v1.111 — estimator-robustness campaign (IEKF under noise + GNSS dropout)

[avrabe]

The first genuinely stochastic dispersion campaign

The attitude and fail-safe campaigns are deterministic sweeps. This one is a true Monte-Carlo: the noise is the dispersion. It drives the real relay_iekf::Iekf against a static-hover truth across 600 seeded-random trials, each a dispersed noise + GNSS-dropout realisation (Box-Muller Gaussian).

Dispersion vector (per trial)

• gyro noise σ ∈ [0.001, 0.02] rad/s, accel σ ∈ [0.02, 0.30] m/s², GNSS σ ∈ [0.10, 1.00] m
• one GNSS-dropout window (start 5–8 s, duration 1–3 s): the filter must dead-reckon on the IMU, then reconverge when GNSS returns

Invariants (all 600 trials, 0 failures)

metric	worst measured	bound
tilt error (gravity aiding)	0.016 rad (~0.9°)	< 0.05
dead-reckoning drift during dropout	9.5 m	< 40 (divergence catch)
reconverged position error	1.71 m	< 2.5
peak NEES (nees_position chi-square consistency)	56.9	< 90 (overconfidence catch)
NaN	none	—

Using the IEKF's own NEES statistic is the aerospace estimator-consistency check — it catches an overconfident covariance (error large relative to reported uncertainty), which a pure error-magnitude bound would miss.

Discipline

• Reproducible from (seed, index); the worst NEES trial (56.9) is identical at 600 and 2000 trials — confirms the seeding.
• Runs under the Closed-loop simulation named gate + workspace test; rivet steps marked # bench-only (heavy IEKF sim, ~49 s).
• FV-FALCON-SIMMC-003 (method: simulation, verifies IEKF-P01 / EKF-P05).
• Also bumps FV-FALCON-SIMMC-002 implemented→verified — a v1.110 release-tracking correction (its CI is green), per the rivet 0.23 discipline now in use for planning/tracking.

Honest scope

Static-hover truth (cleanly isolates noise-rejection + dropout reconvergence). Maneuvering-truth + the spoof-monitor path are the next estimator-campaign slices.

🤖 Generated with Claude Code

[temper-pulseengine]

Automated review for PR #250

pulseengine/relay:feat/estimator-robustness-campaign → pulseengine/relay:main

Verdict: 💬 Comment

Summary: The changes in this pull request are focused on adding a new test for an estimator robustness campaign, which is implemented and verified. The test uses a Monte Carlo simulation to validate the safety of the NET across a dispersed regime sweep. This ensures that the estimator performs well under various conditions and meets the required safety standards.

Findings: 0 mechanical (rivet) · 2 from local AI model.

Findings (2):

1. artifacts/verification/FV-FALCON-SIMMC-002.yaml:13

       status: verified
   

   The test is verified, indicating that it has been successfully executed and validated.

2. artifacts/verification/FV-FALCON-SIMMC-002.yaml:25

       description: >
   

   The description provides a clear explanation of the purpose and functionality of the test.

---

Generated by a local AI model and post-validated against a strict JSON contract. Each finding includes the verbatim line being criticised — verify by reading the file at the cited location.

Reviewed at d3ddaf9