FGO FIX-rate observation-quality plan

Objective

Close the remaining approximately 0.7955 percentage-point gap to the current +2 pp correct-FIX-distance goal without increasing wrong-FIX distance on a held-out Tokyo run. Raw FIX labels are not the optimization target: a change that promotes more wrong integer solutions fails even when the reported FIX rate rises.

The next experiment targets the float observation model before LAMBDA. It does not further tune TDCP thresholds and does not promote the existing multi-epoch or conditional multi-band shadow candidates.

Evidence already consumed

Do not repeat the following rejected directions:

• temporal integer agreement, BSR, and ratio alone;
• held-out carrier rows from the same urban measurement basin;
• constrained graph cost using the same active factors;
• conditional multi-band AR;
• direct temporal-carrier promotion or a TDCP-based FIX gate; and
• uncapped satellite-badness feedback, which starved ambiguity resolution.

The current three-run logs also separate the remaining opportunity:

• run1 correct FLOAT epochs are split mainly between no/insufficient candidates and ratio rejection;
• run2 and run3 correct FLOAT epochs are dominated by ratio rejection; and
• lowering the ratio threshold is unsafe because the rejected population also contains many wrong candidates.

The working hypothesis is therefore that biased DD pseudorange factors are degrading the float state and ambiguity covariance before integer search. The candidate generator should be improved before any acceptance rule is relaxed.

Research basis

• Song et al. (2025) compare pseudorange change with Doppler-predicted range change, then compare DD pseudorange change with IMU/odometer-predicted DD geometry. Measurements that fail are downweighted before FGO: https://arxiv.org/html/2510.00524v1.
• Wen and Hsu's GraphGNSSLib supplies an open FGO/RTK reference using DD pseudorange, carrier phase, Doppler, and LAMBDA, and points to graduated non-convexity work for urban outlier mitigation: https://github.com/weisongwen/GraphGNSSLib.
• Gao et al. connect pseudorange, carrier, and Doppler in an urban RTK factor graph and perform ambiguity resolution after the float optimization: https://doi.org/10.33012/2022.18190.
• Teunissen-style integer aperture remains the safety model: observation cleanup may improve the float distribution, but it does not authorize relaxing the existing ambiguity acceptance tests.

Proposed experiment

Add an opt-in, monitor-only predicted DDPR quality trace. For every DD pseudorange row with an uninterrupted target/reference pairing, calculate two causal innovations:

1. DD Doppler consistency: compare the epoch-to-epoch measured DD pseudorange change with integrated DD Doppler range change.
2. IMU geometry consistency: compare the same measured change with the DD geometric-range change from the one-step IMU-predicted antenna pose.

Both tests must use satellite double differences. This cancels common rover and base receiver-clock terms and avoids importing the raw clock-jump failure seen by the temporal-carrier investigation. The current epoch's optimized pose must not be used to classify its own input factor.

The monitor exports, per DD row:

• epoch, target, reference, signal, and arc/pair age;
• measured DDPR change;
• Doppler- and IMU-predicted change;
• signed innovation and propagated sigma for each witness;
• the previous/current DDPR residual against the causal predicted geometry;
• elevation and rover/reference SNR; and
• proposed action (keep, downweight, or unavailable).

No graph factor, covariance, ambiguity set, hold state, reported position, or FIX/FLOAT status reads the monitor result.

Only after the monitor passes its gates may a separate experimental switch apply a bounded sigma inflation to DD pseudorange factors. Hard exclusion is out of scope for the first activation experiment. Carrier factors and LAMBDA thresholds remain unchanged.

Dataset and holdout protocol

1. Development: Tokyo run1 only. Select normalization and bounded downweighting parameters.
2. Validation: freeze all rules, then run Tokyo run2 once. A failure ends the experiment; do not tune against run2.
3. Final holdout: run Tokyo run3 only after run2 passes. Do not unlock or adjust any threshold after seeing run3.
4. Preserve the existing baseline CSVs and problem caches. Every experimental output uses a new prefix and records the exact command and cache fingerprint.

Reference truth is used only for offline scoring. Runtime classification may use GNSS observations, Doppler, IMU prediction, and causal state, but never reference.csv.

Gates

Gate 0: monitor non-interference

• status, ECEF position, ratio, fixed ambiguity count, and AR outcome are identical to baseline at every epoch;
• common rover/base clock-jump unit tests do not create a DD innovation; and
• missing Doppler, pair changes, gaps, or non-finite covariance fail to unavailable, never bad.

Gate 1: witness usefulness

Using truth-position DDPR residuals only as offline labels, the frozen monitor must, on run2, capture at least 50% of gross DDPR rows while flagging no more than 5% of clean rows. Report the complete precision/recall curve; the stated operating point is the minimum continuation gate, not a post-hoc optimum.

Gate 2: short solver A/B

On a representative 500-epoch run1 slice:

• zero additional wrong FIX epochs;
• no loss of correct FIX epochs;
• no additional smoother failure or non-finite epoch; and
• wall-clock overhead at or below 10%.

Gate 3: full-run activation

With parameters frozen after full run1:

• wrong-FIX distance must not increase on run2 or run3 individually;
• correct-FIX distance must not decrease on either holdout;
• aggregate correct-FIX distance must close the remaining 0.7955 pp goal;
• official score and all-solution <50 cm distance must not regress by more than 0.1 pp on any run;
• fixed horizontal RMS and P95 must not regress by more than 5%; and
• matched positioning distance and solver-failure counts must not regress.

If the improvement is positive but below the target, keep the switch experimental and document the measured contribution. If any safety condition fails, retain telemetry and reject activation.

Implementation sequence

1. Checkpoint the completed temporal-carrier work, then create a dedicated observation-quality branch so the two experiments remain reviewable.
2. Add solver-independent DDPR innovation records and deterministic unit tests for geometry, clock cancellation, gaps, pair changes, and missing witnesses.
3. Wire a default-off CLI monitor and normalized per-factor CSV.
4. Run baseline/monitor A/B and Gate 1 without changing the graph.
5. If Gate 1 passes, add bounded DDPR sigma inflation behind a second default-off switch and run the staged A/B protocol.
6. Promote nothing unless every full-run gate passes; otherwise record the negative result and move to a different observation model.

Development checkpoint

The default-off monitor passed Gate 0 on both the first 500 epochs and the full 11,905-epoch run1: status, ECEF position, ratio, fixed ambiguity count, and AR outcome were byte-for-byte identical to the preserved baseline CSV. The focused clock-cancellation, fail-closed, and fixed-lag non-interference tests also pass.

The original rule requiring both normalized temporal innovations to exceed a common threshold did not meet the full-run development target. At 0.2 sigma it captured 50.973% of gross truth-labelled rows but flagged 12.531% of clean rows; at 0.3 sigma the rates were 40.675% and 5.634%. This rule is rejected.

The causal absolute residual against the one-step IMU-predicted geometry is a substantially stronger development score. Freezing the rule abs(current predicted DDPR residual) > 1.5 m on run1 captures 85.522% of 14,498 gross rows while flagging 2.848% of 153,347 clean rows. The 30,023 rows between 1 m and 4 m are an explicit ambiguity band and are excluded from Gate 1 scoring. This 1.5 m threshold is now frozen for the run2 validation; it must not be adjusted using run2 or run3.

With that rule unchanged, run2 passes Gate 1: it captures 5,705 of 5,894 gross rows (96.793%) and flags 477 of 166,600 clean rows (0.286%). The monitor remains non-interfering across all 9,147 run2 epochs, with exact agreement in status, ECEF position, ratio, fixed ambiguity count, and AR outcome. This unlocks the bounded-downweight experiment on run1. Run3 remains sealed until the activation rule, inflation scale, and all other parameters have been frozen.

The bounded activation sweep on the first 500 run1 epochs rejected 3x and 5x because each introduced one wrong FIX. The 2x variant introduced no wrong FIX but exchanged one previously correct FIX epoch for another and therefore failed the strict per-epoch no-loss gate. The 1.5x and 1.75x variants changed no FIX/FLOAT labels. The 1.25x variant was the only improving safe candidate: one additional correct FIX, no lost correct FIX, no wrong FIX, and no NONE or non-finite epoch. Consequently 1.25x is the sole candidate advanced to the full-run1 activation A/B; the stronger multipliers are permanently rejected.

The full-run1 A/B rejects that remaining 1.25x candidate. It increased correct-FIX distance by 148.467 m (+1.439 percentage points) and improved the official 50 cm distance by 144.574 m (+1.401 pp), but it also increased wrong-FIX distance by 189.135 m (+1.833 pp) and lost 619 previously correct FIX epochs. This violates the zero-wrong-FIX-increase stop condition even though the net correct-FIX count increased. Therefore the activation switch was removed, run2/run3 activation was not run, and only the non-interfering telemetry is retained. The run3 observation-quality labels remain sealed; no activation result was inspected there.

As the next independent observation-model check, the existing receiver-clock- drift-free single-difference Doppler velocity factors were evaluated on the same run1 500-epoch development slice at 0.5 m/s (the less aggressive setting favoured by the earlier RTK study). They added 18,356 Doppler rows but changed no FIX/FLOAT decision: both baseline and candidate had 499 correct FIX and zero wrong FIX. Meanwhile fixed horizontal RMS regressed from 0.0422 m to 0.0539 m (about 28%). With no correct-FIX gain and a clear accuracy regression, this model was not advanced to a full run. The temporary harness sigma override was removed as well.

Stop conditions

Stop before activation if the score is mainly explained by the optimized current pose, if DD formation does not cancel injected receiver-clock jumps, if run2 misses Gate 1, or if the first bounded-downweight run adds any wrong FIX. These conditions prevent another expensive threshold search over a non-independent witness.