Skip to main content

Results: iteration 3 (second zig, stopped at validation)

Question: does a problem with more information in the bearings — a second zig, both turns across the line of sight (generator 0.3.0, spec 007) — bring the 2c models within the G2 speed threshold? This is the first Part B lever of SRD Annex A (B.2.1), chosen after the owner's realism check in the viewer.

The models are the 2c configuration, trained from scratch on seed 204. A pre-registered validation rule (spec 007 FR-006) decides whether the pair goes to seed 204's test split: the LSTM's G2 speed interval must reach below 1 kt.

Stopped at validation: speed did not reach the threshold

The LSTM is at 1.10 kt [1.06, 1.16] at zig + 600 s on the σ = 1° slice, against 1 kt, so the rule stops the iteration and the test split is not scored. The forecast recorded before training was 1.1–1.3 kt.

Classical gained more. It reaches 0.34 kt on the same rows; the LSTM is 3.3 × classical on speed, against 2.4 × in iteration 2c.

The BiLSTM is the closest any model has come: 1.02 kt [0.98, 1.05], reported only. Range and course pass with the largest margins of the project.

Full record: validation.md, the decision record and the write-up.

Seed-204 validation split (σ = 1°)​

EstimatorRange, zig + 300 sCourse, zig + 300 sSpeed, zig + 600 s
G2 threshold< 20%< 10°< 1 kt
lstm-instantaneous (iteration 3)8.8% [8.5, 9.1]6.19° [5.89, 6.41]1.10 kt [1.06, 1.16]
bilstm-time-late (iteration 3)7.2% [7.0, 7.5]4.36° [4.18, 4.54]1.02 kt [0.98, 1.05]
lstm-instantaneous (2c, reported)11.4%6.53°2.01 kt
bilstm-time-late (2c, reported)17.7%7.72°1.93 kt
classical3.4%1.43°0.34 kt
naive40.2%89.7°4.23 kt

n = 4,883 scenarios at zig + 300 s (99% of the slice) and 3,165 at zig + 600 s (64%). The 2c models were trained on the two-leg problem and are out of distribution here.

Zig + 300 s falls mostly during the second turn, so it is not like-for-like with iterations 1–2c; by zig + 600 s nearly every scenario is on leg 3 (spec 007, "Reading the gate under three legs").