Syntonic Nav: Against Truth-Tuned Rivals
Year: 2026
Type: preprint
Abstract: A Filter with No Tuning against EKF2 and EKF3 Tuned on the Truth, Pre-registered Simulation Syntonic Nav takes every noise, prior and gate of a navigation filter from records of the sensors; nothing in it is tuned. In its version 2.0, its latest form, SYN5 v4b, was compared with PX4's EKF2 and ArduPilot's EKF3 at their shipped default values, and held all 30 sealed checks. That comparison leaves open whether the advantage is only the defaults' mismatch with the vehicle. This note answers it with the strongest rivals the bench allows: EKF2 and EKF3 whose noise, bias and gate parameters were tuned against the true state, an oracle no human tuner has, on the same kinds of scenario they were then tested on. Eleven parameters of EKF3 and eleven of EKF2 were searched in log space on spent seeds; the tuned sets, the criteria and written predictions were frozen; one confirmation ran on 20 fresh seeds over 14 scenarios and was judged once. SYN5 v4b, unchanged, held 39 of the 42 criteria. It diverged in none of 280 runs, while the truth-tuned EKF3 diverged on 4 of 20 seeds with a faulty IMU and the truth-tuned EKF2 on all 20. Its velocity error was 0.50 to 0.67 of the better oracle's in every scenario, and its vertical error the lowest of all filters everywhere. The three failures, each exactly at the 70 % seed boundary, were tilt during a 55 s GNSS denial (ratio 1.19), horizontal position with correlated receivers (1.06) and drift during a denial with a large bias walk (1.10). The tuned EKF3 is 11 to 28 % better on tilt in most scenarios, but not on enough seeds to count as better. Tuning halved EKF3's tilt and velocity errors: much of the gap to the defaults was mismatch. The predictions written before the run gave the whole claim about 0.02. A second part treats the tilt gap. A diagnosis traced it to the gyro bias at half throttle, which SYN5 starts at the idle value and learns slowly. A one-change fix, SYN5 v5a, widens that bias's prior by the measured shift between levels, with no new constant. Verified on 20 more fresh seeds, it held 39 of 42 criteria (v4b 38 on the same seeds): it brings tilt to the oracle's level in the seven scenarios without an IMU fault or walk, is worse than v4b nowhere, and leaves the gap in the two bias-walk scenarios, which have another cause. This is a verification on the scenario types the fix was designed on, not a held-out result. Simulation only; the code is withheld until the hardware tests are complete. Authorship and collaboration statement. All prose in this corpus was drafted by large language models under the author's direction and revised in author-directed loops. The author set the questions, chose the hypotheses, designed and ran every experiment, decided every retraction, and designed the corpus as a system (immutable hub-and-spoke record, concept and version identifiers, errata register, hash-frozen firmware, pre-registration gates). Verification is independent of both: the algebraic core is machine-checked in Lean 4/Mathlib, numerical results follow pre-registered protocols, and flight logs are public. Where model-generated reasoning text survives in a published document, it is an artifact of this method and is recorded in the Corpus Status Register. The author alone approves every statement and is accountable for it; no language model is an author. This note is simulation only: no Lean proof and no flight log belong to it; its code and data are committed by SHA-256 and withheld until the hardware tests. RETURN TO HUB : https://doi.org/10.5281/zenodo.17254395
Author: Bronsard, Jean-Pierre
Language: English
Subfield: Aerospace Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225569704