Skip to content

ppc_current_status.md

Latest commit

 

History

History
312 lines (250 loc) · 14.6 KB

ppc_current_status.md

File metadata and controls

312 lines (250 loc) · 14.6 KB
Raw

PPC Current Status

Stale (2026-10-02 note). This page describes the retired ranker / scorer contract (Phase71, 86.205492%). That number is not comparable with the current honest scorer (missing epochs stay in the denominator, #159), whose canonical value is 59.222040% (docs/ppc_causal_float_selector.md). For current PPC numbers use plan.md §2. The plan.md link below now points at the handoff entry point; the chronological log it refers to is archive/plan_history_2026-04_to_2026-07.md.

Last updated: 2026-05-19.

This is the short current-state document for PPC work. The long chronological log remains in plan.md.

Current Conclusion (2026-05-19)

Canonical production best: Phase71 OSM road route = 86.205492% OFFICIAL.

Phase71 keeps the Phase43 per-run conditional setup and adds an OSM road-centerline corrected candidate (xd_gici_osmroad_hs) only on nagoya/run2. The gain is isolated to n/r2; the other five production runs are unchanged.

run Phase43 Phase71 delta
tokyo/run1 90.841510% 90.841510% +0.000000pp
tokyo/run2 95.410067% 95.410067% +0.000000pp
tokyo/run3 88.949168% 88.949168% +0.000000pp
nagoya/run1 83.700284% 83.700284% +0.000000pp
nagoya/run2 64.426589% 65.669779% +1.243190pp
nagoya/run3 92.662146% 92.662146% +0.000000pp

Official average:

Phase43: 85.998294%
Phase71: 86.205492%
Delta: +0.207198pp
TURING 85.6% delta: +0.605492pp

Key implementation files:

  • experiments/materialize_phase70_osm_road_centerline_candidate.py
  • experiments/build_phase70_osm_road_ranker_overlay.py
  • experiments/scripts_run_phase70_osmroad_neutral_check.sh
  • experiments/scripts_run_phase71_osmroad_production.sh

The Phase71 production script regenerates the OSM candidate and ranker overlay under /tmp by default. The prep-only smoke command reproduced triggered_epochs=359, good/bad=30/0, and overlay added_rows=359 after the artifact-regeneration change.

GICI rtk_imu_tc candidate injection — selector-layer probe (2026-06-01)

GICI forppc2024 rtk_imu_tc NMEA outputs were verified across all 6 runs at 95–100% <1 m (fix4 99–100%) by experiments/eval_gici_tc_ppc2024_batch.py (PR #72; +18 s leap, coverage-aware best-deployable pick per run). To test injection, experiments/materialize_gici_tc_ppc_candidate.py converts the best per-run NMEA into a selector-pool candidate (libgnss_diag_phase10/gici_tc/, .pos + synthesised gate-passing diag .csv — GICI exposes no RTK ratio/RMS, so the diag marks every emitted epoch as a trusted external fix; this is a deliberate "full-trust" probe).

sim_ppc_phase_csv_addcand.py --discover-diag-dirs --only-labels xd_gici_tc (policy phase11dd, base pool from ppc_ctrbpf_fgo_phase11dd_full_p2k_runs.csv):

run base +gici_tc Δ
tokyo/run1 66.40% 85.79% +19.39pp
tokyo/run2 84.86% 92.82% +7.96pp
tokyo/run3 80.03% 78.10% −1.93pp
nagoya/run1 64.33% 76.24% +11.91pp
nagoya/run2 40.21% 64.75% +24.54pp
nagoya/run3 59.15% 86.83% +27.68pp
aggregate 70.65% 81.17% +10.52pp

Including RTK-float (fix=5) epochs beats --fix4-only (+10.52pp vs +4.44pp): on this weak base, GICI float still beats the base fill.

Caveat — not yet a production win. This base (phase11dd, 70.65%) is far weaker than the Phase71 production pipeline (86.21% OFFICIAL). Compared against the production per-run numbers (tokyo1 90.84%, tokyo2 95.41%, tokyo3 88.95%, nagoya1 83.70%, nagoya2 65.67%, nagoya3 92.66%), full-trust gici_tc would regress 5/6 runs and roughly tie the OSM lever on the weak run (n/r2 64.75% vs 65.67%). The selector layer is also below the PF+ranker layer that yields OFFICIAL, and past one-layer gains have been absorbed by the full pipeline (cf. v3-ranker / Viterbi negatives).

Decisive next test: inject gici_tc into the production pool via the full exp_ppc_ctrbpf_fgo.py replay (Phase43/71 candidate set + ranker) — especially n/r2-targeted, where production is only ~65% — and measure OFFICIAL, rather than trusting the selector-layer delta. A fair (non-fabricated) gate for gici_tc is also open work, since it has no native ratio/RMS.

Resolution (2026-06-01): gici_tc is already fully in production; no upgrade exists

Tracing the production pool settled the question:

  1. gici_tc is already deployed. The Phase43/71 pool already carries 19 GICI candidates (xd_gici_def = libgnss_diag_phase19/gici_tc_esdfix, plus xd_gici_* for the gici_full_* variants). They are part of the 86.21% OFFICIAL — not an un-injected lever.
  2. The newly-batch-evaluated outputs are byte-identical to production. Same-config apples-to-apples (production gici_full_X vs forppc2024 gici_open_ws/test_nagoya2_X) matches exactly on every config (combo4 95.0%, window5 84.8%, hisnr 82.4%, … same n/p50/<1 m). The production candidates are these forppc2024 runs.
  3. The earlier "n/r2 +6.17 pp" was a config-mismatch artifact — production gici_tc_esdfix (def, 88.85%) vs the new combo4 (95.0%). But combo4 is already in the pool as xd_gici_c4 (95.0%). There is no better build to swap in, so a swap treatment is identical to baseline.
  4. The n/r2 selector mode ranker_gici_cluster_override is now committed (2026-06-17): implemented in experiments/ppc_gici_override.py and wired into exp_ppc_ctrbpf_fgo.py's argparse choices, per the documented Phase 43/71 rule (high-risk GICI pick → re-pick within the xd_gici family toward a tight ≥6-member 50 cm cluster within 0.8 m; cluster count priority, RMS tie-break). Logic is covered by tests/test_ppc_gici_override.py. Caveat: exact score replay of the 86.21% run still needs the regenerated candidate pool (the libgnss_diag_phase10/19 dirs + selector_ranker_predictions_v5_nlos.csv are regenerable via the plan.md CLIs, not all checked in), and some thresholds in the original uncommitted build were under-documented — the committed defaults follow plan.md (12 / 6 / 0.8 / 0.5 m).

Conclusion: the GICI candidate-injection avenue is closed (already realized). The eval_gici_tc_ppc2024_batch.py / materialize_gici_tc_ppc_candidate.py tooling (PR #72/#73) stands as candidate-quality verification, not a new lever. Remaining PPC headroom is in the n/r2 selector/ranker layer, not the GICI pool.

Previous Conclusion (2026-05-15)

Former canonical best: Phase 19aw K=3 = 83.42% OFFICIAL (rms_prefilter_k on top of 17-variant gici-open pool + status=5 gate + velocity bridge)、+11.94pp / +5,500m from Phase 11ep canonical 71.48%、TURING gap 2.18pp remaining。 詳細: rms_prefilter_breakthrough_2026_05_15.md。

1 セッションで 3 stage breakthrough:

  1. Fix=4 gate 撤廃 + status=5 path (Phase 19ap): selector が Float candidate (rms<=0.3m) を初めて受理 → +2.71pp (74.81%)
  2. Velocity/IMU bridge (Phase 19at): last good anchor + PF velocity × Δt synthetic candidate (rms=0.2, max_dt=6.0s) → +0.16pp (74.97%)
  3. rms_prefilter_k=3 (Phase 19aw): selector ranking 直前で top-3 by residual_rms に フィルタ → +8.44pp (83.42%)。 cluster-bias の温床だった composite formula residual / (... * abs_max^c) が高 abs_max 候補を促進していた problem を 19 lines のコードで切断。

過去のメモ (Phase 19al 76.83%) は metric ミス (pooled→per-run-averaged) の補正前数値。 Phase 19at = 74.97% が正しい補正後 baseline、 Phase 19aw = 83.42% が当時の頂点。

Previous (earlier 2026-05-15)

17-variant gici-open TC FGO pool で 76.83% aggregate (補正前) / 74.97% OFFICIAL (補正後)。 詳細: gici_open_phase19_breakthrough.md。

2 段 breakthrough (1 セッション):

  1. Format bug fix (Phase 19l): nmea_to_ppc_pool_esdfix.py に output_added=1 列 追加で gici TC FGO が PPC selector で初めて pick されるように → +1.62pp。 過去 "gici 0.00pp redundant" 結論を完全取り消し。
  2. 17-variant pool (Phase 19s-19al): lever arm / AR ratio / PR/Phase outlier / SNR / elevation / IMU / window の異なる 17 gici config を pool 同居、 PPC selector が per-epoch best を自動選択で 更に +1.45pp。

Pool 飽和確定 (5 連続 marginal/regression、 Phase 19am-an)。 残 gap 8.77pp の close には architectural breakthrough (CLAS/madoca PPP-AR / city-model NLOS / triple-freq cascade / multi-base) が必要、 ~月単位 task。

n/r2 構造的限界: 全 19 variants で 44.21% に張り付き、 全 knob で改善なし。 残 epoch (55.79%) は libgnss++ + gici 全 candidate でも recoverable でない。

(legacy) phase11er n/r2 rescue 状態 (2026-05-14)

(以下は gici breakthrough 前の状態記録。 phase11er 自体は今も 17-variant pool に 吸収されて生きている)

  • libgnss++ is not losing to RTKLIB demo5 on the checked-in public PPC coverage benchmark. It wins all six Tokyo/Nagoya public runs in the existing RTKLIB demo5 comparison.
  • The real gap is harder: libgnss++ still scores only 20-30% official PPC in hard Nagoya spans, and gnss_gpu currently relies on that solver output as a hybrid floor/candidate source.
  • phase11er is a safe RTKDiag candidate-emission policy for nagoya/run2. It improves the hard lowcase, but it does not prove that CT-RBPF/FGO itself has become strong.
  • The next useful work is fallback-span diagnosis and candidate coverage, not more global selector complexity.

libgnss++ vs RTKLIB demo5

Existing third_party/gnssplusplus docs report:

run gnss++ official RTKLIB demo5 official delta
tokyo/run1 34.9% 0.0% +34.9pp
tokyo/run2 69.0% 16.9% +52.1pp
tokyo/run3 60.6% 35.6% +25.0pp
nagoya/run1 49.5% 22.4% +27.1pp
nagoya/run2 20.9% 11.0% +9.9pp
nagoya/run3 27.4% 7.6% +19.7pp

Local scorer checks on existing .pos files:

profile / pos-dir weighted honest PPC note
libgnss_rtk_pos_v5 50.72% current gnss_gpu hybrid floor
libgnss_diag_phase10/dev_demo5_trusted_o3 54.10% nagoya/run2 24.49%
libgnss_diag_phase10/demo5_continuous_nojump 52.83% nagoya/run2 29.24%

Important distinction: dev_demo5_* and demo5_continuous_* are local libgnss++ profile outputs, not RTKLIB demo5 raw output.

phase11er

phase11er is a conservative nagoya/run2 RTKDiag rescue policy:

  • allowed only on nagoya/run2
  • residual selection over seven candidate labels
  • ratio_min=1.0
  • residual_rms_max=50.0
  • candidate-to-hybrid gate: 10 m
  • emit candidate, fallback to hybrid
  • all non-nagoya/run2 runs block candidates and pass through hybrid

Verified p2k result:

scope hybrid phase11er delta
6-run honest p2k 14.224688% 14.354783% +0.130095pp
nagoya/run2 full p2k 10.544309% 11.815481% +1.271172pp
nagoya/run2 segment p2k 45.322161% 50.785987% +5.463826pp

Internal diagnosis on nagoya/run2 p2k:

emitted source epochs pass <=0.5m fail >3m median error p95 error
rtkdiag_candidate 1298 1220 40 0.101m 0.744m
rtkdiag_fallback_hybrid_rtkdiag 654 1 647 26.354m 31.909m
pf_rtkdiag 42 0 42 27.011m 48.063m
pf_min_sats 6 0 6 23.560m 45.204m

Interpretation: candidate epochs are strong; fallback epochs are the bottleneck.

Current Artifacts

  • experiments/results/phase71_osmroad_production_summary.csv
  • experiments/results/phase71_osmroad_block_other_runs_summary.csv
  • experiments/results/phase70_osmroad_neutral_check_summary.csv
  • experiments/results/ppc_phase57_gap_nagoya_run2_internal_epochs.csv
  • experiments/results/ppc_ctrbpf_fgo_phase43_prod_*_full_runs.csv
  • experiments/results/ppc_phase71_osmroad_prod_*_full_runs.csv
  • experiments/results/selector_ranker_predictions_v5_nlos.csv
  • experiments/results/ppc_compare_libgnss_v5_runs.csv
  • experiments/results/ppc_compare_dev_demo5_trusted_o3_runs.csv
  • experiments/results/ppc_compare_demo5_continuous_nojump_runs.csv

Legacy phase11er artifacts:

  • experiments/results/ppc_phase11er_policy_all_p2k_runs.csv
  • experiments/results/ppc_phase11er_internal_n2_p2k_runs.csv
  • experiments/results/ppc_phase11er_internal_n2_p2k_internal_epochs.csv
  • experiments/results/ppc_phase11er_internal_n2_p2k_state_summary.csv
  • experiments/results/ppc_phase11er_internal_n2_p2k_state_groups.csv
  • experiments/results/ppc_phase11er_internal_n2_p2k_state_spans.csv

See ../experiments/results/README.md for artifact conventions.

Next Tasks

  1. Keep Phase71 production replay reproducible without committing /tmp materialized OSM candidates or large overlay CSVs.
  2. Phase72: test whether road/map constraints can generalize beyond the n/r2 OSM span without perturbing the five neutral runs.
  3. Find stronger absolute candidate sources for stable low-residual wrong solutions that ranker/consensus logic cannot detect.
  4. Reproduce RTKLIB demo5 raw baseline through the third_party/gnssplusplus/docs/ppc_reproduction.md coverage-matrix path and align the denominator with the local scorer.
  5. Compare libgnss++ profiles at span level, especially on nagoya/run2.

Reproduction Snippets

Current libgnss++ floor scorer:

python3 experiments/exp_ppc_libgnss_hybrid.py \
  --skip-solvers \
  --pos-dir experiments/results/libgnss_rtk_pos_v5 \
  --spp-dir experiments/results/libgnss_spp_pos \
  --results-prefix ppc_compare_libgnss_v5

Legacy phase11er skeleton:

base=experiments/results/libgnss_diag_phase10
labels=fgo_v14_snr38,full_ratio15_lock3_trustedseed_rtkout3oGem3,dev_demo5_trusted_o3,n2_nobds,fgo_v1,full_ratio15_lock3_trustedseed_rtkout3mlc1,full_ratio15_lock3_trustedseed_rtkout5
posdirs=${base}/fgo_v14_snr38,${base}/full_ratio15_lock3_trustedseed_rtkout3oGem3,${base}/dev_demo5_trusted_o3,${base}/n2_nobds,${base}/fgo_v1,${base}/full_ratio15_lock3_trustedseed_rtkout3mlc1,${base}/full_ratio15_lock3_trustedseed_rtkout5
python3 experiments/exp_ppc_ctrbpf_fgo.py \
  --results-prefix ppc_phase11er_policy_all_p2k \
  --runs all \
  --methods rbpf+dd+gate+hybrid,rbpf+dd+gate+hybrid+rtkdiag_pf \
  --n-particles 5000 \
  --start-epoch 0 \
  --max-epochs 2000 \
  --hybrid-pos-dir experiments/results/libgnss_rtk_pos_v5 \
  --rtkdiag-candidate-pos-dirs "$posdirs" \
  --rtkdiag-candidate-diag-dirs "$posdirs" \
  --rtkdiag-candidate-labels "$labels" \
  --rtkdiag-candidate-run-index-policy phase11er