Acquisition evidence and integration checks

Before converting a vendor file

A readable array is not proof that it is a correctly decoded FID. Confirm complex quadrature, direct-dimension length, acquisition order, digital filtering/group delay, spectral width, observation frequency, transmitter offset, and observed nucleus from vendor parameters. Keep the vendor files alongside the conversion record. Hash the source acquisition files and record the reader/version and any corrections; the helper hashes the supplied NPZ or NMRPipe file and processing JSON, not an earlier vendor acquisition.

A processed frequency-domain file must not be fed through this FID pipeline. Echo, nonuniform sampling, real-only acquisition, indirect dimensions, and data requiring receiver-specific phase cycling need a different processing derivation. The bundled helper rejects real arrays and multidimensional arrays but cannot distinguish an incorrectly labeled complex spectrum from a complex FID.

For Bruker, digital-filter removal is an acquisition-specific operation, not a universal first step. Determine whether it has already been applied; applying it twice distorts phase. Upstream bruker.read and bruker.remove_digital_filter are possible building blocks, but no vendor-reader workflow is claimed as tested here. Consult the Bruker reference before adapting it and verify against a trusted processed reference. remove_digital_filter(dic, data, truncate=True, post_proc=False) returns data, not a (dictionary, data) pair. post_proc=True returns a frequency-domain correction and must not be passed into this FID helper. Its default truncation may discard useful data; record the resulting point count and correction settings.

Verify frequency sign before fitting phase

The helper pairs proc_base.fft with the descending unit_conversion.ppm_scale() axis. The canonical FID therefore uses the negative complex exponential for positive ppm offsets. If a known reference appears mirrored about the carrier, revisit quadrature and sign convention. Reversing a plotted axis does not correct the data-to-ppm mapping. A reference offset correction should update the carrier and be recorded, rather than moving a plotted label without moving integration limits.

For the even zero-filled size N, the first sample is carrier_ppm + spectral_width_hz / (2 * observation_mhz) and each step is -spectral_width_hz / (N * observation_mhz) ppm. The final sample is one digital step above the lower bandwidth edge; the two Nyquist edges are not both sampled. unit_conversion receives carrier in Hz, whereas the NMRPipe FDF2CAR header is ppm. proc_base.em receives line_broadening_hz / spectral_width_hz, not Hz directly. proc_base.ps multiplies the complex spectrum by the positive phase exponential in degrees; first-order phase is measured across indices, with no implicit pivot.

Processing comparisons that change conclusions

• Reprocess with less line broadening to see whether nearby peaks remain separable.
• Compare no-baseline and justified signal-free baseline regions; broad resonances can be mistaken for a baseline. First-order polynomial correction is intentionally narrow.
• Inspect the imaginary channel during manual phase adjustment. A visually positive spectrum alone does not establish absorption-mode phase across the full bandwidth.
• Vary integration bounds and baseline anchors to assess area stability. Overlap, truncation ringing, solvent suppression, and acquisition dead time may dominate error.
• Treat automatic local maxima as candidate peaks. Negative peaks, multiplet grouping, line-shape fitting, isotope patterns, and assignments are not performed.

The synthetic regression signal has two exponentially decaying components with identical linewidth and known amplitude ratio. Its exact area benchmark is a mathematical processing check, not external validation of a spectrometer, sample preparation, or quantitative assay.

Supported NMRPipe boundary

The NMRPipe reader requires one direct FDF2 dimension, complex samples, an untransformed FTFLAG, and a header size matching the decoded complex array. Spectral width (Hz), observation frequency (MHz), and carrier (ppm) must match the explicit processing JSON within float32 header rounding tolerance. Selected header fields are preserved in the report. The sign convention is still explicit: the file format alone does not establish the acquisition's physical frequency sign or whether vendor conversion conjugated it.

FDF2TDSIZE must also match the stored point count. This prevents known prior zero filling or truncation from being reported as acquired time. The centered axis checks use CENTER = floor(N / 2) + 1 and ORIG = CAR * OBS - SW * (N - CENTER) / N, following nmrglue 0.12's NMRPipe source. The ORIG tolerance allows float32 header rounding, including cancellation near zero. An inconsistent ORIG cannot be ignored: pipe.make_uc uses it to derive the carrier. Headers can still be wrong or incomplete; these checks do not authenticate acquisition history. For NPZ input, acquired duration assumes every supplied point was acquired.

Synthetic canonical files generated with nmrglue were round-trip validated. An upstream 2,176-byte fixture generated with simTimeND / nmrPipe -fn SET separately verifies stored real/imaginary order and calibration. Its source and checksum are recorded in the repository test fixtures. This exercises the actual binary reader and recovered spectral coordinates, but does not establish that every experimental converter writes equivalent headers or samples. Digital-filter removal and earlier apodization are not inferred or undone. Check the upstream acquisition and conversion record before applying the processing settings. Neither nmrPipe nor nmrDraw was available for an independent live comparison. Linear baseline least squares is tested; resonance line-shape fitting and automatic phase fitting are outside this helper's scope.