Known gaps

In short: every case either engine does not port raises an exception naming the source line it would need; none guesses. The one result that does not fail loudly is the HVPS particle count, which disagrees with an archived product and carries a runtime warning. Everything else here is a refusal with a workaround, or a limit of the method rather than the port.

See also Reference for what is complete.

UIOOPS v3.5 (the default engine)

The engine refuses these cases, each where the original MATLAB itself dies or needs a file no release ships.

PECAN projects take a parameter file you supply. getPECANparams.m ships in no release, so sizedist.run stops on a PECAN project until you pass pecan_params= (--pecan-params): a CSV of spiral start_s, end_s and intar_threshold_s, the format the refusal message names (arrayprobe_uiops.pecan_params). This is an input, not a gap the package pursues.

The v3.1.4 port is not in this package. It is the separate validation anchor that reproduced the archived products; this package ships the v3.5 and SODA-2 engines, so use v3.5 for new data.

v3.5 has no reference pair and cannot get one. No archived product was written by v3.5, so it is validated against Octave runs of its own sources (whole-program reference outputs, zero disagreeing items; the imgProc_sm reference output is distributed, the sizeDist and combine reference outputs are not). The ellipse fit (iRectEllipse=1) agrees with the compiled CGAL binary within 1e-2 on the axes and 1e-3 on the angle, not exactly.

SODA-2

The driver (arrayprobe soda2 run <config>) runs the SPEC path and the CIP/PIP path under the SEA and DMT formats end to end. Reprocessing from an existing PBP file is ported (arrayprobe_soda2.reprocess_pbp) and reached from Python, not the CLI.

aspect_ratio.pro is checked with a stand-in above two slices. GDL has no ERODE; those cases run against a shim with an unconfirmed border rule, labelled zero_outside_shim_unconfirmed and not counted as oracle evidence; the border rule is being settled from IDL's documentation of ERODE. SODA-2 is compared beside UIOOPS, never used as its oracle.

HVPS particle counts disagree with an archived product

Do not publish HVPS counts or concentrations as validated output. The UIOOPS HVPS chain lands on the archived product's exact time grid and bin ladder. Under one fixed inter-arrival call, the 1 µs default for every flight, it accepts a different number of particles, by a different factor per flight; the per-flight call that matches the archive is described below the table:

The difference is in the accepted particle set, in a minority of seconds where the archive's reject ratio is higher; the 2D-S, through the same chain, reproduces exactly.

What is inferred: the HVPS inter-arrival thresholds. The archived HVPS products were written in July 2020 by a campaign variant of UIOOPS (its sizeDist.m:64 reads a per-flight inter-arrival threshold file), and no threshold file from that run survives. The archive's counts are matched under one of two calls per flight: the 1 µs default on 2017-08-26 (0.9987) or the port's own threshold file on 2016-09-06 (1.0043). Nothing in the data decides which call a flight got, so this is a per-flight run configuration, inferred by comparison with the archive and recorded with the output, not an independent reproduction. The 0.43 % residual on 2016-09-06 sits in the blocks where the fits agree and its cause is open; 2017-08-13 is untested under the call its job log records. Across all six 2016 flights, with the port's threshold file fitted in five blocks, four flights (2016-09-06, -12, -14 and -20) match the archive to within 0.5 %; 2016-09-10 (+9.96 %, its blocks fitted across two raw files) and 2016-09-25 (−69.7 %) stay open.

The Octave oracle's limit. The unmodified IntArrAnalysis_time.m under Octave writes the port's thresholds in all five blocks on 2017-08-26 and in blocks 2-5 on 2016-09-06. Block 1 of 2016-09-06 differs (Octave 2.5653e-4 s, the port 4.12125e-4 s) because Octave ignores the fit's 'Robust','on' (the campaign variant's IntArrAnalysis_time.m:153) and the port honours it; with the robust option off, the port gives Octave's threshold exactly. Octave therefore cannot confirm the robust fit MATLAB ran. The archive favours the port's value: under it, block 1's hours differ from the archive by one particle. What would settle it: the threshold file from the archive's own run, which does not survive; for the robust fit alone, re-running 2016-09-06 with Octave's block-1 threshold — if block 1's residual grows, MATLAB's threshold is the robust one (not yet run).

Open limits and what would close them

Five limits, each stated with its cause and what to do. The first two are inputs you supply, not gaps the package pursues; the last three need an input from outside the package.

Other limits

• A fitted inter-arrival threshold needs scipy>=1.16. Below it the tau-derived threshold lands 47.6 % from the Octave reference (Installation); the default call uses the fixed per-probe value and is unaffected.
• Parallel decoding can die at worker start-up on Windows (WinError 6, BrokenProcessPool). It is opt-in (--s2-jobs, --jobs); run_flight.py retries, then falls back to the byte-identical serial decode and records it in a .decode sidecar. Root cause not established.
• Method limits however faithful the port. O'Shea et al. (2019, AMT 12, 3067–3079) advise against monoscale OAP data below ~100 µm, above this package's ≈55 µm used-bin floor (Glossary); McFarquhar et al. (2017, Meteor. Monogr. 58, ch. 11) report UIOOPS and OASIS sizing one particle up to 20 % apart; Jaffeux et al. (2022, AMT 15, 5141–5157) report 58.2 % (2D-S) habit agreement between a classifier and ten scientists' majority label, which bounds visible algorithm–human agreement and is not an inter-annotator figure.