For UIOOPS developers

What porting UIOOPS found about the MATLAB itself — written to be useful without adopting this port.


Findings	123, each with `file:line` and how it was established
Stop a run	6
Change a number	19
Clean bills	checks that found nothing wrong, reported as such

See also Enhancements for which have an opt-in correction.

A port is an unusually thorough code review. The upstream-findings list reports 123 findings and several clean bills, each with a file:line, how it was established (measured / derived / read), what it reaches, and a suggested edit. Read-only items are flagged for your check, never asserted. Much of this code is inherited; several suspicions that proved wrong on measurement are reported too.

Some of these you already fixed

The order was: port UIOOPS v3.1.4 bug-for-bug → validate against the archived product, MATLAB's own output → write up the findings → then port UIOOPS v3.5, on a lineage that diverged from v3.1.4 years ago. Some findings had been already repaired upstream, independently, before this list was sent.

Nine so far. Six are set out in the findings document with the v3.5 line and what changed: the diode-32 hollow test (packing diodes 1-32 only, in v3.5 a whole-array boolean AND), the int16 diameter saturation (int16 × double collapsing the size — the cast moved to the read, dated in your own source), the double-scaled area-ratio concentration (conc_AreaR scaled by the diode size twice, the second scaling commented out with the reason beside it), the all-zero image_hollow (image_hollow written but identically zero, in v3.5 carrying a real value), the never-emitted habit C (habit 'C' unreachable behind a stub, the stub gone), and the inconsistent 2DC diode size (the 2DC diode size disagreeing between two files, reconciled to 0.03). A seventh, the late-start hole window (the hole-position window comparing an index to a span, so a late-starting particle is never scanned for holes), went in the same rewrite as the diode-32 hollow test.

Four of the nine were fixed the way this review would have suggested.

Two more came from porting v3.5's own sizeDist.m. The missing mean-ratio transpose — the three mean_area_ratio/mean_aspect_ratio_* variables and their REJ_ twins written without the transpose their neighbours carry (sizeDist.m:2124, :2147) — is fixed at :2412, :2434, under the comment "Fixed data structure for netCDF export ~ Joe Finlon 02/13/19", the same date as the fix for the int16 diameter saturation. And the first-event-only reacceptance scan, which is only half done:

Finding	In v3.1.4	In v3.5
The first-event-only reacceptance scan — the shattering reacceptance loop scans only the first event (MATLAB's `a:b` uses only a vector operand's first element)	`sizeDist.m:1036`, `:1169`	Fixed in one branch of three

The first-event-only reacceptance scan is still actionable. v3.5's new 1-Hz threshold branch (sizeDist.m:1416-1419) replaces the bare colon range from find(rBegin) to find(rEnd) (v3.1.4 sizeDist.m:1036) with one bracketed from the first rBegin index to the last rEnd index, under a dated comment — repaired on purpose. The same construct still stands at :1213 and :1346, so two of the three reacceptance paths in the current release keep the defect.

This set is read, not generated, and is not claimed complete. Our classifier establishes only whether a finding's code is present in a later release, and a bug rewritten to fix it still reads as present, so there may be more than nine.

Verifiable without Python

Every entry is checkable in MATLAB or Octave on your own data. Three examples:

A second that appears twice in the TAS file loses all its particles. one_sec_times is the TAS grid floored to whole seconds, so two housekeeping packets in one second give two identical entries — the first spanning an empty interval while its particle block is still consumed. 6.6–18.2 % of the seconds in every file measured are duplicated; on one dense file that is 1.2 million particles, and correcting the grid adds 5.9 % to its used bins. To check, floor your .tas.csv times to seconds and count diff(sort(t)) == 0.
The last record of every raw file is never a primary. The decode loop reads the pair (N, N+1), probes one byte for EOF and breaks, so particles beginning in the final record are dropped: 0–235 particles per file across 24 raw 2D-S files. In several it is the five-word packet header that straddles the record end, which the 2017 bounds guards do not cover.
CGAL_EllipseSize aborts the interpreter on real 2D-S data. With iRectEllipse = 1 and the real MEX routines, the unmodified imgProc_sm dies on the first image record: six of its sixteen particles are one pixel wide or tall, and a one-pixel shadow is collinear. The first five size bins hold 99.4-100.0 % of all particles, so degenerate geometry is common.

Selected findings


A duplicated true-airspeed second loses every particle in that second	6.6–18.2 % of the seconds in every file measured; the archived products show the same loss
The zero-image branch abandons a whole image record	measured at 2.35 % of records on a dense file — about 1.6 % of particles
`sum_IntArr` and `sample_vol` are rewritten for seconds that accept no particles	160 of 313 particle-bearing seconds on one flight
A file can report `Shattering Algorithm = 'Applied…'` when no shattering rejection ran	the output cannot say which call produced it
The Korolev out-of-focus correction is dead code in every tree	the depth and size factor are computed and never applied
Reject code `'t'` is unreachable	the looser threshold is tested second, so it never fires; the netCDF description states the two thresholds the other way round
`spot_edge(147)` is the literal `0757` — a missing decimal point	bounded consequence, and measured to affect no particle on six flights

The document also measures what the fixes were worth, running pre-fix and post-fix code side by side: two of the UIOOPS authors' own repairs matter more than their changelog lines suggest, and one means particle counts in older products are a lower bound, not recoverable from the product.

A companion page for the same audience is among the published pages.