Iteration 5: a hybrid passes G3; the network adds robustness, not accuracy

G2 left the models usable but about 3 × worse than the classical least-squares solver. G3 asks for parity: within 1.2 × classical on every metric, and a lower gross-error rate.

Iteration 5 trained nothing. It combined the two estimators: classical's estimate, except where classical fails or the network disagrees by more than 50% on range, where the network's is used.

G3 passes for that hybrid on a fresh test set of 20,000 scenarios (seed 207):

• Parity. Every median and 90th percentile of range, course and speed is within 0.97–1.02 × classical.
• Gross errors. Five minutes after the zig they fall from 1.19% to 0.57% (−0.62 pp [−0.82, −0.42]).
• Validation agreed. It forecast 1.25% → 0.55% on seed 205.

Records: experiments/20261028-i5-g3/. Figures: iteration-5 results.

How we got here

The owner first chose a model lever for G3: a per-leg steady-track fit as an extra input (Annex A B.2.3). Three pieces of research reshaped the plan before anything was trained:

1. The per-leg fit doesn't help where it matters. On a long first leg it measures the range-rate ratio well, about 28% error. At the gate points the current leg is only about four minutes old, and the same quantity is pure noise: 37–64 /h of error against a true value near 1 /h. Classical does better because it fits all the legs together.
2. Starting the solver from the network changes nothing. It gave classical's answer in 99.9% of rows. Classical's gross errors are not bad starting points. They are cases where the least-squares optimum itself is far from the truth: near-unobservable geometry, where any start converges to the same wrong answer. In 91% of those rows the network was still within 50% of the true range.
3. So the hybrid has to sometimes overrule least squares. The disagreement switch does, without a tuned parameter. Its 50% threshold is the gross-error bound itself.

Reading it critically

• It is a hybrid result. The network alone is still 0.956 kt on speed against classical's 0.307 kt. Nothing here says a network matches classical.
• No accuracy gain. The hybrid's medians are about 1% worse than classical's on course and speed: measurable, and immaterial against 1.2. What it adds is fewer gross errors: the 1.4% of rows where it switches.
• The switch aims at the metric that is gated. The threshold equals the gross-error definition. That was fixed in the decision record before seed 207 existed, but it means the lower gross rate is partly by design.
• It costs something at short range. At zig + 300 s the switch fixes 82 classical gross errors and causes 21. Those 21 are mostly under 4 km, where classical is sharp and the network's relative range error is large.
• The gain is early only. Ten minutes after the zig the gross rates are 0.17% against 0.14%: no evidence of a difference. The "lower gross-error rate" is shown at the pre-registered five-minute point only.
• It grows with noise. At σ = 2° gross errors fall from 4.63% to 1.80%. That is the case SRD §5 calls interesting for ML over classical methods.
• Still simulation. The target holds course and speed (SRD §3), and the data are mock.

What next

All four gates are passed: G0, G1, G2, and G3 for the hybrid. The proof of concept has its answer.

• The network alone extracts usable range, course and speed, but is about 3 × worse than classical.
• As a fallback it makes least squares more robust in near-unobservable geometry, without costing accuracy.

SRD §5 says G3 defines the bar for the next phase. Natural next steps are target manoeuvres (where classical's constant-velocity model breaks) and an EKF reference. Both are the owner's choice.