CLAS Public Validation Datasets

Current develop status (2026-06-19): the current CLI no longer exposes the historical --claslib-parity entry point. Native CLAS runs use gnss ppp --clas-osr directly, or gnss clas-ppp --profile clas, which forwards --clas-osr after expanding QZSS L6/Compact SSR inputs. The iter55 millimetre-level table below is retained as historical reference; the active default-flip decision remains the A5 STOP described in clas_dd_filter_a5.md.

This page records the public CLAS datasets used for the native CLASNAT parity checks. It separates accepted regression gates from public exploratory runs that are useful for robustness work but are not yet millimetre-level passes.

Public Source

The current CLAS parity data comes from the public CLASLIB repository:

https://github.com/QZSS-Strategy-Office/claslib.git

The local validation copy used during iter55 was:

Its Git remote was confirmed as:

origin  https://github.com/QZSS-Strategy-Office/claslib.git

The files below are from that repository's data/ directory.

Accepted Regression Gate

Case	Public files	Window	Native mode	Acceptance status
2019-08-27, QZSS CLAS sample	0627239Q.obs, sept_2019239.nav, 2019239Q.l6	2019-08-27, 16:00 GPST window	Historical iter55 --claslib-parity native CLASNAT path	Historical accepted gate. Iter55 measured 0.00556 m RMS 3D against the CLASLIB trajectory over the last 100 epochs of a 2000-epoch run, and 0.01014 m RMS 3D over the last 100 epochs of a 300-epoch smoke window.

This is the dataset behind the historical --claslib-parity mm-level regression. It is public because the input files are shipped in the public QZSS-Strategy-Office/claslib repository, not because they are private local capture files.

The same 2000-epoch iter55 run measured 0.00251 m RMS 3D against the fixed ECEF reference (-3957235.3717, 3310368.2257, 3737529.7179) over its last 100 epochs.

Public Exploratory Runs

Case	Public files	What was checked	Result
2018-11-25, CLASLIB bundled sample	0161329A.obs, tskc2018329.nav, 2018328X_329A.l6	CLASLIB reference generation and 300-epoch native CLASNAT run	CLASLIB reference solved, but native CLASNAT stayed in fallback status for the checked window. Direct trajectory diff against CLASLIB was about 1.336 m RMS 3D over the last 100 aligned epochs, so this is not an accepted <20 mm gate.
2019-12-15, CLASLIB bundled ST12 sample	0627349AB.bnx, sept_2019349.nav, 2019349B.l6	CLASLIB reference from BINEX, then convbin to RINEX 3.02 with a GNSS signal mask before native CLASNAT	CLASLIB reference solved, but native CLASNAT did not reach the mm gate in the checked window. This remains an input-format and robustness probe, not an accuracy gate.

No additional bundled CLASLIB public sample reached <20 mm native CLASNAT RMS in iter55. Those probes are still useful because they confirm that the additional samples are public and reproducible, and they expose the next robustness gaps without weakening the accepted 2019-08-27 gate.

Reproduce The Current Native CLAS Run

Clone CLASLIB and point the commands at its public data/ directory:

git clone https://github.com/QZSS-Strategy-Office/claslib.git /tmp/claslib
DATA=/tmp/claslib/data

Build the current native PPP binary:

cmake -S . -B build
cmake --build build --target gnss_ppp -j4

Run the current CLAS wrapper. --profile clas expands the raw QZSS L6 file and forwards the resulting sampled corrections to gnss ppp --clas-osr:

python3 apps/gnss.py clas-ppp \
  --profile clas \
  --obs "$DATA/0627239Q.obs" \
  --nav "$DATA/sept_2019239.nav" \
  --qzss-l6 "$DATA/2019239Q.l6" \
  --qzss-gps-week 2068 \
  --out /tmp/gnsspp_2019239.pos \
  --summary-json /tmp/gnsspp_2019239_summary.json \
  --max-epochs 300

gnss clas-ppp also forwards --antex <antennas.atx> and --receiver-antenna-type <type> to the underlying gnss ppp run for receiver/satellite antenna parity probes. The public 0627239Q.obs header does not declare a receiver antenna type, so A4b receiver-antenna probes should pass the native ANTEX antenna key explicitly, for example --receiver-antenna-type "TRM59800.80 NONE". The CLASLIB ANTEX file uses CRLF line endings, so the native ANTEX loader trims trailing carriage returns before matching labels. Set GNSS_PPP_CLAS_RX_ANTENNA=1 to materialize those receiver terms into the CLAS OSR receiver_antenna_m component and the aggregate PRC/CPC terms. The default path leaves them unchanged while A4b parity is measured. The A4b probe also enables GNSS_PPP_CLAS_ATMOS_GRID_MATRIX=1 and GNSS_PPP_CLAS_ATMOS_LIFECYCLE=1 so native STEC/trop materialization follows the CLASLIB-style grid selection and regenerated atmosphere-token lifecycle. It also sets GNSS_PPP_CLAS_TROP_CLIMATOLOGY=1, which limits the CLAS grid-trop hydro/wet zenith-baseline change to this A4b parity lane while default CLAS output keeps the legacy constants.

For the A4b GPS L2W identity probe, enable the diagnostic gates and write a native zero-difference code component dump:

GNSS_PPP_CLAS_DD_FILTER=1 \
GNSS_PPP_CLAS_CODE_ROW_PARITY=bias,full-prc \
GNSS_PPP_CLAS_RX_ANTENNA=1 \
GNSS_PPP_CLAS_ATMOS_GRID_MATRIX=1 \
GNSS_PPP_CLAS_ATMOS_LIFECYCLE=1 \
GNSS_PPP_CLAS_TROP_CLIMATOLOGY=1 \
GNSS_PPP_CLAS_CODE_DUMP=/tmp/clas_a4b_native_code_dump.csv \
python3 apps/gnss.py clas-ppp \
  --profile clas \
  --obs "$DATA/0627239Q.obs" \
  --nav "$DATA/sept_2019239.nav" \
  --qzss-l6 "$DATA/2019239Q.l6" \
  --qzss-gps-week 2068 \
  --antex "$DATA/igs14_L5copy.atx" \
  --clas-atmos-selection freshness-first \
  --receiver-antenna-type "TRM59800.80     NONE" \
  --compact-code-bias-composition-policy base-only-if-present \
  --compact-code-bias-bank-policy delayed-15s-bank \
  --compact-bias-row-materialization selected-satellite-base-extend \
  --out /tmp/gnsspp_clas_a4b_native.pos \
  --summary-json /tmp/gnsspp_clas_a4b_native_summary.json \
  --max-epochs 300

Self-diff the dump to inspect identity provenance before comparing against a CLASLIB-side component dump:

python3 scripts/analysis/clas_zd_component_diff.py \
  /tmp/clas_a4b_native_code_dump.csv \
  /tmp/clas_a4b_native_code_dump.csv \
  --base-label native \
  --candidate-label native \
  --row-type code \
  --json-out /tmp/clas_a4b_native_code_selfdiff.json \
  --details-csv /tmp/clas_a4b_native_code_selfdiff.csv

python3 - <<'PY'
import json
payload = json.load(open("/tmp/clas_a4b_native_code_selfdiff.json"))
print(json.dumps(payload["identity_provenance"], indent=2, sort_keys=True))
PY

Disposable CLASLIB dumps usually use numeric sys/prn/RTKLIB code fields instead of native G14/C2W-style row keys. Normalize those dumps before using them as the oracle side of the component diff:

python3 scripts/analysis/claslib_zd_component_export.py \
  /tmp/claslib_zd_dump.csv \
  --output /tmp/claslib_zd_dump.normalized.csv \
  --stage-label post

python3 scripts/analysis/clas_zd_component_diff.py \
  /tmp/claslib_zd_dump.normalized.csv \
  /tmp/clas_a4b_native_code_dump.csv \
  --base-label claslib \
  --candidate-label native \
  --stage post \
  --row-type code \
  --sat G14 \
  --freq 1 \
  --rinex-code C2W \
  --duplicate-policy mean \
  --json-out /tmp/clas_a4b_native_vs_claslib_code.json \
  --details-csv /tmp/clas_a4b_native_vs_claslib_code.csv

Unmodified CLASLIB rnx2rtkp can also write an .osr file when run with its OSR output option. That file carries OSRRES(ch*) rows with tow, sys, prn, and L1/L2/L5 component columns, but no GPS week or exact observation code. For the current A4b GPS L2W probe, the exporter maps GPS slot 2 to C2W/L2W; provide the GPS week explicitly so the diff key is not ambiguous. The normalized iono_scaled_m field is reconstructed from the CLASLIB PRC closure instead of the raw .osr iono display column, so it matches the correction actually applied in PRC1/PRC2/PRC5. iono_l1_m is the frequency-slot-specific L1-equivalent of that reconstructed value, not a copy of the L1 slot's display or closure value. The normalized stec_tecu field is derived from that L1-equivalent delay with the GPS L1 TECU-to-meter factor. The exporter also writes claslib_iono_source=prc_closure, claslib_raw_iono_l1_m, and claslib_raw_stec_tecu, so raw CLASLIB .osr display ionosphere can be inspected separately from the effective PRC-closure ionosphere used by the default diff. This lets the diff separate TECU materialization drift from frequency-scaling drift:

python3 scripts/analysis/claslib_zd_component_export.py \
  /tmp/claslib_solution.osr \
  --output /tmp/claslib_solution.osr.normalized.csv \
  --stage-label post \
  --gps-week 2068

On the 300-epoch 2019 sample smoke, current develop produces 5,350 native code rows. The GPS L2W slice has 300 rows, all with exact bias identity and exact observation matches, and zero observation-family or code-bias fallback rows. Use base-only-if-present with the latest preceding base bank for this A4b probe: the subtype-6 network code-bias rows are replacement rows for the subtype-4 base values, not additive deltas. On the same 300-epoch window, the direct network policy produced periodic G14/C2W code_bias_m outliers such as -0.70 m where CLASLIB reports +0.76 m. The A4b policy above restores those rows to the base value and reduces the G14/C2W code_bias_m diff from mean_abs=0.1795 m, rms=0.4000 m, max_abs=1.4600 m to mean_abs=0.1319 m, rms=0.3029 m, max_abs=0.7800 m.

The row materialization option also extends the latest base code-bias bank onto phase-only subtype-6 network rows. Without it, the 300-epoch G14/C2W smoke has 49/300 native rows with code_bias_present=0 and code_bias_m=0; with it, the same slice has 0/300 missing code-bias rows. On the short CLASLIB ZD oracle window bundled with this probe, that phase-only-row extension reduces the G14/C2W code_bias_m diff from mean_abs=0.0764 m, rms=0.2295 m, max_abs=0.7600 m to mean_abs=0.0091 m, rms=0.0135 m, max_abs=0.0200 m, and the aggregate PRC diff from mean_abs=0.0915 m, rms=0.2369 m, max_abs=0.7819 m to mean_abs=0.0243 m, rms=0.0266 m, max_abs=0.0419 m. The gated receiver-antenna materialization keeps the exact GPS L2W row on CLASLIB's receiver-antenna slot; on the same 300-epoch G14/C2W slice this reduces receiver_antenna_m from mean_abs=0.0249 m to mean_abs=0.0021 m. The gated trop-climatology baseline then reduces trop_correction_m from mean_abs=0.0283 m, rms=0.0284 m, max_abs=0.0289 m to mean_abs=0.0011 m, rms=0.0012 m, max_abs=0.0017 m. Aggregate PRC mean/RMS also improves from mean_abs=0.0292 m, rms=0.0324 m to mean_abs=0.0146 m, rms=0.0212 m; the remaining max_abs=0.0695 m is an iono/code-bias dominated row, not a trop regression. The remaining top-row deltas are therefore dominated by ionosphere and residual code-bias lifecycle terms rather than the 30-second code-bias sign flip, phase-only code-bias gaps, receiver-antenna slot mismatch, or fixed trop zenith constants.

The diff JSON also includes top_component_deltas, the largest individual component deltas across all matched rows, and top_row_component_breakdowns, which groups all component deltas for the same ZD key. Use those fields to distinguish the global largest component from the most unbalanced row before changing the gated correction model. The optional CI summary schema ci_optional_clas_zd_component_diff.v7 and later summaries lift both the global top delta, the leading row-breakdown component, and a per-component max-delta map into summary metrics, so release artifacts show the next single-component target without manually opening the full diff JSON. Schema ci_optional_clas_zd_component_diff.v11 additionally surfaces top-row values for highlighted components. Present-on-both components, such as stec_tecu, iono_l1_m, iono_scaled_m, code_bias_m, and trop_correction_m, carry CLASLIB/native/delta values in the CI summary; present-on-one-side components still expose the native atmosphere, clock, and code-bias reference epochs when the CLASLIB export cannot provide comparable fields. Schema ci_optional_clas_zd_component_diff.v12 also highlights CLAS atmosphere network/grid provenance when both normalized sides provide those columns; ci_optional_clas_zd_component_diff.v13 adds native code-bias reference TOW.

CI can regenerate the native side of this A4b probe without a pre-staged native CSV:

python3 scripts/ci/run_clas_a4b_native_selfdiff.py

The wrapper sparse-checks out the public CLASLIB data/ directory at the pinned GNSSPP_CLAS_A4B_CLASLIB_REF, runs the native command above, and self-diffs the G14/C2W code rows. It emits ci_clas_a4b_native_selfdiff_summary.json, native_code_dump.csv, native_code_dump_summary.json, native_summary.json, and selfdiff.{json,csv}. The wrapper first validates the native dump with clas_zd_component_summary.v2; the CI summary schema is ci_clas_a4b_native_selfdiff.v3. The summary uses status: passed only when the native run reaches the configured epoch count, the native dump summary passes and covers the dump rows, the dump summary independently verifies GPS L2W exact observation/bias identity with zero fallback rows, and the self-diff has no unmatched rows or component deltas. Set GNSSPP_CLAS_A4B_DATA_ROOT to use a pre-existing CLASLIB data checkout, or GNSSPP_CLAS_A4B_AUTO_FETCH=0 to record blocked_infrastructure instead of fetching public data. CI treats that blocked state as a failure by default; set GNSSPP_CLAS_A4B_FAIL_ON_BLOCKED=0 only for diagnostic-only local runs.

The optional CLAS ZD oracle/native diff runs after this native self-diff in CI. CI also builds pinned CLASLIB as a test-only tool, runs unmodified rnx2rtkp -s on the same public A4b window, normalizes its .osr file, and uploads ci_claslib_osr_zd_export_summary.json, claslib.nmea.osr, claslib_osr.normalized.csv, and claslib_osr_summary.json:

python3 scripts/ci/run_claslib_osr_zd_export.py

When GNSSPP_CLAS_ZD_CANDIDATE_CSV is unset, the workflow sets GNSSPP_CLAS_ZD_NATIVE_CSV=output/clas_a4b_native_selfdiff/native_code_dump.csv so the generated public-data native dump becomes the candidate side. When GNSSPP_CLAS_ZD_BASE_CSV is unset, the workflow uses output/claslib_osr_zd_export/claslib_osr.normalized.csv as the CLASLIB base side and defaults the optional diff filter to post/code/G14/f1/C2W; explicit workflow variables still override those defaults. The normalized CLASLIB CSV includes test-only grid provenance derived from clas_grid.def and the CLASLIB .osr receiver latitude/longitude, and the runner fails if GPS L2W rows are missing those provenance columns. When GNSSPP_CLAS_ZD_COMPONENTS is unset, the remote diff compares the component-level fields prc_m, prc_component_sum_m, prc_closure_residual_m, stec_tecu, iono_l1_m, iono_l1_from_stec_m, iono_l1_stec_closure_residual_m, iono_scaled_m, iono_scale, iono_scaled_closure_residual_m, trop_correction_m, code_bias_m, receiver_antenna_m, relativity_m, atmos_ref_tow, clock_ref_tow, code_bias_ref_tow, and atmos_clock_gap_s, plus CLAS grid id/count/weight provenance. atmos_grid_distance_m remains available for explicit diagnostic runs, but it is not part of the default component list because receiver-coordinate differences can otherwise dominate the top-row summary without representing a correction-value regression. The derived prc_closure_residual_m value is computed by the diff tool as PRC - (trop + relativity + receiver antenna + scaled iono + code bias), so it diagnoses PRC convention/rounding gaps without changing either input CSV schema. The derived iono_l1_stec_closure_residual_m is iono_l1_m - stec_tecu * GPS_L1_TECU_TO_METERS, which separates TECU-to-meter unit mistakes from upstream STEC materialization mismatches. The derived iono_scaled_closure_residual_m is iono_scaled_m - iono_scale * iono_l1_m, which separates frequency-scaling mistakes from upstream STEC/L1-delay mismatches. The aggregate applied_pr_corr_m remains available for explicit diagnostic runs, but the default sign-off points the top-row evidence at the underlying ZD correction component. The GitHub step summary highlights stec_tecu, iono_l1_m, iono_scaled_m, iono_l1_from_stec_m, iono_l1_stec_closure_residual_m, iono_scaled_closure_residual_m, code_bias_m, trop_correction_m, and prc_closure_residual_m from the per-component max-delta map, and reports missing counts for highlighted atmosphere-reference fields when CLASLIB cannot provide a comparable value. The native clas_zd_component_summary.v2 snapshot also reports numeric min/max stats for atmos_ref_tow, clock_ref_tow, code_bias_ref_tow, atmos_clock_gap_s, atmos_lifecycle_tow, atmos_selected_satellite_count, atmos_valid_grid_count, atmos_stec_grid_value_count, and atmos_selected_grid_stec_value_count, so reviewers can see which lifecycle bank was materialized before opening the raw CSV. For the leading-row breakdown, v11 summaries also include CLASLIB/native/delta values for highlighted present components and the native value of each highlighted missing field, which ties the largest PRC/iono/trop delta back to the selected reference epoch.

For the G14/C2W code-bias timing issue, native SSR interpolation forbids future code-bias samples and the A4b L6 expansion uses delayed-15s-bank so base-bank lookup uses the latest bank effective at tow - 15 s. Native therefore no longer applies the 230430 bank to TOW 230426..230429, and it holds the previous bank through the first half of each 30-second bank. On the CLASLIB .osr slice this moves G14/C2W code_bias_m from mean_abs=0.00857 m, rms=0.01309 m to mean_abs=0.00800 m, rms=0.01265 m, with mismatch count moving from 120/280 to 112/280 common rows. The L6 expansion then emits delayed base-bank refresh rows for selected subtype-6 network replacements, moving the same G14/C2W slice to 16/280 mismatches with mean_abs=0.00114 m and rms=0.00478 m; aggregate PRC moved to mean_abs=0.00879 m and rms=0.01481 m. Current develop regenerates the same public-data slice with zero code_bias_m delta on all 280 common G14/C2W rows, including the former 5-second boundary mismatches. CI hard-gates this component at 0.0 m while the remaining ionosphere and aggregate PRC differences stay diagnostic. The CLASLIB .osr snapshot summaries also include claslib_raw_iono_l1_m and claslib_raw_stec_tecu numeric stats as explicit diagnostics; they are intentionally outside the default native diff component set. If either generated side is missing, the optional diff reports blocked_infrastructure, not a passing sign-off.

Build with CLASLIB linked only when you need oracle-backed unit tests:

cmake -S . -B build-claslib \
  -DCLASLIB_PARITY_LINK=ON \
  -DCLASLIB_ROOT_DIR=/tmp/claslib
cmake --build build-claslib --target run_tests -j4

Run the oracle helper parity tests when that optional build is available:

./build-claslib/tests/run_tests --gtest_filter='*ClasnatParity*'

Expected helper coverage after iter55 is 17 tests, including filter, lambda, tropmodel, and stec_grid_data.