Results: iteration 2c (G2, final attempt)

Question: do the two levers that each moved speed on validation reach the G2 speed threshold together? Iteration 2c (spec 005) combines the speed-weighted loss of iteration 2b with the capacity the scan pointed to (two stacked 128-unit recurrent layers, from scratch) and judges G2 on a third fresh test split (seed 203). It is the last iteration the SRD allows under G2 (§7, amended 2026-10-15 before 2b's test scoring).

The carried set is the 2c models, declared before training; the 2b models are scored beside them on the same split and reported, never gated. Thresholds, slice, coverage and the pass / fail / inconclusive rule are as on the iteration-2 page.

G2 failed; the project stops at G2 (agent review panel, round 2)

Speed fails for both models, by less again. The instantaneous model is at 1.47 kt [1.37, 1.52] and the time-late model at 1.53 kt [1.47, 1.61], against 1 kt; the 2b models on the same scenarios are at 1.78 kt and 1.77 kt, and classical at 0.61 kt, at the Cramér–Rao bound.

Range and course are the best of the project and pass for both models: 13.5% / 7.36° instantaneous and 12.1% / 6.12° time-late. The time-late model is now better than the instantaneous one on range and course, and worse on speed.

Full record: review.md, the validation note validation.md and the no-go write-up.

Experiment 20261016-i2d-g2 (attempt 3; 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: 1.47 kt [1.37, 1.52] and 1.53 kt [1.47, 1.61]); range and course pass for both; iteration 2c is the last allowed under G2 (SRD §7): no-go; 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: 1.47 kt [1.37, 1.52] and 1.53 kt [1.47, 1.61]); range and course pass for both; iteration 2c is the last allowed under G2 (SRD §7): no-go; see review.md. Panel record: review.md.

Check	Result	Detail
g2-instantaneous-range	pass	instantaneous lstm-instantaneous median range error 0.1348 vs threshold 0.2 on sigma=1deg at zig+300s; 95% CI [0.1295, 0.1385] (bootstrap, 2000 resamples, n 4689); model n 4689, excluded 339 (93.3% of the slice reach zig+300s; decision 13); naive 0.3718 (n 4689, excluded 339), classical 0.0876 (n 4689, excluded 339)
g2-instantaneous-course	pass	instantaneous lstm-instantaneous median course error 7.3552° vs threshold 10° on sigma=1deg at zig+300s; 95% CI [7.0453, 7.7018] (bootstrap, 2000 resamples, n 4689); model n 4689, excluded 339 (93.3% of the slice reach zig+300s; decision 13); naive 90.4661 (n 4689, excluded 339), classical 3.4095 (n 4689, excluded 339)
g2-instantaneous-speed	fail	instantaneous lstm-instantaneous median speed error 1.4654 kt vs threshold 1 kt on sigma=1deg at zig+600s; 95% CI [1.3731, 1.5225] (bootstrap, 2000 resamples, n 3038); model n 3038, excluded 1990 (60.4% of the slice reach zig+600s; decision 13); naive 4.4366 (n 3038, excluded 1990), classical 0.6132 (n 3038, excluded 1990)
g2-time-late-range	pass	time-late bilstm-time-late median range error 0.1210 vs threshold 0.2 on sigma=1deg at zig+300s; 95% CI [0.1169, 0.1260] (bootstrap, 2000 resamples, n 4689); model n 4689, excluded 339 (93.3% of the slice reach zig+300s; decision 13); naive 0.3718 (n 4689, excluded 339), classical 0.0876 (n 4689, excluded 339)
g2-time-late-course	pass	time-late bilstm-time-late median course error 6.1210° vs threshold 10° on sigma=1deg at zig+300s; 95% CI [5.8716, 6.4076] (bootstrap, 2000 resamples, n 4689); model n 4689, excluded 339 (93.3% of the slice reach zig+300s; decision 13); naive 90.4661 (n 4689, excluded 339), classical 3.4095 (n 4689, excluded 339)
g2-time-late-speed	fail	time-late bilstm-time-late median speed error 1.5298 kt vs threshold 1 kt on sigma=1deg at zig+600s; 95% CI [1.4680, 1.6117] (bootstrap, 2000 resamples, n 3038); model n 3038, excluded 1990 (60.4% of the slice reach zig+600s; decision 13); naive 4.4366 (n 3038, excluded 1990), classical 0.6132 (n 3038, excluded 1990)
g1-instantaneous	pass	lstm-instantaneous median range error 0.1455 vs limit 0.1881 (0.5 × naive 0.3761) at zig+300s; n 9340 (naive 9340), excluded 660 (naive 660); model/naive 0.387, 95% CI [0.375, 0.397] (paired bootstrap, 2000 resamples, n 9340)
g1-time-late	pass	bilstm-time-late median range error 0.1306 vs limit 0.1881 (0.5 × naive 0.3761) at zig+300s; n 9340 (naive 9340), excluded 660 (naive 660); model/naive 0.347, 95% CI [0.336, 0.357] (paired bootstrap, 2000 resamples, n 9340)
time-late-not-worse	fail	3/19 points worse: t=1020s 0.1313 > 0.1195; t=1080s 0.1325 > 0.1236; t=1140s 0.1361 > 0.1284 (4 later points with n < 500 reported, not gated)
instantaneous-beats-control	pass	lstm-instantaneous vs cnn-causal-control median range error — zig+300s: 0.1455 vs 0.2838, final: 0.1372 vs 0.2904
time-late-beats-control	pass	bilstm-time-late vs mlp-control median range error — zig+300s: 0.1306 vs 0.2733, final: 0.1402 vs 0.2861
shuffled-labels-at-naive	pass	lstm-instantaneous-shuffled median range error 0.3863 = 1.027 × naive 0.3761 (allowed 1 ± 0.1) at zig+300s; course error median 77.6° vs naive 90.4° (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	0.328	[0.314, 0.343]	9340
bilstm-time-late	0.347	[0.336, 0.357]	9340
cnn-causal-control	0.755	[0.738, 0.772]	9340
lstm-instantaneous	0.387	[0.375, 0.397]	9340
lstm-instantaneous-shuffled	1.027	[1.012, 1.042]	9340
mlp-control	0.727	[0.709, 0.745]	9340

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	4689	37.2% [36.3%, 38.3%]	145.4%	90.5° [87.5°, 92.9°]	4.36 kt [4.22 kt, 4.47 kt]	33.7%
classical	4689	8.8% [8.3%, 9.1%]	54.7%	3.4° [3.2°, 3.6°]	1.22 kt [1.16 kt, 1.29 kt]	11.2%
bilstm-time-late	4689	12.1% [11.7%, 12.6%]	35.6%	6.1° [5.9°, 6.4°]	1.67 kt [1.63 kt, 1.74 kt]	4.7%
cnn-causal-control	4689	27.8% [27.0%, 28.4%]	83.5%	22.4° [21.6°, 23.1°]	3.76 kt [3.64 kt, 3.88 kt]	20.7%
lstm-instantaneous	4689	13.5% [12.9%, 13.9%]	36.6%	7.4° [7.0°, 7.7°]	1.80 kt [1.74 kt, 1.87 kt]	4.8%
lstm-instantaneous-shuffled	4689	38.3% [37.3%, 39.2%]	116.5%	77.5° [74.7°, 80.4°]	4.34 kt [4.23 kt, 4.46 kt]	32.8%
mlp-control	4689	27.1% [26.4%, 27.8%]	65.8%	17.0° [16.5°, 17.7°]	3.28 kt [3.20 kt, 3.37 kt]	18.8%

By slice

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

Median range error by slice:

Estimator	0-5pct	5-10pct
naive	37.7% n 4814	37.6% n 4526
classical	12.2% n 4814	12.5% n 4526
bilstm-time-late	13.2% n 4814	12.9% n 4526
cnn-causal-control	28.6% n 4814	28.1% n 4526
lstm-instantaneous	14.9% n 4814	14.2% n 4526
lstm-instantaneous-shuffled	38.5% n 4814	38.7% n 4526
mlp-control	27.5% n 4814	27.1% n 4526

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	33.8%	53.0%	26.6°	3.49 kt	14.6%
cnn-causal-control	9303	52.7%	64.6%	53.2°	4.07 kt	59.4%
lstm-instantaneous	9303	35.4%	55.3%	29.7°	3.52 kt	17.9%
lstm-instantaneous-shuffled	9303	64.6%	73.8%	87.1°	4.26 kt	84.0%
mlp-control	9303	51.4%	66.8%	45.4°	3.94 kt	53.6%

Earlier attempts

Kept on record; the latest attempt is shown above.

Experiment	Machine verdict	Decided by	Model / naive, 95% interval
20261015-i2c-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: 1.77 kt [1.68, 1.83] and 1.77 kt [1.67, 1.86]); range and course pass for both; see review.md	classical 0.316 [0.303, 0.331] bilstm-time-late 0.430 [0.419, 0.443] cnn-causal-control 0.739 [0.721, 0.758] lstm-instantaneous 0.418 [0.405, 0.432] lstm-instantaneous-shuffled 1.000 [0.989, 1.016] mlp-control 0.687 [0.672, 0.707]
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]