[v3.2.0] Release Notes (release candidate, synced from RELEASE_NOTES_v3.2.0.md)

[BenjaminTJohnson]

Synced 2026-08-06 with RELEASE_NOTES_v3.2.0.md at tag v3.2.0-rc1. The file in the
repository is authoritative:
https://github.com/JCSDA/CRTMv3/blob/v3.2.0-rc1/RELEASE_NOTES_v3.2.0.md
Partner-testing trackers: #351 (GSI/OMD), #352 (jedi-bundle). Feedback window
runs through Tuesday 2026-09-08.

CRTM v3.2.0 Release Notes

Status: release candidate. The library code is frozen and the coefficient
tarball is rolled, verified and published.

Coefficient data. Three efforts landed: the IASI-NG regeneration (SpcCoeff
and ODPS TauCoeff, regenerated 2026-08-01 by crtm-coeffgen on the LBLRTM
backend), the GeoXO gxi rework (gxi_geoxo and v.gxi_geoxo, same date), and
the ABI ODPS family regeneration (2026-08-03/04, see below). Note that
gxs_geoxo_lw and gxs_geoxo_mw are unchanged v3.1.4 inheritances and were not
part of the GeoXO rework.

The tarball was verified against the staging tree file by file: 1440 files,
all netCDF, zero differences, extracting to fix_REL-3.2.0.0/fix/.

Scope of the coefficient change

Every file classified against the v3.1.4 baseline (fix_REL-3.1.2.0, md5
0e5888cae80aa674b2e67ecd4490317d, the tarball v3.1.4 itself pinned). Method
and full artifacts: test-data-release/coeff_delta_REL-3.2.0/.

family	identical	metadata-only	data-changed	new	retired
SpcCoeff	396	53	14	75	1172
TauCoeff	0	705	11	93	1883
NLTECoeff	0	6	1	32	0
EmisCoeff	3	0	0	19	2
CloudCoeff	1	0	0	9	15
AerosolCoeff	0	0	0	4	0
ACCoeff	0	15	0	2	2
BeCoeff	1	0	0	0	0
test_data	0	0	0	0	531
total	401	779	26	234	3605

Reading that table:

• 234 files are new and 26 have changed numbers. Those 26 are the only
  files that can move a brightness temperature for a sensor you already use.
• 779 are metadata-only: the provenance backfill wrote Title, History,
  Comment and Profile_Set_Id where they had never been written. No radiometric
  change.
• 3605 "retired" overstates the loss. 2558 of those are format drops, not
  retirements: the Big_Endian/Little_Endian binary split disappears because the
  tree is now uniformly netCDF. Of the genuine netCDF retirements, 14 are
  renames with verified counterparts (JPSS-2 became NOAA-21 at launch, so
  atms_j2 is now atms_n21, and similarly for its siblings), 361 are bulk
  per-detector and passband variants, and 7 are withdrawals with individual
  justification. The reconciliation is in
  coeff_delta_REL-3.2.0/retirement_reconciliation/RETIREMENT_RECONCILIATION.md.
  If a sensor you use appears to have vanished, check the rename list first.

The ABI infrared family was regenerated late in the cycle. All six products
(abi_g16, abi_g17, abi_g18, abi_g19, abi_gr, abi-81K_g17; SpcCoeff
and ODPS TauCoeff each) carry new numbers. The staged abi_g19 was a 2024 STAR
test article whose fitted-CO2 extrapolation produced a measured -1.67 K channel
16 bias against 1279 GOES-19 clear-ocean superobs, plus spurious stratospheric
water Jacobians in the window channels; it had shipped since REL-3.1.2.0. The
family was rebuilt at the 2026.5 gas epoch with per-flight-model CWG SRFs
authenticated against NOAA NCC. ABI users should expect a brightness
temperature change and should not carry forward bias corrections trained on the
old coefficients. The v.abi_* visible halves were not part of this pass.

file	fix_REL-3.2.0.0.tgz
size	3,377,422,223 bytes
md5	88995873986cf2b077808a75d1c56f83
published	2026-08-06, https://bin.ssec.wisc.edu/pub/s4/CRTM/

It is smaller than the June 2026 tarball because the July campaign retired and
replaced coefficient files and the tree is now uniformly netCDF.

A default build works. The checksum above is what test/CMakeLists.txt
and Get_CRTM_Binary_Files.sh pin, and it is what the server serves, so
cmake .. downloads and verifies the correct tree with no extra arguments.
The 2026-08-06 archive was verified locally before promotion: all 1440 members
byte-identical to the staging tree (which itself is hash-verified against the
validated coefficient sources in the campaign ledger). Upload confirmed
against the server on 2026-08-06: Content-Length: 3377422223, with the first
and last MiB byte-compared clean against the local roll. The size alone
(3,377,422,223 vs 3,377,517,263) distinguishes it from the superseded
2026-08-05 copy. Building against an unpacked tree remains supported:

cmake .. -DFIX_FILE_PATH=<path-to>/fix_REL-3.2.0.0/fix

Developer-facing detail for every item below (commits, affected tests, TB
impact) is in REL-3.2.0_changes_vs_develop.md. Changes are stated relative
to CRTM v3.1.4; the v3.1.4 tag and the develop branch point differ only in
the default coefficient-data location (no library code differences), so the
same catalog applies against either baseline. A per-sensor inventory of the
shipped coefficient files is in REL-3.2.0_coefficient_inventory.md.

Highlights

• netCDF coefficient transition. netCDF is now the canonical coefficient
  format: SpcCoeff, TauCoeff (ODPS/ODAS/ODSSU/Zeeman), CloudCoeff,
  AerosolCoeff, and emissivity LUTs all load from .nc files, and the fix
  tarball ships netCDF only. Regression baselines are netCDF as well.

• PARMIO MW-water emissivity (new, default-on at and above 200 GHz). A
  LUT-driven physical-reference ocean emissivity backend for sub-mm sounders
  (AWS, TROPICS class). When PARMIO.MWwater.EmisCoeff.nc is present
  (CRTM_Init auto-loads it; it ships in the fix tarball), MW-water channels
  at or above 200 GHz route to PARMIO; below 200 GHz (every legacy sensor)
  the FASTEM path is byte-identical to v3.1.x.

• TELSEM2 MW-land emissivity atlas (new, opt-in). Enabled by the new
  CRTM_Init argument Use_MWland_Atlas=.TRUE. (which auto-resolves the
  default-named TELSEM2.MWland.EmisCoeff.nc from the coefficient path) or by
  passing an explicit MWlandCoeff_File. Without the opt-in, MW land
  emissivity uses NESDIS_LandEM exactly as before, even if the atlas file is
  present in the coefficient directory.

• Level-resolved downwelling/upwelling radiance profiles (new outputs).
  Opt-in via Options%Compute_Down_Radiance_Profile /
  Compute_Up_Radiance_Profile; fully differentiated (FWD/TL/AD/K) and
  combined correctly for fractional cloud. The surface downwelling radiance
  RTSolution%Down_Radiance is a first-class output (always populated on the
  emission path; enabled via Options%Compute_Down_Radiance for scattering).

• Vector radiative transfer (Options%n_Stokes > 1) substantially repaired.
  The surface is now handed to the solver in the Stokes basis rather than
  (V,H); the third and fourth Stokes components survive the surface aggregation
  and the azimuthal accumulation instead of being zeroed at both ends; the
  clear-sky path gained a vector solver, so a cloud-free polarimetric run no
  longer returns Q = U = V = 0; the fractional-cloud clear/cloudy combine is
  differentiated across all Stokes components; and RTSolution%Radiance is now
  the channel-polarized measurement rather than Stokes I (see behavior change
  14). Tangent-linear, adjoint and K are verified at n_Stokes = 4 against
  finite differences, the adjoint dot-product identity and K against AD, for
  both atmospheric and surface control variables including wind direction.
  This path has not been validated against an external model; see the known
  issues before using U or V quantitatively.

• Analytic MW-land surface Jacobians (issue Missing Jacobians for physical inputs (CRTM v3.x) #281) — the NESDIS_LandEM
  microwave land path (< 80 GHz) now returns analytic TL/AD/K sensitivities for
  every physical land-state variable: LAI, vegetation fraction, soil
  moisture, soil temperature, and land (skin) temperature. Canopy_Water_Content
  has an exactly-zero Jacobian (the forward never consumes it). This also
  corrects the Land_Temperature Jacobian below 80 GHz, which previously omitted
  the emissivity's skin-temperature dependence (through the LandEM gsect0
  thermal ratio) and was therefore too large; it now matches finite differences.
  Forward radiances are unchanged.

• MW scene-ozone transmittance component (GROUP_MW_O3, Group_Index=7)
  for microwave sensors.

• UV scene-NO2 transmittance component (GROUP_UV_NO2, Group_Index=8,
  issue GROUP_UV_NO2 (Group_Index=8): scene-variable NO2 ODPS transmittance component for UV-VIS sensors #340) for UV-VIS air-quality spectrometers (TEMPO, GEMS class). A
  sixth ODPS component carries scene-variable NO2 absorption: supply an NO2
  profile in the Atmosphere (HITRAN id 10, ppmv) and NO2-sensitive radiances
  respond; omit it and the coefficient file's reference climatology applies.
  Full FWD/TL/AD/K support, verified by a machine-precision predictor-level
  transpose test and an end-to-end TL/AD/K parity test. Group 1/2/3/7
  coefficient files never reach the new code.

• UV sensors can now run the forward operator (issue UV sensors (Sensor_Type 4) cannot run the forward operator: no UV branch in the CRTM_SfcOptics dispatch #339). The
  surface-optics dispatch previously had no UV (Sensor_Type 4) branch, so
  every UV SpcCoeff (the shipped OMPS family included) failed CRTM_Forward
  with "Unrecognised sensor type". UV channels now share the VIS Lambertian
  surface-optics path, and a UV-only sensor list loads the VIS surface
  emissivity LUTs at CRTM_Init. MW/IR/VIS behavior is unchanged.

• New sensor: gems2_amethyst (Weather Stream GEMS2 24-channel
  microwave sounder, 118.75 GHz oxygen bank plus 160-183.31 GHz humidity
  bank, on the GEMS2-Amethyst smallsat). Generated with crtm-coeffgen
  (MonoRTM, ECMWF84); brightness-temperature validation against
  line-by-line truth averages 0.04 K. WMO ids are invalid-value sentinels
  until C-5/C-8 assign codes; the 118.75 GHz line-center channels carry no
  Zeeman treatment. Unrelated to the Korean GEMS UV spectrometer
  (gems_gk2b) despite the acronym.

• FY-3 microwave family completed (12 sensors, FY-3C through FY-3G).
  New coefficient pairs for MWHS-2 (FY-3C/D and the E-variant on FY-3E/F),
  MWTS-2 (FY-3C/D), MWTS-3 (FY-3E/F), MWRI (FY-3C/D), MWRI-2 (FY-3F;
  instrument failed in 2025, coefficient serves historical reprocessing),
  and MWRI-RM (FY-3G). Generated with crtm-coeffgen from the NWP-SAF
  passband definitions and validated with forward, weighting-function, and
  adjoint physics checks; the 57 GHz line-splitting channels reproduce the
  AMSU-A weighting-function progression.

• AMSR3 humidity-channel bandwidths corrected to WMO OSCAR. The bundled
  AMSR3 spectral response function was too wide on its three highest channels:
  165.5 GHz by a factor of 1.25, 183.31 +/- 7 by 2.35, and 183.31 +/- 3 by 2.72,
  all against the per-sideband figures published by WMO OSCAR. The other 18 of
  21 channels matched OSCAR exactly and are unchanged, and STAR independently
  revised the same three. Against 2152 collocated observations the 183.31 +/- 3
  bias improves from +4.270 K to +3.422 K and 183.31 +/- 7 from +3.118 K to
  +2.906 K, with channels 1 to 18 unmoved. The larger effect is on the
  Jacobians: the 183.31 +/- 3 water vapour Jacobian error against a
  reference-free tiled-channel truth falls from 9.468 percent to 0.338 percent,
  which retires the belief that ODPS could not represent a wide double-sideband
  channel. That was never an ODPS limitation, only a band 2.72 times too wide.
  Recorded caveat: 165.5 GHz moved the wrong way on observation-minus-background
  (+3.951 K to +4.168 K), but that channel sits inside a 3.4 to 4.2 K
  common-mode bias shared by every variant including STAR's, so
  observation-minus-background cannot arbitrate it; the change rests on OSCAR,
  on STAR's independent revision, and on the 18-of-21 exact match. Spectroscopy,
  training profiles and algorithm are unchanged, and only the TauCoeff differs
  (the regenerated SpcCoeff is identical in every data variable).

• INSAT-3DS visible sensors completed. v.imgr_insat-3ds and
  v.sndr_insat-3ds previously shipped a SpcCoeff with no TauCoeff and
  could not pass CRTM_Init; both now carry TauCoeffs generated from the
  measured ISRO SRFs (crtm-coeffgen, ECMWF84) and regenerated SpcCoeffs
  whose centroids match the previously shipped files to better than
  0.03 nm.

• CRTM-Exp cloud-optics schema (experimental, opt-in). A new
  habit-resolved cloud LUT format selected explicitly with
  Cloud_Model='CRTM-Exp'; the default cloud path is unchanged.

• SNICAR visible snow reflectance LUT (new, opt-in). A SNICAR-based VIS-snow
  reflectance table, SNICAR.VISsnow.EmisCoeff.nc, ships in the fix tarball
  alongside updated IR snow emissivity modules. The default snow surface path
  (NPOESS) is unchanged: you get SNICAR only by asking for it by filename, since
  there is no VISsnowCoeff_Scheme argument (unlike MWwaterCoeff_Scheme).

  err = CRTM_Init( Sensor_Id, ChannelInfo, &
                   VISsnowCoeff_File = 'SNICAR.VISsnow.EmisCoeff.nc', &
                   File_Path         = coeff_path )

  VISsnowCoeff_Format selects the file format and defaults to netCDF.
  An unrecognised filename prefix (anything other than NPOESS or SNICAR)
  is a hard CRTM_Init failure rather than a silent fallback to no snow table.

  Maturity, stated plainly: this table is shipped, loadable, and exercised
  by one radiance-level test (test_SNICAR_VISsnow_Physics) that pins the
  grain-size, depth and density response against an invariant NPOESS control
  through the full radiative transfer path. That is the extent of what is
  verified. Two known limitations remain: the table's angle dimension is
  labelled "Solar Zenith Angle" in the file but is interpolated at the RT
  view/quadrature angles, so the solar zenith angle never reaches the table;
  and the forward path applies no bounds guard, so snow states outside the
  table extrapolate silently while the tangent-linear and adjoint return
  exactly zero. It carries no validation package of the kind behind the ABI or
  IASI-NG coefficient work. Treat it as experimental and evaluate it against
  your own cases before operational use.

• ODPS transmittance-algorithm modernization (issue ODPS group system modernization: load-time validation, group registry, per-component kernels, file-roster dispatch #343). The ODPS group
  system was rebuilt on a single group registry with load-time validation of
  Group_Index and the Component_ID/Absorber_ID rosters (malformed or
  mislabeled coefficient files are now rejected at load with a clear message
  instead of computing garbage), per-component predictor kernels for FWD, TL,
  and AD, and file-roster-driven dispatch. Results are bit-identical for valid
  coefficient files; Zeeman-reserved group indexes are refused (the historical
  OMPS "Group 4" failure mode).

• Long coefficient paths (issue CRTM string lengths for filesystem paths are too small #238). Coefficient file paths are now
  carried in deferred-length strings instead of fixed 80/128/256-character
  buffers, so deep installation paths no longer truncate silently;
  initialization through a ~300-character path is regression-tested.

• Fastem1 SST Jacobian corrected. On the legacy Fastem1 MW-water path
  (Options%Use_Old_MWSSEM=.TRUE., frequency >= 20 GHz) the emissivity's
  sea-surface-temperature derivative was silently dropped, so
  Surface_K%Water_Temperature carried only the skin-emission term. The
  Jacobian is now complete and validated against finite differences. The
  default (FastemX) path was never affected.

• Intel builds fixed: TELSEM2 atlas load and the PRA polarization angle.
  Two defects that a GNU-only test suite had been passing, both found by adding
  an ifx build to the release verification.

  On Intel, CRTM_Init segmentation faulted whenever the TELSEM2 atlas was
  requested, on any machine with the stock 8 MB Linux stack limit. The atlas is
  large: n_data is 2,770,889, so cell_number is 11 MB and emissivity is
  155 MB. nf90_get_var takes an assumed-shape dummy and passes it down to an
  F77 layer that takes an assumed-size one, and the compiler bridges the two
  with a contiguous copy-in temporary. That temporary is created inside the
  netCDF library's own compiled code, so it follows the flags netCDF was built
  with and not CRTM's, which is why no CRTM compiler flag can prevent it. The
  reader now takes the atlas in bounded slices, so every temporary stays small
  however netCDF was built. This is a read-path change only; the values loaded
  are identical.

  Separately, the PRA_POLARIZATION surface-optics branch divided zero by zero
  at nadir scan angle. The shared denominator reduces exactly to
  |sin(phi)|*sqrt(1 + sin(theta_f)^2), and both numerators vanish with it, so
  the expression was undefined there and returned whatever the compiler folded
  it to: GNU gave a polarization weight of 1, which selects the opposite
  polarization to the correct limit of 0, and Intel gave a NaN that propagated
  into the radiance, the weighting functions and the adjoint. The singularity is
  removable, and passing the two numerators to ATAN2 removes it rather than
  special-casing it. This affects gems2_amethyst and gems2_beryl only, both
  new in this release, so no previously released sensor changes behavior. The
  expression had been duplicated in the forward, tangent-linear, adjoint and
  Stokes-projection paths and is now one shared function.

• netCDF SpcCoeff now loads its NLTECoeff and ACCoeff siblings. The
  binary SpcCoeff reader streams both substructures inline from the same file
  via a DATA_PRESENT indicator; the netCDF layout instead stores them as
  separate <sensor>.NLTECoeff.nc and <sensor>.ACCoeff.nc. In v3.1.4 the
  runtime read path never looked for them, so a netCDF run silently attached
  neither. The machinery existed but was wired only into
  SpcCoeff_netCDF_to_Binary and SpcCoeff_Binary_to_netCDF, so placing the
  siblings next to the SpcCoeff did not help either, contrary to a
  reasonable expectation. v3.2.0 resolves the sibling from the canonical
  fix/NLTECoeff/netCDF/ layout and falls back to a flat co-located layout.

  Measured on AIRS by running both libraries against identical coefficients:
  brightness temperatures differ by up to 36.4 K on the 4.3 um shortwave CO2
  channels in daylight, confined to channels 1900-2114 (the NLTE set), with
  identical optical depths, and exactly zero difference at night where the
  NLTE correction is inactive. The Jacobian effect is proportionally larger:
  peak dTB/dT on the affected channels changes by a median of 23 percent and
  up to 90 percent.

  The antenna-correction half was measured the same way, on AMSU-A with
  Options%Use_Antenna_Correction = .TRUE. and a non-zero iFOV: deleting
  amsua_n19.ACCoeff.nc changes v3.1.4's answer on zero of 3780 rows and
  v3.2.0's on all of them, by up to 1.225 K with a mean of 0.611 K.
  Smaller than the NLTE effect but it touches every channel and enters as a
  systematic bias rather than a scatter.

  In the v3.1.4 coefficient tree this reaches 7 sensors with an NLTE sibling
  (AIRS and its module variants, CrIS B3, IASI B3) and 17 with an
  antenna-correction sibling (AMSU-A, AMSU-B, MHS). Users reading binary
  coefficients were never affected, since that reader streams both
  substructures inline.

• A visible-sensor hang without cloud coefficients is gone. Under v3.1.4 a
  clear-sky visible forward run with Load_CloudCoeff=.FALSE. and
  Load_AerosolCoeff=.FALSE. initialized normally, entered the first
  CRTM_Forward call and did not return: one call exceeded 100 seconds of CPU
  where the full 84-profile, three-angle run otherwise finishes inside that.
  Loading the two coefficient sets avoids it, and v3.2.0 runs the same
  configuration to completion. Reported for the benefit of anyone who hit this
  and worked around it; the trigger is identified, but the underlying
  non-convergence was not traced and it was not proven that the v3.1.4 call
  never terminates.

• Runtime OpenMP control. OMP_NUM_THREADS is honored at run time (no
  longer captured at configure time).

• Expanded self-checking test coverage. New baseline-independent checks
  include general TL-vs-FD and adjoint-consistency tests across the three
  main sensor types (finite difference TL/AD ctest #280), multi-sensor single-call bit-consistency, ODPS
  group-validation and long-path initialization tests, a DDA-ARTS ICE_CLOUD
  behavior pin, multi-sensor OMPS UV and TEMPO UV+VIS physics verifications
  (each registered when its pre-release coefficient pairs are present), and
  OpenMP thread-count consistency tests (Thread safety comprehensive examination #111).

Coefficient changes

The shipped tree is 1440 files, all netCDF. Every one was hash-classified
against the v3.1.4 baseline (fix_REL-3.1.2.0), with NaN bit-patterns
canonicalised and VLEN strings hashed by content:

classification	files	meaning
identical	403	byte-equivalent content
metadata-only	789	provenance attributes written or backfilled; no physics change
data-changed	14	the actual work list
new	234	products that did not exist in v3.1.4

The 14 data-changed files are the only ones whose numbers moved:
airs_aqua.NLTECoeff; SpcCoeff for iasi-ng_metop-sg-a1,
metimage_metop-sg-a1, mws_metop-sg-a1, v.abi_g18, v.imgr_insat-3ds,
v.metimage_metop-sg-a1, v.sndr_insat-3ds, v.viirs-i_n21; and TauCoeff for
iasi-ng_metop-sg-a1, metimage_metop-sg-a1, mws_metop-sg-a1,
v.metimage_metop-sg-a1, v.viirs-i_n21.

On the large "retired" count, which is easy to misread. The census records
3605 retired paths, but 2612 of those are .bin files removed by the
deliberate binary-format drop, not products withdrawn. Of the 993 netCDF
retirements, the overwhelming majority are per-detector and spectral-shift
variant families withdrawn deliberately, plus 14 renames whose counterparts
are verified present (JPSS-2 became NOAA-21 at launch, and similar), and 7
individually justified withdrawals. No operational product disappeared without
a counterpart.

Per-sensor detail (id, platform, generation date, provenance, ACCoeff/NLTECoeff
presence, validation status) is in REL-3.2.0_coefficient_inventory.md. The
evidence behind the data-changed entries — line-by-line truth comparisons,
observation closure, Jacobian statistics, cross-sensor validation and a
standing adversarial audit — is in
test-data-release/coeff_delta_REL-3.2.0/, indexed by its DELTAS.md.

Breaking and behavior changes

1. RTSolution file formats changed incompatibly. The netCDF reader
   requires variables absent from files written by earlier versions
   (per-element RT_Algorithm_Name, Reflectance, Downwelling_Radiance,
   the n_Layers global attribute), and the binary record grew. RTSolution
   files written by pre-3.2.0 code cannot be read; regenerate archived files.

2. Options%Obs_4_downward_P removed (compile-breaking). Migrate to
   Options%Compute_Down_Radiance / Compute_Down_Radiance_Profile and read
   RTSolution%Down_Radiance / RTSolution%Downwelling_Radiance(:).

3. Coefficient wrapper I/O defaults flipped Binary → netCDF
   (CloudCoeff_ReadFile/WriteFile/InquireFile and the analogous wrapper
   modules). External callers reading .bin files through these routines must
   now pass netCDF=.FALSE. explicitly.

4. CRTM_ChannelInfo_Subset hard-fails on duplicate or non-member channel
   lists (previously silent misbehavior: stalled merges and silently
   deactivated channels).

5. DDA-ARTS cloud optics: ICE_CLOUD now scatters (the legacy
   non-scattering shortcut applies only to Mie-TAMU tables), and its default
   DDA habit changed from IceSphere to IconCloudIce. Users of DDA-ARTS
   CloudCoeff tables with ICE_CLOUD in their profiles will see different
   brightness temperatures (validated against a sub-mm sounder; tropical O−B
   at 325 GHz moved from ~+13 K to ~−0.6 K). Default (Mie-TAMU) cloud optics
   are bit-identical.

6. PARMIO is presence-activated; TELSEM2 is opt-in. Placing
   PARMIO.MWwater.EmisCoeff.nc in the coefficient directory switches MW
   water emissivity physics at and above 200 GHz; removing it restores FASTEM.
   CRTM_Init prints an INFORMATION message when a sensor with ≥ 200 GHz
   channels initializes without the PARMIO LUT. The TELSEM2 MW-land atlas, by
   contrast, is never loaded unless requested (Use_MWland_Atlas=.TRUE.
   or an explicit MWlandCoeff_File); a present-but-not-requested atlas file
   is ignored. Note that with TELSEM2 opted in, all land-parameter Jacobians
   (LAI, vegetation, soil moisture, soil/land temperature) are zero; the
   atlas is a climatology and does not depend on them; the analytic
   NESDIS_LandEM Jacobians apply only when the atlas is not loaded.

   The 200 GHz figure is a safety floor, not physics. It was placed where the
   traditional sounding sensors stop so that enabling PARMIO could not disturb
   operational channels. Options%Use_PARMIO_MWSSEM drops the floor and runs
   PARMIO everywhere its table has data. Table coverage is checked separately
   and is never relaxed, including when a caller opts in, because the
   alternative is a confident number computed at the wrong frequency: the
   interpolator otherwise clamps silently, and a 204.78 GHz channel was being
   evaluated at 229 GHz. Channels where PARMIO is wanted but uncovered fall
   back to FASTEM.

7. OpenMP threading. CRTM_Init reads OMP_NUM_THREADS at run time; if
   it is unset or empty, CRTM defaults to a single thread via
   OMP_SET_NUM_THREADS(1). Because that call is process-global, it also
   affects OpenMP regions of the host application after CRTM_Init — export
   OMP_NUM_THREADS explicitly in threaded host applications (DA systems
   embedding libcrtm). The configure-time capture of OMP_NUM_THREADS (and
   the per-test ENVIRONMENT overrides) are gone: the environment at run time
   is what counts.

8. Binary coefficient files are on the way out. The v3.2.0 fix tarball
   ships no .bin coefficient files and the test suite no longer exercises
   the binary coefficient read path. The binary readers remain in the library
   for users with existing binary trees, but they should be considered
   deprecated (removal expected in a later v3.2.x, per the README).

9. Duplicate _j2 sensor aliases removed from the fix tree. Six
   sensors shipped twice under both a _j2 and an _n21 name for the same
   satellite (JPSS-2 = NOAA-21). Five of those pairs are identical in every
   data variable; the sixth (atms_j2 / atms_n21) agrees to
   floating-point round-off only (largest difference 5.7e-14 in
   Frequency, with the derived Planck and band-correction coefficients
   differing in their last bits from independent computation), which is
   radiometrically identical but not literally bit-identical. Otherwise the
   pairs differ only in the internal Sensor_Id string and the creation
   timestamp. A seventh entry, v.viirs-m_j2, had no _n21 counterpart in
   v3.1.4 and is a straight rename rather than a de-duplication. The _j2
   copies are gone; use the _n21 name:

   removed	use instead
   atms_j2	atms_n21
   atms_j2-srf	atms_n21-srf
   cris-fsr_j2	cris-fsr_n21
   viirs-i_j2	viirs-i_n21
   viirs-m_j2	viirs-m_n21
   v.viirs-i_j2	v.viirs-i_n21
   v.viirs-m_j2	v.viirs-m_n21

   cris-fsr_j2.NLTECoeff.nc went with them (identical to
   cris-fsr_n21.NLTECoeff.nc, and orphaned once its SpcCoeff was removed).

   CRTM_Init resolves coefficients by filename, so any caller configured
   with a _j2 sensor id must be updated or initialization will fail. The
   duplication was a maintenance hazard as much as dead weight: nothing in the
   tree recorded that the two names were meant to be twins, so regenerating
   one would have silently left the other stale.

   Note this does not apply to the visible-channel files that share
   content across detector variants (v.imgrD1..D8_gNN, v.sndrD1..D4_gNN,
   v.mi-l/m_coms). Those are genuinely distinct sensors whose parent IR
   channels differ; they share a common visible channel by instrument design
   and are all retained.

10. Twelve further sensor renames. These are name changes only: in every
    case the replacement file is identical to the removed one in every data
    variable, and only the filename and the internal Sensor_Id differ.
    CRTM_Init resolves coefficients by filename, so a caller configured with
    an old name will fail to initialize and must be updated.

    removed	use instead	why
    viirs-i_j1	viirs-i_n20	JPSS-1 became NOAA-20 at launch. Every one of these files already carried WMO_Satellite_Id 225, which is NOAA-20 in WMO C-5, so the _j1 name contradicted the file's own content.
    viirs-m_j1	viirs-m_n20	as above
    v.viirs-i_j1	v.viirs-i_n20	as above
    v.viirs-m_j1	v.viirs-m_n20	as above
    v.viirs-dnb_j1	v.viirs-dnb_n20	as above
    mwi_metop-sg-a1	mwi_metop-sg-b1	platform correction. MWI flies on Metop-SG-B, not Metop-SG-A.
    tms_tomorrow-s01_v4	tms_tomorrow-s01_v4-STAR	lineage relabel, so the v4 delivery is tagged to the organization it came from. Six sensors, s01 through s06.

    The six tms_tomorrow-sNN_v4 entries follow the same pattern and are not
    listed individually.

11. Five products withdrawn. Two were duplicates under a nonsensical name
    and three were never-flown or notional instruments:

    withdrawn	why
    airs_g13	identical in every data variable to airs281_aqua, which still ships. Its own WMO_Satellite_Id is 784 (Aqua) and its sensor id is 420 (AIRS), so the g13 suffix never described the content.
    iasi_g13	identical in every data variable to iasi616_metop-a, -b and -c, all of which still ship and which are distinguished from each other only by their WMO satellite ids (4, 3 and 5), as they should be for one instrument design on three platforms. iasi_g13 carried WMO satellite 1022, which identifies no platform. iasi_g13.NLTECoeff.nc went with it.
    ssmis_f20	DMSP F-20 was cancelled and never launched. The file carried WMO_Satellite_Id 1023, the invalid-value sentinel, because no satellite id was ever assigned.
    zssmis_f20	the Zeeman companion to the above, withdrawn with it.
    atms-ng_v1	a notional next-generation ATMS with 1169 channels and placeholder WMO ids (satellite 1, sensor 1). No such instrument exists.

    Nothing that still ships is lost by any of these: users of airs_g13 or
    iasi_g13 should switch to the correctly named file, which holds the same
    numbers.

12. Twelve unnamed CloudCoeff development artifacts removed. The v3.1.4
    tree carried test_new.bin_type0 through test_new.bin_type10 and
    test_new.bin_MIESNOW under CloudCoeff/Little_Endian/, and the netCDF
    transition converted them along with everything else. They carry no title,
    history or comment attribute of any kind, nothing in the library or the
    test suite references them, and their names describe a conversion run
    rather than a product. Removing them takes about 540 MB off the tarball.
    The named cloud tables are all retained, including the microphysics-scheme
    variants (CloudCoeff.GFDLFV3, CloudCoeff.Thompson08, CloudCoeff.WSM6)
    and the TAMU tables, none of which the test suite exercises either.

13. OMPS replaced by per-platform NOAA-20 and NOAA-21 products. The two
    shipped OMPS files were unusable and mislabelled, and have been retired in
    favor of four regenerated products:

    removed	replaced by
    u.omps-npAllFOV_j2	u.omps-np_n20 (151 ch), u.omps-np_n21 (158 ch)
    u.omps-tcAllFOV_j2	u.omps-tc_n20 (196 ch), u.omps-tc_n21 (198 ch)

    Three separate defects motivated this:

    • The files could not be loaded at all. Both carried
      Group_Index=4, which has been Zeeman-reserved with zero components
      since 2008, so CRTM_Predictor_Create failed outright. The replacements
      are Group_Index=8 (GROUP_UV_NO2), the UV variant carrying a scene-NO2
      component, and all four now pass CRTM_Init.
    • The platform labels were wrong, in opposite directions.
      u.omps-npAllFOV_j2 was labelled NOAA-21 (WMO 226) but its channel set
      matches the NOAA-20 grid (rms 0.04 nm, max 0.07 nm; every other
      platform's grid is at least 7 times farther); NOAA-21's nadir profiler
      natively has 158 channels reaching 245 nm, not 151. u.omps-tcAllFOV_j2
      really was NOAA-21 content, but indexed with a three-channel offset.
    • Only one platform was represented where two instruments exist.

    Channel numbering now follows each platform's own SRF. For total column,
    old channel N corresponds to u.omps-tc_n21 channel N+3; the old file
    omitted n21 channels 1 to 3 (298.1 to 298.9 nm) and extended three channels
    past its red end. Channel selections carried over from the old files must
    be re-mapped, not reused.

    The per-platform channel sets are verified against primary sources: the
    JPSS NOAA-21 OMPS SDR validated-maturity record (nadir profiler 158
    channels, nadir mapper 198) and the published instrument table in Yan et
    al. 2024, doi:10.3390/rs16234488 (nadir profiler SNPP 147 / NOAA-20 151 /
    NOAA-21 158; nadir mapper 196 / 196 / 198). Note the same record flags
    NOAA-21 nadir mapper radiances below 302 nm (roughly u.omps-tc_n21
    channels 1 to 10) as not validated for operational use.

14. RTSolution%Radiance is now the channel-polarized measurement when
    n_Stokes > 1, not Stokes I. The emergent Stokes vector is projected onto
    the channel's polarization, so a vertically polarized channel reports I+Q
    where it previously reported I. Brightness_Temperature, which is derived
    from it, moves with it. RTSolution%Stokes is unchanged and still holds the
    physical (I, Q, U, V). Anyone already running n_Stokes > 1 on a polarized
    channel will see the reported radiance and brightness temperature change by
    the polarization difference, which over ocean is order 20 percent of the
    signal. The projection weights are taken from the scalar path's own
    polarization handling, so the two now agree: a run with the channel forced
    to pure vertical or pure horizontal reproduces the corresponding scalar
    radiance to machine precision. Nothing at n_Stokes = 1 changes.

15. RTSolution_AD%Radiance and RTSolution_K%Radiance are now honoured as
    input seeds on the vector path. Previously %Radiance was an output alias
    for Stokes(1) but not an input one, and seeding it for an n_Stokes > 1
    run silently produced a zero Jacobian; only %Stokes or
    %Brightness_Temperature worked. Both now work. Code that seeded
    %Radiance and %Stokes(1) together to work around this will now double
    count.

16. RT_Algorithm_Id = RT_SOI with n_Stokes > 1 is now an error.
    Previously the vector branch was taken before the algorithm selector was
    consulted, so the caller silently received ADA results labelled as SOI. SOI
    has no vector solver, so this is now rejected with a message naming RT_ADA.

17. CRTM_MWwaterCoeff_Load_FASTEM discards the previously loaded scheme and
    reports failure. Switching scheme, for example FASTEM6 to FASTEM4, used to
    leave the shared coefficient structure deallocated while the function
    returned SUCCESS, because the underlying setter rejects a shape mismatch by
    destroying the target. Scheme switching now works, and a failed load returns
    FAILURE instead of SUCCESS. Related: a coefficient dimension mismatch used to
    terminate the program with a Fortran runtime formatting error rather than
    reporting cleanly.

18. Surface reflectance from category tables is clamped to the physical
    range. The 4-point Lagrange interpolation across SEcategory reflectance
    spectra overshoots [0,1] where the tabulated spectrum has sharp structure.
    The NPOESS VIS snow table holds exact zeros at 4000 and 5000 cm-1 with
    positive neighbours, which produced a reflectance near -0.03 at 2.25
    micron for both snow types, a negative top-of-atmosphere radiance (about
    -0.17 mW/(m2 sr cm-1) for VIIRS M11 over old snow), and a NaN brightness
    temperature from the inverse Planck. The interpolant is now clamped in
    SEcategory_Emissivity, and the visible-path direct-reflectivity limiter
    clamps below zero as well as above one. The defect is present in v3.1.4
    (reproduced end to end there), so this fixes an inherited problem rather
    than a 3.2.0 regression. Results change only where the clamp engages;
    in-range channels are untouched.

19. CRTM_VISsnowCoeff_Load returns FAILURE for an unrecognised
    classification prefix. The classification is parsed from the filename
    text before the first dot (NPOESS or SNICAR); any other prefix, or a
    filename with no dot at all, used to print a FAILURE message and return
    SUCCESS, so CRTM_Init succeeded with no visible snow table loaded and
    the snow surface optics quietly returned whatever the SfcOptics structure
    already held. Both arms now return FAILURE, and
    Compute_VIS_Snow_SfcOptics reports FAILURE if it is ever reached with
    neither table loaded.

20. CRTM restores the caller's OpenMP nesting policy. CRTM_Forward,
    CRTM_Tangent_Linear and CRTM_K_Matrix raise max-active-levels to run
    their nested channel loop. That setting is global to the OpenMP runtime and
    was never put back, so a host doing its own threading found its nesting
    policy silently replaced by a CRTM compute call, which can turn the host's
    own nested regions from serialised into thread-spawning. CRTM_Adjoint
    never sets the level, so it inherited whatever the previous forward or
    K-matrix call left behind. Each routine now saves and restores the value on
    every exit path. No computational code is touched and no result changes.
    The defect is present in v3.1.4, so this fixes an inherited problem rather
    than a 3.2.0 regression.

21. Channels are no longer split across threads when the split cannot pay for
    itself. Channel-level threading gives every thread its own optical-depth,
    surface-optics and RT scratch structures, sized by the layer count and
    stream maxima rather than by how many channels the thread receives. When
    threads outnumbered profiles, CRTM divided the channels among all spare
    threads regardless of how few each would get, and for small sensors that
    cost far more than it saved: a single 22-channel microwave profile on 16
    threads ran about 30 times slower than the same build on one thread. The
    thread count is now capped so each channel-thread owns at least
    MIN_CHANNELS_PER_CHANNEL_THREAD channels, and CRTM no longer spawns a
    thread team merely to count the available threads. Spare threads below the
    threshold are left idle deliberately.

    This only ever lowers the chosen thread count, so it cannot create nesting
    where there was none, and it is a no-op wherever profiles already absorb
    every thread. Results are unchanged; only the thread decomposition differs.
    The defect is present in v3.1.4 and is not a 3.2.0 regression.

    Who was affected: hosts that call CRTM with one profile at a time while
    also setting OMP_NUM_THREADS greater than 1. GSI does pass a single
    profile per call, so a hybrid MPI/OpenMP GSI configuration was exposed; a
    pure-MPI configuration with OMP_NUM_THREADS=1 never entered the path.
    JEDI and UFO were never exposed, because they pass the whole observation
    batch as profiles.

    Measured effect on the single-profile case (forward, wall clock, against
    the same build on one thread): 22-channel microwave sensor on 16 threads
    0.03x to 1.00x and on 8 threads 0.10x to 1.00x; 399-channel infrared
    sounder on 8 threads 0.92x to 2.12x; 2211-channel infrared sounder on 8
    threads 1.96x to 2.79x. Cases where profiles already met or exceeded the
    thread count are unchanged at 4.4x to 5.4x.

Known issues and limitations

• Sub-mm thin frozen cloud (≈ 325 GHz): optically thin frozen-cloud
  layers can produce nonphysical TBs through the adding-doubling/MOM path
  (small-τ matrix conditioning with high phase-function truncation orders).
  Affects sub-mm scattering scenes only; under investigation.
• The FASTEM to PARMIO handover at 200 GHz is a step, not a blend. The two
  models are independent and are not reconciled at the boundary, so a sensor
  with channels either side of 200 GHz sees a discontinuity in ocean
  emissivity there. Measured at -1.42 K mean in brightness temperature for
  TROPICS channel 12. This is a consequence of switching models at a frequency
  rather than a defect in either one, and it is why the floor sits where no
  operational sounding channel crosses it. PARMIO's own published validation
  stops at 165.5 GHz, so its whole default dispatch range is above the range
  its authors validated; the table labels that band
  extrapolated-experimental in its confidence_label variable, and callers
  should read that label rather than assume the table is uniform.
• SNICAR visible snow LUT (opt-in): two known limitations, detailed with
  the feature under Highlights: the table's angle dimension is labelled
  "Solar Zenith Angle" in the file but is interpolated at the RT
  view/quadrature angles (the solar zenith angle never reaches the table),
  and the forward path applies no LUT bounds guard while the tangent-linear
  and adjoint return exactly zero out of bounds. Both are deferred beyond
  v3.2.0; the angle question is pending confirmation with the table's
  author before either the file metadata or the surface-optics dispatch is
  changed.
• ODSSU netCDF supports the ODPS sub-algorithm only (the shipped SSU
  files are ODPS-based; ODAS-based SSU coefficient files remain binary-only).
• Options binary I/O does not persist the new fields (n_Stokes and the
  radiance-profile switches); they will be added with the next format
  revision.

Polarimetric (Options%n_Stokes > 1) radiative transfer

The vector path is functional and internally verified, and its surface sign
convention has been checked against RTTOV, but the emergent radiance has not
been validated end to end against an independent radiative transfer model or
against observations. Treat it as a capability under development rather than
a production-ready product. The specific limitations follow.

• No external reference for the emergent radiance. Every check on the
  radiance itself compares CRTM against CRTM: tangent-linear against finite
  differences, the adjoint dot-product identity, K against AD, physical
  invariants such as the polarization bound, and agreement with the scalar path
  in the limits where the two must agree. None of that can catch a convention
  that is consistently wrong. One such convention has now been settled
  externally: the sign of the third and fourth Stokes components at the
  surface was compared against RTTOV's FASTEM5 and matches at every relative
  wind azimuth, so U and V may be used quantitatively without first
  establishing the sign yourself. That validates one interface. The solver, the
  polarized phase matrix and the transport still have no independent reference,
  so a top-of-atmosphere U or V remains unvalidated end to end.
• The default microwave water surface has no polarimetric model. The
  CRTM_Init default is FASTEM6, whose azimuth model parameterises the
  vertical and horizontal components only and returns the third and fourth
  Stokes components as identically zero. A polarimetric surface requires
  MWwaterCoeff_Scheme = 'FASTEM4' (or FASTEM5), or PARMIO, which is used
  automatically at and above 200 GHz when its lookup table is present. Either
  argument selects the model: MWwaterCoeff_Scheme is the direct selector, and
  MWwaterCoeff_File is now honoured as one too, with the scheme recovered
  from the leading component of the file name. An explicit scheme wins if the
  two disagree, and the disagreement is reported rather than resolved silently.
  In v3.1.4 and earlier, MWwaterCoeff_File selected nothing at all and a
  request for FASTEM4 through it silently left FASTEM6 loaded.
• Microwave only. The coupled (V,H) to Stokes surface branch exists only in
  the microwave section of the surface optics. Infrared and visible sensors
  populate the first Stokes component alone, so a vector run there returns
  Q = U = V = 0 from the surface regardless of geometry.
• Aerosols contribute no polarization. The shipped AerosolCoeff carries a
  single phase element, so a vector run mixes polarized cloud scattering with
  unpolarized aerosol scattering. This is deliberate and is not blocked, since
  a scalar aerosol table must not prevent a polarized cloud run.
• The surface reflects no U or V. Both FASTEM and PARMIO set the third and
  fourth Stokes reflectivity components to zero, so U and V are emitted by the
  surface but never reflected. This is exact for clear sky, where the
  downwelling reaching a specular surface is unpolarized and symmetry about the
  meridional plane forbids a reflected U. It is an approximation once the
  downwelling is itself polarized by scattering.
• Polarimetric cloud lookup tables are not validated. The n_Stokes > 1
  scattering path requires a six-phase-element table (the experimental
  CRTM-Exp scheme). Those tables have not been validated for full-Stokes
  work, so polarized scattering results carry that uncertainty independently of
  the code.
• Phase-matrix normalization is inconsistent for below-diagonal polarized
  blocks. Normalize_Phase scales each row's intensity and polarized
  elements together and then applies an intensity-only symmetry copy, so a
  block below the diagonal ends up with its (1,1) element carrying one row's
  normalization and its polarized elements another. This does not produce a
  polarization-bound violation with the shipped coefficients, where the
  measured worst ratio is 0.65, but it is not the correct polarized symmetry
  treatment.
• Mixed-polarization channels are projected at the sensor angle. For the
  V/H-mixed, constant-mixed and PRA polarizations the vector path applies the
  polarization mixing once, to the emergent radiance at the sensor angle, which
  is where a receiver projects. The scalar path instead applies it to the
  surface emissivity at every quadrature angle. The two coincide when there is
  a single angle, and differ slightly for a scattering mixed-polarization
  channel.
• plus45L, minus45L, RC and LC polarizations are treated as
  vertical, inherited unchanged from the scalar path so that the two agree.
  These are placeholders, not the true projections.
• RT_Algorithm_Id = RT_SOI is not supported with n_Stokes > 1 and is now
  rejected; see the behavior changes above. SOI has no vector solver.

[BenjaminTJohnson]

Release-impacting update to the coefficient data section: after the
2026-08-01 tarball roll, the ABI assumption audit replaced the ENTIRE ABI
ODPS family in the staging tree (abi_g19 on 2026-08-03; abi_g16, abi_gr,
abi_g17, abi_g18, abi-81K_g17 on 2026-08-04; all Version 2, generated by
crtm-coeffgen at the 2026.5 gas epoch with NCC-authenticated per-flight-model
SRFs). Rationale and validation are in issue #347 and its comments; the
headline defect was the shipped abi_g19 being a 2024 test article with a
-1.67 K CO2-band O-B bias and unphysical window-channel water Jacobians,
present since REL-3.1.2.0.

Consequences for this release:

1. The rolled tarball (md5 7cd36fb18e3c69d5f4399a31009cc4ce) NO LONGER
   MATCHES the staging tree. It needs to be re-rolled and the file-by-file
   verification repeated before publishing. Twelve files differ (six
   TauCoeff/ODPS/netCDF and six SpcCoeff/netCDF ABI files).
2. abi_g18 is in the standard regression suite: its FWD/TL/AD/K baselines
   must be reseeded after the coefficient swap (expected TB movement up to
   ~0.65 K at channel 16; a coefficient improvement, not a code change).
   The release-notes line about abi_g18 WMO ids (272/617 vs -999 sentinels)
   is now moot: the new file carries the correct ids natively.
3. Suggested release-notes addition for the coefficient section: "The
   complete ABI family (GOES-16/17/18/19, the generic GOES-R set, and the
   GOES-17 81K warm-focal-plane variant) was regenerated at the current
   well-mixed-gas epoch with authenticated per-flight-model SRFs, replacing
   heritage files that under-responded to present-day CO2; the GOES-19 file
   additionally replaces a 2024 test article (issue abi_g19 ODPS coefficients are a test article: CO2-epoch bias and spurious stratospheric water Jacobians; validated replacement available #347)."

Developer detail has been added to REL-3.2.0_changes_vs_develop.md (new
section 13) and the per-sensor rows in REL-3.2.0_coefficient_inventory.md
are updated, including resolution of the abi_g19 "flagged 2026-08-01"
verification note.

[BenjaminTJohnson]

Release-impact update (2026-08-05, supplementing the earlier comment):

1. The ABI family in the fix tree is now TauCoeff Version 3 (OPTRAN effective-target fix; hashes in the revised body of abi_g19 ODPS coefficients are a test article: CO2-epoch bias and spurious stratospheric water Jacobians; validated replacement available #347). The tarball re-roll requirement stands and now covers these files; any tarball rolled before 2026-08-05 is stale.
2. cris-fsr_n21.NLTECoeff.nc was replaced 2026-08-05 with the post-fix regeneration (Version 2; byte md5 72681c2f3306c9e92393302a99442c19; prior Version 1 archived at test-data-release/staged_backup_20260805_cris_nlte with MD5SUMS).
3. HOLD for the release roll: the following staged ODPS files are known-defective (details in the regeneration campaign records) and their replacements are being generated and gated now -- viirs-i_j4, viirs-m_j4, mersi2_fy3d, mersi3_fy3f, agri_fy4a, agri_fy4b, gxi_geoxo, metimage_metop-sg-a1, irs_mtg-s1. Recommendation: do not roll the release tarball until their dispositions are decided; each replacement arrives with self-fit, Jacobian, field-diff, and O-B (native or proxy) evidence plus exact hashes.
4. abi_g18 FWD/TL/AD/K regression baseline reseeding is still owed before the suite will pass against the new coefficients.

[BenjaminTJohnson]

Release-impact update (2026-08-06): the 2026-08-05 HOLD is CLEARED. All dispositions are decided and executed; the tree is ready for the tarball re-roll.

1. The nine held sensors (viirs-i_j4, viirs-m_j4, mersi2_fy3d, mersi3_fy3f, agri_fy4a, agri_fy4b, gxi_geoxo, metimage_metop-sg-a1, irs_mtg-s1) were replaced 2026-08-06 with full regenerations (SpcCoeff + TauCoeff pairs, Version bumped): epoch-2026.5 global-factor training gases, OPTRAN fitted to the effective water target, Jacobian-aware OPTRAN acceptance. Every pair passed self-fit, K-matrix Jacobian vs the replaced file, field-diff with documented exceptions, consistency (incl. NLTE pairing for IRS), CRTM load, and O-B (native for IRS on real MTG-S1 radiances; channel-matched proxy for the not-yet-flying imagers, waivers recorded).
2. Six OPTRAN-refresh TauCoeffs were also replaced: mwi_metop-sg-b1, msi_earthcare, and the four VIS NO2-capable products v.tempo_is40e, v.oci_pace, v.olci_s3a, v.olci_s3b. For the VIS four, the staged Group-8 (GROUP_UV_NO2) products embedded the pre-fix raw-target OPTRAN; they were rebuilt by re-running the NO2 merge on the refreshed bases, so the NO2 component is bit-identical to what was staged in July and only the OPTRAN water-line block changed (Version 1 to 2). All four rebuilt files load and run through CRTM K cleanly.
3. cris-fsr_n21 Spc/TauCoeff remain the heritage files by decision (observation-space verdict on 611 real CrIS superobs favored heritage); the post-fix NLTECoeff replacement from 2026-08-05 stands and pairs with the heritage SpcCoeff.
4. Accounting: 24 files replaced; prior staged files archived with MD5SUMS at LBL/work/regen_2026-08/staged_backup_20260806; two-hash ledger (byte md5 + variable-content sha) regenerated at LBL/work/coeff_ledger/LEDGER.tsv with every staged file verified identical to its validated source.
5. Still owed before release close: tarball re-roll (any tarball rolled before 2026-08-06 is stale), post-roll solar + ledger re-sweep of the extracted tarball, and the abi_g18 FWD/TL/AD/K regression baseline reseed.