Results: iteration 5 (hybrid estimator, G3 passed)

Question: can a network help reach G3, "within 1.2× classical on every metric, and lower gross-error rate" (SRD §5)? The models alone are 2.1–3.5 × classical. Spec 009 judged a hybrid (Annex A B.2.6, decision 24).

• The estimator. Classical's least-squares estimate, replaced by the frozen iteration-4 LSTM's where classical fails or the two ranges differ by more than 50%. See the algorithm note.
• Training. None.

G3 passed, for a hybrid

All seven checks pass on seed 207's test split (20,000 scenarios, σ = 1°):

• Ratios. Every median and P90 of range, course and speed is within 0.97–1.02 × classical, against 1.2.
• Gross errors, five minutes after the zig. They fall from 1.19% to 0.57%: −0.62 pp [−0.82, −0.42].

What it is not. It is not a network matching classical. The LSTM alone is still about 3 × classical on speed. The hybrid's accuracy is classical's (its medians are about 1% worse). What the network adds is robustness where least squares lands far from the truth.

Records:

• the experiment, whose review.md holds the panel;
• the decision;
• the research;
• the write-up.

Seed-207 test split, σ = 1° (median / P90)

Estimator	Range, zig + 300 s	Course, zig + 300 s	Speed, zig + 600 s	Gross, zig + 300 s
hybrid	3.36% / 14.6%	1.38° / 17.5°	0.310 / 1.95 kt	0.57%
classical	3.35% / 14.8%	1.37° / 18.1°	0.307 / 1.92 kt	1.19%
lstm-instantaneous alone	7.16% / 20.8%	4.75° / 23.5°	0.956 / 2.99 kt	0.99%
naive	40.7% / 162%	91.0° / 162°	4.22 / 7.68 kt	38.1%

n = 9,834 at zig + 300 s, 6,483 at zig + 600 s.

G3 checks (paired bootstrap 95% intervals, 2,000 resamples)

Check	Hybrid / classical, or difference	95% CI	Limit
range median / P90	1.003 / 0.989	[1.000, 1.009] / [0.976, 0.999]	1.2
course median / P90	1.011 / 0.966	[1.003, 1.018] / [0.932, 0.983]	1.2
speed median / P90	1.010 / 1.015	[1.002, 1.016] / [1.001, 1.030]	1.2
gross-error rate	−0.620 pp	[−0.824, −0.417]	0

Where the switch helps and hurts

At zig + 300 s, 140 σ = 1° rows switch (1.4%):

• Helps (82 rows). A classical gross error is fixed. These cluster at long range (beyond 12 km, gross errors fall 95 → 23), at low bearing rate and after shallow ownship turns.
• Hurts (21 rows). A good classical row becomes gross. These are mostly at short range (under 4 km, gross errors rise 2 → 20), where classical is sharp and the network's relative range error is large.

At zig + 600 s, which is reported and not gated, the gross rates are 0.17% for the hybrid against 0.14% for classical: 11 against 9 rows, no evidence of a difference.

At σ = 2°, also reported and not gated, the benefit is larger: gross errors fall from 4.63% to 1.80% at zig + 300 s.