Phase 4b: a hybrid passes G5 — a learned range with a tuned tracker's velocity

Phase 4's recurrent tracker T2 beat a tuned IMM tracker overall but lost on straight legs, so it did not reach the test. The owner chose a follow-on: a hybrid of the two frozen estimators, judged by G5 unchanged. The hybrid takes T2's range always, and T2's course and speed only where the two estimators' courses differ by more than 45°. Elsewhere it takes the IMM's.

It passed, on a fresh 40,000-scenario test split (seed 213), decided by the agent review panel in two rounds:

Hybrid / IMM, 97.5% interval	Limit	Result
Range, all gate rows	< 1	0.86 [0.84, 0.88]
Course, all gate rows	< 1	0.90 [0.89, 0.92]
Range, settled rows	< 1.1	1.02 [0.99, 1.06]
Course, settled rows	< 1.1	0.98 [0.97, 0.99]
Speed, settled rows	< 1.1	0.99 [0.99, 1.00]

It also makes 3.4 percentage points fewer gross range errors than the IMM.

Records: the gate record and the panel review.

How it was set up

• Nothing was trained. Both estimators were frozen: the IMM since decision 29, T2 since phase 4. tma-g5 checked T2's weights by file digest.
• Why a course disagreement. The IMM lags a manoeuvre by minutes, and while it lags its course disagrees with T2's. A disagreement above θ is therefore a sign that T2's velocity is the better one.
• Choosing θ. θ = 45° was chosen on validation by a criterion stated in advance: the largest smallest margin over the five checks. Every θ from 25° to 90° passed there, so 45° sits on a plateau, not a spike.
• Freezing it. The decision, θ and the test seed were committed and pushed before the test split was scored.
• Forecast: an 85% chance of passing.
• No sign of over-fitting. The three checks that depend on θ moved by less than 0.01 between validation and test.

What it does not show

The panel insisted on these, and they matter as much as the pass.

• It is a hybrid, not a learned tracker. On its own, T2 still fails the straight-leg checks (phase 4). The result shows that a learned estimator adds value beside a tuned tracker, as G3's hybrid did for classical least squares.
• It is three numbers per row, not a track. Range comes from one estimator and velocity from another, so together they do not satisfy either one's bearing-rate relation. Measured afterwards, the cost is small but real: on settled rows, projected 5 min ahead, its bearing is a median of 1.5° off, against 1.0° for the IMM and 1.6° for T2 alone.
• Before a target manoeuvres, its range is worse. On settled rows before any manoeuvre, the hybrid's range error is 6.0% against the IMM's 4.1% (1.47 ×). After a manoeuvre it is 10.3% against 15.2% (0.68 ×). The pooled settled check passes only because the second group carries it, even though the first group is 55% of settled rows. For a target holding course, the commonest case, a tuned filter is better.
• Against the EKF it would fail two settled checks: course 1.44 × and speed 1.21 ×. The IMM was the pre-registered reference, chosen on overall range, so the gate stands. But "better than a tuned tracker on straight legs" is not shown. Batch least squares is also better on settled range: 6.6% against 7.6%.
• The settled velocity checks pass largely by construction. The hybrid uses the IMM's velocity on most rows.
• The rule family was chosen after seeing phase 4's result. The fresh seed controls the choice of θ, not the choice of family.

The long view

The programme's question was whether a network trained on simulated bearings could give usable range, course and speed. After G0 to G5, the answer is consistent.

• Where networks win. They are better where tracking is hardest: after target manoeuvres, and late in long engagements. In range they reach about 0.7 × a tuned tracker after a manoeuvre.
• Where they lose. On clean geometry, a target holding course after an ownship zig, a tuned filter or a batch fit is better, especially on speed.
• What works. Combining them works, as a hybrid that takes each one's strength. A network alone does not.

Real or recorded data remain the untested step.