AuraSync-Engine V4 — Detect → Correct → Recover (simulation prototype)

An event-camera state estimator and controller for an inverted pendulum (cart-pole) that is hit by sudden impulses. V4 adds a way to detect when the learned observer is confidently wrong, correct it from independent event geometry, and recover with an uncertainty-aware safety supervisor.

Everything here runs in simulation on a 2-core CPU. No hardware result is claimed. The V3 README is in docs/V3_README.md.

V4.1 update: geometry-first, evidence-aware, selective neural assistance

Result (held-out seeds 4000+, 272 paired stress trials):

• Adding the network did not help. V4.1 against the same pipeline without the network (D-ev): survival −1.8 points [−7, +3], safe recovery −0.4 [−6, +5].
• The evidence fix is the real gain. D-ev against Baseline D: survival +7.7 [+2, +13], safe recovery +9.2 [+4, +14], and occlusion survival 45% → 84%.
• Recommended default: D-ev, which needs no network and 0.73 cores (V4: 1.83).

Details: docs/V41_REPORT.md · audit: docs/V41_AUDIT.md · code: aurasync/v41/pipeline.py, benchmark_v41.py, analyze_v41.py, v41/ · tests: tests/test_v41.py.

The idea in one analogy

The V3 neural observer is a ship's navigator who dead-reckons. After a big wave he is sure of his position, and wrong. In V3, that was the cause of every fall.

V4 adds three things:

• a lookout who takes direct bearings on the mast: a RANSAC line fit on events less than 3 ms old;
• a rule for when to believe him: a CUSUM consistency detector plus a gated Kalman update, with re-initialisation only after several consistent bearings;
• a way to warn the captain (the supervisor) when nobody is sure.

Headline results (test seeds, never used for any choice)

What holds up:

• V4 vs V3 + shield: survival +41.7 points [+34, +49], safe recovery +30.4 [+23, +37].
• Faults V3 cannot survive: V4 survives the injected sign flip and the confident 15° error on 16/16 trials (V3 + shield: 44% and 38%). Its alarm fires a median 4 ms after the fault.
• HDR: 100% survival, where V3 and the frame camera get 0%.

What does not:

• Baseline D, the same detect-and-correct idea without the neural network, is level with or better than V4 in 14 of 15 non-nominal conditions. In nominal conditions its safe recovery is 7.1 points higher [2, 13], at a sixth of the compute.
• Occlusion is the only condition where V4 is clearly ahead (69% vs 44% survival).
• Correction is essential. Detection plus uncertainty inflation without it is worse than V3: 8% nominal survival.
• Uncertainty inflation showed no benefit.
• The first V4 design (the network overwrites the state) had the lowest error and 27% survival.
• Compute: default V4 needs about 1.5 CPU cores of compute per second of control.

Full report with CIs, ablations, failure analyses and limitations: docs/V4_REPORT.md.

Repository map

Reproduce

pip install torch numpy scipy matplotlib --break-system-packages
python3 -m unittest discover -s tests -v       # 12 tests; 4 need the V3 checkpoints + dev set
python3 v4/repro_v3.py                         # V3 reproduction check
python3 v4/collect_dev.py && python3 v4/eval_dev.py     # validation set (about 3.4 GB in data/)
python3 v4/select_v4.py && python3 v4/tune_v4.py        # selection and tuning (validation only)
WHICH=all python3 benchmark_v4.py              # test benchmark, resumable (about 2 h on 2 cores)
python3 analyze_v4.py                          # tables, CIs, figures
python3 v4/latency_v4.py                       # timings (run on an idle CPU)

Runs are deterministic (fixed seeds, 1 thread per worker), and the V3 checkpoints in checkpoints/v3/ are reused unchanged. Rules that hold throughout:

• Ground truth: estimators never read it. Replay tests pass s_true=None.
• Configuration: stored in results/v4/tuning_v4.json (supervisor), checkpoints/v3/calibration.json and results/v3/tuning.json (inherited).

Integrity notes

• Simulation only. No real event sensor, actuator or timing hardware.
• Latency. The 1 kHz control loop is not a demonstrated 1 ms sensor-to-actuator latency. See docs/V4_LATENCY.md.
• No priority claims. No prior-art search was done, and nothing here is claimed as first of its kind, formally safe or patentable.
• Sample size. Stress and scenario cells use 16 trials and mostly one seed, so differences below about 20 points there are not established.