Lossy vs lossless

mscompress always wraps the final output in ZSTD (which is lossless), but the pre-compression transform you apply to the m/z and intensity arrays can be lossy:

Path	What it does	When to use
Lossless	ZSTD only. Original double arrays preserved bit-for-bit.	Quantitative re-analysis, regulatory contexts, "I'm not sure yet."
Cast	Cast 64-bit doubles to 32-bit (or 16-bit) floats before ZSTD.	When 32-bit precision is plenty for downstream analysis.
Delta	Replace each value with the delta from the previous one, then quantize to 16/24/32 bits.	m/z arrays — they're nearly monotonic and small deltas pack well.
Log	Log-2 transform before ZSTD.	Intensity arrays with multi-decade dynamic range.
Bitpack / VBR	Variable-bit packing of quantized values.	When you need maximum compression and have characterized the precision loss.

You configure the algorithm per stream — m/z and intensity transforms are chosen independently. See Choosing a profile for recommendations and the algorithm catalog for per-algorithm details on precision, scale factors, and tradeoffs.

Warning

Lossy paths are not reversible. If you compress with delta16, the decompressed m/z values will not exactly match the original — they'll be quantized to ~127 ppm (depending on the scale factor). Always keep the original mzML if you might need bit-exact data later.