% random non-opening scenario gives mean effective throughput 0.9100. Across 20 declared designs, conditional means span 0.6185 to 0.9568; these are synthetic Gaussian-PSF scenarios, not forecasts for a planned observation.