Results: iteration 2b (G2, second attempt)

Question: does weighting speed in the loss get the iteration-2 models to the G2 speed threshold? Iteration 2b (spec 004) keeps the data, the inputs and the architectures of iteration 2 and changes one thing: every model is continued from its committed phase-B weights with the speed term of the loss weighted 4×, to 220 epochs at most with patience 6. The best epoch is re-scored under the new loss before training resumes, so the continuation has to beat the old weights on the new scale.

The gate is judged on a fresh test split (seed 202, same count and ranges), pinned by content digest in the training plan. The iteration-2 test split (seed 201) decided the first attempt and is not used again. The 2b models are the carried set by a rule fixed before training; the phase-B weights are scored beside them on the same split and reported, never gated.

The thresholds, the σ = 1° slice, the zig + 600 s coverage and the pass / fail / inconclusive rule are as on the iteration-2 page.

G2 failed again (agent review panel, round 2)

Speed still fails for both models, by less. The instantaneous model is at 1.77 kt [1.68, 1.83] and the time-late model at 1.77 kt [1.67, 1.86], against 1 kt; phase B on the same scenarios is at 2.02 kt and 2.31 kt. Classical reaches 0.60 kt, at the Cramér–Rao bound.

Range and course pass for both models: 14.6% / 8.22° instantaneous and 15.8% / 8.66° time-late. The time-late model's course failure from iteration 2 is gone, and it now passes its G1 ratio check cleanly. Its "not worse than instantaneous" check fails at 2 of 19 points without an interval; the paired intervals contain the slack. Read it as the G1 caveat persisting: the time-late model is slightly worse than the instantaneous one at 17 of 19 points, not better.

By range band the models are ahead of classical beyond 10 km only at zig + 300 s, and on the σ = 1° slice only because classical's non-converged scenarios count against it; at zig + 600 s classical is better in every band. Target speeds above ownship's 15 kt map to normalised targets above 1, where the models' shrinkage is worst; the output scaling was not changed in 2b.

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

Experiment 20261015-i2c-g2 (attempt 2; earlier attempts below): machine verdict G2 fail (agent review panel, round 2 (no rejections; no open blocker or major findings): G2 FAIL (SRD §5); speed fails for both variants (CI wholly above 1 kt: 1.77 kt [1.68, 1.83] and 1.77 kt [1.67, 1.86]); range and course pass for both; see review.md).

Decided by agent review panel, round 2 (no rejections; no open blocker or major findings): G2 FAIL (SRD §5); speed fails for both variants (CI wholly above 1 kt: 1.77 kt [1.68, 1.83] and 1.77 kt [1.67, 1.86]); range and course pass for both; see review.md. Panel record: review.md.

Check	Result	Detail
g2-instantaneous-range	pass	instantaneous lstm-instantaneous median range error 0.1461 vs threshold 0.2 on sigma=1deg at zig+300s; 95% CI [0.1413, 0.1520] (bootstrap, 2000 resamples, n 4682); model n 4682, excluded 329 (93.4% of the slice reach zig+300s; decision 13); naive 0.3787 (n 4682, excluded 329), classical 0.0845 (n 4682, excluded 329)
g2-instantaneous-course	pass	instantaneous lstm-instantaneous median course error 8.2203° vs threshold 10° on sigma=1deg at zig+300s; 95% CI [7.8796, 8.5362] (bootstrap, 2000 resamples, n 4682); model n 4682, excluded 329 (93.4% of the slice reach zig+300s; decision 13); naive 90.4037 (n 4682, excluded 329), classical 3.3111 (n 4682, excluded 329)
g2-instantaneous-speed	fail	instantaneous lstm-instantaneous median speed error 1.7669 kt vs threshold 1 kt on sigma=1deg at zig+600s; 95% CI [1.6839, 1.8334] (bootstrap, 2000 resamples, n 2973); model n 2973, excluded 2038 (59.3% of the slice reach zig+600s; decision 13); naive 4.2236 (n 2973, excluded 2038), classical 0.6025 (n 2973, excluded 2038)
g2-time-late-range	pass	time-late bilstm-time-late median range error 0.1579 vs threshold 0.2 on sigma=1deg at zig+300s; 95% CI [0.1521, 0.1622] (bootstrap, 2000 resamples, n 4682); model n 4682, excluded 329 (93.4% of the slice reach zig+300s; decision 13); naive 0.3787 (n 4682, excluded 329), classical 0.0845 (n 4682, excluded 329)
g2-time-late-course	pass	time-late bilstm-time-late median course error 8.6621° vs threshold 10° on sigma=1deg at zig+300s; 95% CI [8.2859, 9.0289] (bootstrap, 2000 resamples, n 4682); model n 4682, excluded 329 (93.4% of the slice reach zig+300s; decision 13); naive 90.4037 (n 4682, excluded 329), classical 3.3111 (n 4682, excluded 329)
g2-time-late-speed	fail	time-late bilstm-time-late median speed error 1.7729 kt vs threshold 1 kt on sigma=1deg at zig+600s; 95% CI [1.6731, 1.8582] (bootstrap, 2000 resamples, n 2973); model n 2973, excluded 2038 (59.3% of the slice reach zig+600s; decision 13); naive 4.2236 (n 2973, excluded 2038), classical 0.6025 (n 2973, excluded 2038)
g1-instantaneous	pass	lstm-instantaneous median range error 0.1579 vs limit 0.1886 (0.5 × naive 0.3772) at zig+300s; n 9342 (naive 9342), excluded 658 (naive 658); model/naive 0.418, 95% CI [0.405, 0.432] (paired bootstrap, 2000 resamples, n 9342)
g1-time-late	pass	bilstm-time-late median range error 0.1620 vs limit 0.1886 (0.5 × naive 0.3772) at zig+300s; n 9342 (naive 9342), excluded 658 (naive 658); model/naive 0.430, 95% CI [0.419, 0.443] (paired bootstrap, 2000 resamples, n 9342)
time-late-not-worse	fail	2/19 points worse: t=840s 0.1390 > 0.1318; t=960s 0.1483 > 0.1405 (4 later points with n < 500 reported, not gated)
instantaneous-beats-control	pass	lstm-instantaneous vs cnn-causal-control median range error — zig+300s: 0.1579 vs 0.2787, final: 0.1539 vs 0.2849
time-late-beats-control	pass	bilstm-time-late vs mlp-control median range error — zig+300s: 0.1620 vs 0.2593, final: 0.1685 vs 0.2761
shuffled-labels-at-naive	pass	lstm-instantaneous-shuffled median range error 0.3774 = 1.000 × naive 0.3772 (allowed 1 ± 0.1) at zig+300s; course error median 70.0° vs naive 90.8° (reported, not gated)

Model / naive at the G1 point

Median range error of each model divided by naive's, 5 minutes after the zig, with a paired bootstrap 95% interval. G1 passes only if the whole interval is below 0.5; an interval that straddles 0.5 is inconclusive.

Model	Ratio	95% interval	n
classical	0.316	[0.303, 0.331]	9342
bilstm-time-late	0.430	[0.419, 0.443]	9342
cnn-causal-control	0.739	[0.721, 0.758]	9342
lstm-instantaneous	0.418	[0.405, 0.432]	9342
lstm-instantaneous-shuffled	1.000	[0.989, 1.016]	9342
mlp-control	0.687	[0.672, 0.707]	9342

Error against time since the zig

Dashed lines are the naive and classical baselines; test split, all scenarios.

Metric: Median range errorP90 range errorMedian course errorMedian speed errorGross-error rate

• ━ naive
• ━ classical
• ━ bilstm-time-late
• ━ cnn-causal-control
• ━ lstm-instantaneous
• ━ lstm-instantaneous-shuffled
• ━ mlp-control

Metrics by noise slice and point

Slice: overalldropout_band=0-5pctdropout_band=5-10pctsigma=1degsigma=2deg Point: end of windowzig + 300 szig + 600 s

Brackets: bootstrap 95% interval of the median (σ and overall slices at zig + 300 s and zig + 600 s).

Estimator	n	Median range error	P90 range error	Median course error	Median speed error	Gross-error rate
naive	4682	37.9% [36.9%, 38.8%]	151.9%	90.4° [88.1°, 93.2°]	4.17 kt [4.05 kt, 4.27 kt]	34.8%
classical	4682	8.4% [8.1%, 8.9%]	55.3%	3.3° [3.1°, 3.6°]	1.17 kt [1.11 kt, 1.24 kt]	11.1%
bilstm-time-late	4682	15.8% [15.2%, 16.2%]	42.4%	8.7° [8.3°, 9.0°]	2.12 kt [2.05 kt, 2.19 kt]	6.9%
cnn-causal-control	4682	28.0% [26.9%, 28.8%]	72.9%	22.0° [21.0°, 22.7°]	3.67 kt [3.55 kt, 3.79 kt]	18.7%
lstm-instantaneous	4682	14.6% [14.1%, 15.2%]	41.2%	8.2° [7.9°, 8.5°]	2.11 kt [2.04 kt, 2.18 kt]	6.8%
lstm-instantaneous-shuffled	4682	38.4% [37.3%, 39.4%]	123.6%	70.6° [68.3°, 73.3°]	4.15 kt [4.03 kt, 4.25 kt]	33.8%
mlp-control	4682	26.3% [25.5%, 27.1%]	63.4%	15.1° [14.5°, 15.8°]	3.24 kt [3.13 kt, 3.33 kt]	16.1%

By slice

Slice: dropout_bandrange_bandsigmazig_band Point: finalzig+300szig+600s

Median range error by slice:

Estimator	0-5pct	5-10pct
naive	38.7% n 4718	36.8% n 4624
classical	11.8% n 4718	12.1% n 4624
bilstm-time-late	16.1% n 4718	16.4% n 4624
cnn-causal-control	28.6% n 4718	27.1% n 4624
lstm-instantaneous	15.8% n 4718	15.7% n 4624
lstm-instantaneous-shuffled	38.4% n 4718	37.0% n 4624
mlp-control	26.1% n 4718	25.8% n 4624

Held-out geometry (reported, not gated)

Estimator	n	Median range error	P90 range error	Median course error	Median speed error	Gross-error rate
naive	9303	59.6%	69.9%	90.0°	4.24 kt	76.2%
classical	9303	48.8%	∞	50.9°	9.21 kt	49.3%
bilstm-time-late	9303	40.8%	56.6%	32.1°	3.58 kt	24.7%
cnn-causal-control	9303	54.3%	65.6%	49.3°	4.10 kt	63.9%
lstm-instantaneous	9303	38.6%	57.1%	34.5°	3.58 kt	23.2%
lstm-instantaneous-shuffled	9303	63.9%	73.3%	84.3°	4.23 kt	83.0%
mlp-control	9303	50.5%	65.0%	42.9°	3.95 kt	51.3%

Earlier attempts

Kept on record; the latest attempt is shown above.

Experiment	Machine verdict	Decided by	Model / naive, 95% interval
20261014-i2-g2 (review)	fail	agent review panel, round 2 (no rejections; no open blocker or major findings): G2 FAIL (SRD §5); speed fails for both variants (CI wholly above 1 kt), time-late course and g1-time-late also fail; see review.md	classical 0.330 [0.316, 0.344] bilstm-time-late 0.514 [0.500, 0.529] cnn-causal-control 0.747 [0.732, 0.766] lstm-instantaneous 0.472 [0.458, 0.484] lstm-instantaneous-shuffled 1.015 [1.001, 1.029] mlp-control 0.703 [0.686, 0.720]