RTK Tuning Gates

The default RTK pipeline is the sign-off path for the PPC production runs. Additional gates are default-off unless noted here and are intended for diagnostics, profile sweeps, or dataset-specific opt-in profiles such as Odaiba.

Default Guards

The low-speed non-FIX drift guard is part of the default kinematic output path, including the coverage profile. It rejects long FLOAT/SPP fallback drifts only when the surrounding FIX anchors indicate near-stationary motion. Use --no-nonfix-drift-guard to reproduce the raw unguarded fallback stream.

The SPP height-step guard is also default-on in the kinematic output path. It rejects SPP-only vertical spikes above --spp-height-step-min / --spp-height-step-rate while preserving FLOAT and FIXED epochs.

The FLOAT bridge-tail guard is default-on after six-run PPC sign-off. It rejects FLOAT epochs in slow bounded FIX-to-FIX segments when they diverge from the anchor bridge, and uses horizontal FIX-anchor speed for its motion gate. Use --no-float-bridge-tail-guard to reproduce the pre-bridge-tail coverage stream.

The offline status replay also supports a truth-free kinematic integrity layer. The six-fold LOO profile uses a 12 m jump, 200 m/s2 acceleration, and a three-epoch hold. The advanced replay can extend a post-trigger plateau to an eight-epoch maximum while jumps remain <=0.1 m. Its secondary trigger combines jump/acceleration with prefit RMS, AR ratio, suppressed-outlier count, and satellite count; LOO support for its exact thresholds is limited, so it remains exploratory. Low-satellite exoneration requires nsat >=11, prefit RMS <=0.5 m, and NIS/observation <=0.2; kinematic quarantine always overrides that exoneration. These are apply_ppc_status_demotion.py replay options, not native RTK defaults.

The online consensus manager does not let a soft primary-only telemetry warning create a quarantine while its independent estimator is unavailable, because such a state has no evidence source with which to recover. A hard primary suspect remains fail-closed. --allow-soft-suspect-without-shadow restores the legacy diagnostic behavior. For tight-aperture experiments, --require-primary-suspect-for-disagreement requires joint primary telemetry and estimator disagreement to advance the disagreement streak; it is opt-in.

Opt-In Gates

Flag	Purpose	Default
--ar-policy {extended\|demo5-continuous}	AR extras gate. demo5-continuous disables relaxed-hold-ratio / subset-fallback / hold-fix / Q-regularization for demo5-style continuous ambiguity tracking.	extended
--max-subset-ar-drop-steps <N>	Extend the progressive subset-AR search by dropping up to N worst-variance DD pairs. Diagnostic only: Nagoya run2 spot checks showed extra candidates are usually rejected by the jump gate unless paired with riskier validation changes.	6
--max-hold-div <m>	Reject fix if the hold-state diverges from float by more than N meters.	0 disabled
--hold-fix-unchanged-set	Refuse the hold-fix fallback when a current DD pair has no held integer (a satellite joined or an ambiguity was reset), instead of rounding that pair. Fewer wrong FIX, fewer FIX: on PPC + UrbanNav Tokyo, gnss solve --preset low-cost went from 5.77 % to 3.10 % of FIXED epochs more than 0.5 m (3D) off the reference, with 7.5 % fewer FIXED epochs. Also in gnss live and gnss replay (details).	false
--hold-fix-min-matched-pairs <n>	Refuse the hold-fix fallback when fewer than n current DD pairs carry a held integer. The default 4 is the floor it always had. With 12, gnss solve --preset low-cost on PPC + UrbanNav Tokyo went from 5.77 % to 2.85 % of FIXED epochs more than 0.5 m (3D) off the reference, keeping 101 % of the correct FIXED epochs. On held-out SmartPNT-POS and TEX-CUP urban runs it went from 1.69 % to 1.07 %, keeping 98.5 %. Also in gnss live and gnss replay (details).	4
--min-fix-satellites <k>	Refuse an ambiguity-resolution FIX (full set, subset or WL/NL fallback) whose fixed set has fewer than k non-reference satellites. The epoch stays FLOAT unless the hold-fix fallback applies. With 4 added to --preset low-cost --hold-fix-min-matched-pairs 12, held-out GREAT-WHU, SmartPNT-POS and TEX-CUP urban runs went from 0.95 % to 0.37 % of FIXED epochs more than 0.5 m (3D) off the reference, keeping 97.5 % of the correct FIXED epochs. gnss solve on PPC + UrbanNav Tokyo went from 2.85 % to 1.22 %, keeping 97.6 %. Also in gnss live and gnss replay (details).	0 disabled
--max-pos-jump <m>	Reject fix if the epoch-to-epoch position jump exceeds N meters. Truth-validation against PPC reference showed jumps cluster at <5 m for correct fixes or >10 m for wrong FIX. Pass 0 to disable.	5 m
--max-fixed-anchor-age <s>	Stop using the last FIX position as a jump reference after N seconds. This prevents permanent stale-anchor lockout, but Nagoya 3 causal replays showed that time expiry alone either reacquires a wrong basin or loses too much coverage; use only for diagnostics.	0 disabled
--max-fixed-doppler-consensus <m>	When the FIX anchor is expired or the residual/covariance gate is suspect, reject FIX candidates farther than N meters from an independently propagated rover-Doppler position track. Accepted FIX only may re-anchor the track. Nagoya 3 at 10 m reduced wrong FIX from 664 to 9, but long-gap Doppler drift reduced correct FIX from 1,241 to 680, so this is an integrity experiment rather than a sign-off default.	0 disabled
--max-pos-jump-min <m> + --max-pos-jump-rate <m/s>	Reject fix if the jump from the last fixed position exceeds max(min, rate * dt), so vehicle gaps can be tested without a stale absolute distance clamp.	0 / 0 disabled
--max-float-spp-div <m>	Reject FLOAT epochs that diverge from the same-epoch SPP solution by more than N meters, then fall back to SPP/no-solution. Diagnostic gate for PPC FLOAT high-error sweeps.	0 disabled
--max-float-prefit-rms <m> + --max-float-prefit-max <m> + --max-float-prefit-reset-streak <N>	Reset ambiguity states for the next epoch after N consecutive otherwise accepted FLOAT epochs have high DD prefit residual RMS or max residual. The current FLOAT epoch is still reported.	0 / 0 disabled / 3
--min-float-prefit-trusted-jump <m>	Continuity selector for the residual gate. When > 0, high-residual FLOAT resets are allowed only if the FLOAT position has also diverged by at least N meters from the last trusted FIX/FLOAT position.	0 disabled
--max-fixed-prefit-rms <m> + --min-fixed-prefit-outliers <N> + --fixed-prefit-reset-streak <N>	Detect an accepted FIX in a high-prefit-residual basin after suppression, emit the trigger epoch as FLOAT, clear held integers and stale anchors, and restart ambiguity acquisition from the same-epoch SPP position. Both residual thresholds are required. Nagoya 2/3 diagnostics reduced large wrong FIX but also reduced official-score coverage, so this remains experimental and is not used by the PPC sign-off profile.	0 / 0 disabled / 1
--max-fixed-overconfidence-cov-trace <m2>	Add an upper FLOAT position-covariance-trace condition to the fixed-prefit reset, targeting filters that are implausibly confident while raw DD residuals remain large. It has no effect unless the fixed-prefit residual gate is enabled.	0 disabled
--max-update-nis-per-obs <v>	Reject a whole RTK DD Kalman update before state/covariance mutation when normalized innovation squared divided by active observations exceeds N.	0 disabled
--max-fixed-update-nis-per-obs <v>	Reject only FIXED ambiguity candidates when the preceding RTK DD update NIS divided by active observations exceeds N, while retaining the FLOAT solution for that epoch.	0 disabled
--max-fixed-update-post-rms <m>	Reject only FIXED ambiguity candidates when the preceding RTK DD update post-suppression residual RMS exceeds N meters, while retaining the FLOAT solution for that epoch.	0 disabled
--max-fixed-update-gate-ratio <v>	Apply the fixed-only NIS/post-RMS gates only when the AR ratio is finite and at or below N.	0 unconditional when gates are enabled
--min-fixed-update-gate-baseline <m> + --max-fixed-update-gate-baseline <m>	Optional baseline-length window for the fixed-only NIS/post-RMS gates.	0 / 0 disabled
--min-fixed-update-gate-speed <m/s> + --max-fixed-update-gate-speed <m/s>	Optional speed window for the fixed-only NIS/post-RMS gates.	0 / 0 disabled
--max-fixed-update-secondary-gate-ratio <v> + baseline/speed secondary window flags	Optional second OR-window for fixed-only NIS/post-RMS gates.	0 for all fields
--demote-fixed-status-max-ratio <v>	Keep the position solution but output FIX as FLOAT when the AR ratio is finite and at or below N. PPC nis2-ratio4 uses this to reduce measured Wrong/FIX while preserving official-score coverage.	0 disabled
--demote-fixed-status-min-satellites <N>	Keep the position solution but output FIX as FLOAT when fewer than N satellites are used. The audited KF/FGO integrity matrix uses 8 together with the paired low-satellite ratio gate; it changes no trajectory positions and keeps Tokyo 1 above the public FIX target.	0 disabled
--demote-fixed-status-low-satellite-ceiling <N> + --demote-fixed-status-low-satellite-max-ratio <R>	Keep the position solution but output FIX as FLOAT when at most N satellites are used and the AR ratio is at most R. Both options are required. The audited integrity profile combines N=11/R=15 with minimum 8 satellites.	0 disabled
--rtk-snr-weighting + --rtk-snr-reference-dbhz <v> + --rtk-snr-max-variance-scale <v> + --rtk-snr-min-baseline <m>	Low-cost observation model diagnostic. Uses the lower rover/base SNR for each SD link and inflates DD phase/code variance below the reference SNR, capped by the max scale.	false / 45 / 25 / 0
--rtk-snr-elevation-aware + --rtk-snr-code-only	Refinements of --rtk-snr-weighting (need it enabled). Elevation-aware measures the C/N0 deficit against an expected open-sky C/N0 of 33 + 15 sin(el) dB-Hz (capped by the reference) instead of the constant reference, so low-elevation satellites are not penalised twice and zenith-side satellites blocked by an overpass (48 to 33 dB-Hz) are. Code-only applies the inflation to code rows only, so the phase weight and the fix rate are kept. On UrbanNav Odaiba (stopped under an overpass, 10-29 m vertical float error) the default profile Vp95 drops 11.5 to 5.8 m; PPC coverage-profile weighted official is +1.06 pp (elevation-aware) / +0.23 pp (plus code-only) but not uniform per run (nagoya_run3 -7.0 / -2.6 pp), so both stay opt-in.	false / false
--rtk-canyon-carrier-prior	Kinematic RTK. The per-epoch SPP re-seed discards the position/ambiguity cross-covariance, so under a blocked sky the biased code defines the three float combinations that a stationary geometry cannot observe. While a zenith-side (>= 45 deg) satellite's code disagrees by more than the 3 m outlier gate with the position advanced from the previous posterior by a validated time-differenced carrier displacement (code +4..+14 m, carrier intact), that propagation replaces the re-seed (cross-covariances kept; process noise 1 cm/s) and re-initialised ambiguities are seeded from the carrier at the propagated position instead of from code. It starts only from a posterior that was code/carrier consistent for 5 s, releases after 5 s of consistency, lasts at most 180 s per episode and suspends the preset float-prefit reacquisition reset while engaged. Development set UrbanNav Odaiba only when this row was written (PPC and Hong Kong not yet evaluated): stop under the overpass 273750-273860 Vp95 27.2 to 2.0 m (default profile) and 24.4 to 3.9 m (--preset odaiba); epochs outside both stops unchanged; the second stop (274230-274300) is not engaged. Stays opt-in.	false
--cycle-slip-threshold <m> + --doppler-slip-threshold <m> + --code-slip-threshold <m> + --strict-dynamic-slip-thresholds + --adaptive-dynamic-slip-thresholds	Cycle-slip sensitivity sweep. Dynamic RTK keeps protective minimum thresholds by default; strict/adaptive modes are opt-in urban reacquisition experiments.	0.05 / 0.20 / 5.0 / false / false
--adaptive-dynamic-slip-nonfix-count <N>	Non-FIX epochs before adaptive dynamic slip thresholds activate.	3
--adaptive-dynamic-slip-hold-epochs <N>	Epochs to keep adaptive dynamic slip thresholds after activation.	10
--nonfix-drift-max-residual <m> + --nonfix-drift-min-horizontal-residual <m>	Tighten the default low-speed non-FIX drift guard for tail diagnostics while avoiding vertical-only fallback pruning.	30 / 0
--fixed-bridge-burst-guard + --fixed-bridge-burst-max-residual <m>	Reject isolated short FIX bursts when they diverge from the straight bridge between surrounding FIX anchors.	false / 20
--max-consec-float-reset <N>	Auto-reset ambiguities after N consecutive float epochs. 10 is a historical PPC official-score probe; the current PPC sign-off path is --realtime-profile sigma-demote.	0 disabled
--max-consec-nonfix-reset <N>	Auto-reset ambiguities after N consecutive FLOAT/SPP/no-solution epochs.	0 disabled
--max-postfix-rms <m>	Reject fix if the L1 post-fix DD phase residual RMS exceeds N meters.	0 disabled
--enable-wide-lane-ar + --wide-lane-threshold <cycle>	Pre-compute MW wide-lane integers and inject them as Kalman constraints into the LAMBDA search.	false / 0.25
--wide-lane-min-arc-samples <N>	Average the MW wide-lane over each satellite's continuous arc and constrain only when both arcs hold at least N samples. 0 is the legacy single-epoch test, which admits near-random integers under urban code multipath. --preset odaiba uses 100.	0
--enable-wlnl-fallback	Allow the experimental MW wide-lane / narrow-lane fallback after ordinary LAMBDA failure in non-IFLC runs. It reuses --wide-lane-threshold for both integer checks.	false
--enable-bsr-decimation + --bsr-worst-axes <N> + --bsr-max-drops <N>	Add BSR-guided subset AR decimation candidates alongside the variance-drop subset family. PPC high-wrong probes preserved FIX count but did not reduce Wrong/FIX, so this remains diagnostic.	false / 3 / 6

On PPC Tokyo and Nagoya, leave these off unless reproducing a specific sweep. On Odaiba or other urban multipath datasets, --preset odaiba (low-cost + arc-smoothed wide-lane AR) gives the most fixes and the lowest Hmed. Its FLOAT/SPP epochs in two stationary deep-canyon stretches carry up to 28 m vertical error, so its Hp95/Vp95 are above the default profile's on the 2026-10-11 Odaiba run, both over all epochs (5.28/13.89 m against 4.77/11.54 m) and over the epochs both profiles publish (5.11/11.81 m against 4.57/10.52 m) (benchmarks).

For those stationary deep-canyon errors, see Urban canyon RTK options: it compares --rtk-canyon-carrier-prior and the SNR options on Odaiba, Hong Kong and PPC, and warns against combining the carrier prior with the SNR options under --preset odaiba (wrong fixes 12 → 1,009 on Odaiba).