Skip to content

CLAS Brush-up Baseline (2026-07-14)

This note freezes the behavior and evidence contracts used by the CLAS responsibility-splitting work. The brush-up phase must not change these results unless a later, explicitly-scoped model PR updates the baseline.

Source

  • Commit: 1d27bc3 (develop)
  • Worktree: clean before the baseline run

Lightweight regression gates

python3 -m pytest -q \
  tests/test_clas_a4b_native_selfdiff.py \
  tests/test_clas_zd_component_diff.py \
  tests/test_clas_zd_component_summary.py \
  tests/test_claslib_osr_zd_export.py \
  tests/test_claslib_zd_component_export.py \
  tests/test_optional_clas_zd_component_diff.py \
  tests/test_cli_tools.py -k 'qzss_l6 or clas_ppp'

Result: 47 passed, 16 skipped, 208 deselected.

./build/tests/run_tests --gtest_filter='*Clas*:*CLAS*'

Result: 36 passed from six test suites.

Public-data A4b evidence gate

python3 scripts/ci/run_clas_a4b_native_selfdiff.py

Result: passed.

Metric Baseline
Native epochs 300
Native PPP solution rate 100.0%
Native code-dump rows 6550
GPS L2W exact observation rows 300
GPS L2W observation fallback rows 0
GPS L2W exact bias rows 300
GPS L2W bias fallback rows 0
Self-diff common rows 300
Self-diff compared components 16080
Self-diff maximum absolute delta 0.0 m
Self-diff unmatched rows 0

Stable generated-artifact checksums for this source revision:

  • native_code_dump.csv: 0d767bc50c9e170cbb6bad4d9c14716c20448c0b93434c64c4bcfd5af0689427
  • selfdiff.json: 11fccc68ef1d84668a145815034161d647d44dadd82ef4fca2bf270994d7bb7b

The top-level CI summary embeds absolute artifact paths, so its own checksum is not a portable regression contract. Compare its status and metrics instead.