PPC Deployable Smoother Sanity Report

This note records the 2026-05-05 PPC smoother-stack sanity checks. The performance claim in this PR is reference-free at runtime: PPC reference.csv is used for evaluation only, not for choosing a smoother candidate or segment.

Reference-Free Claim

The deployable result to carry forward is the existing PPC coverage profile, not the smoother oracle sweep. On the six public Tokyo/Nagoya PPC runs, that profile uses fixed solver settings and no reference-driven segment selection at runtime:

comparison vs RTKLIB demo5	result
Positioning-rate lead	+17.0 pp
PPC official-score lead	+28.1 pp
P95 horizontal-error delta	-11.96 m

The same stack also moves the truth-validation wrong-FIX ratio in the right direction:

metric	pre-stack default	post-stack default
fix_wrong / total fixes	28.9%	19.4%
fix95%	0.81 m	0.19 m

That is the reference-free runtime performance story: keep the solver profile ahead of demo5 while reducing mis-fix risk.

Deployable Smoother Check

The selector tooling supports --selection-mode priority_first, which fixes selection to candidate rules plus priority order. In this mode, the PPC reference is still used to evaluate the final output, but it is not used to choose the candidate for a segment, and the reference-scored non-regression gate is skipped.

A fully internal smoother probe used:

• anchor: status == FIXED, ratio >= 10, post-suppression RMS <= 0.05 m, and at least 8 satellites
• loss: non-FIX or FIXED with ratio < 6

That deployable-only fixed smoother check produced no public-score gain over its fixed baseline:

check	aggregate
deployable internal smoother fixed recipe	+0.000 m

Because it does not add reference-free gain, it should stay diagnostic for now. Do not present it as a performance improvement.

Mis-Fix Policy

For PPC-facing changes, prefer gates that preserve or reduce wrong FIX instead of chasing extra fixed epochs. In practice:

• Do not select smoother segments using reference-scored deltas.
• Do not enable a smoother/bridge path as a benchmark claim unless it improves deployable score without increasing wrong-FIX indicators.
• Keep priority_first as the deployable selector mode; keep oracle_delta only for research diagnostics.

Diagnostic Images

The trajectory plots are useful for reviewing where bridges, gaps, and mis-fixes appear, but they are not performance evidence by themselves. Red WRONG_FIX points are FIXED solution epochs whose post-run 3D reference error exceeds 0.50 m.

Residual red points are broken down in PPC residual wrong-FIX analysis. In short, the remaining wrong FIX is baseline-selected and does not match a smoother selector rule, so the next deployable lever is a real-time fixed-update quality gate rather than more reference-selected smoothing.

Existing full real-time gate outputs are summarized in PPC existing real-time gate outputs. The current headline profile is the later deployable sigma-demote nis2-ratio4 runtime selector in that note. The images below are historical smoother/filter diagnostics, not the current PR performance claim.

Reference-Selected Sweeps

Full-reference smoother sweeps are kept only as diagnostics. Those runs use PPC reference.csv to choose the best candidate per official scoring segment, so they are not private-score-equivalent and are not part of the reference-free performance claim.