benchmark/tables/release_tables/baton/c1_transfer.md

C1 — minimal AIDE -> BATON transfer of the driver trunk (E10), rendered 20260911T202503Z

Generated by eval/report_c1_transfer.py; every number is read from run outputs / results/paired_compare.csv / results/shift_twin.csv (nothing typed by hand). Chain: aide/features/build_aide_kin11.py (11 BATON channels from the AIDE in-cabin RTMW keypoints; deviations documented in its header) -> train/pretrain_u_trunk_aide.py (the survivor DriverBranch trunk in_ln + GRU + mode, self-supervised 2 s -> next 1 s of the 11 channels, Laplace NLL, AIDE train / val) -> train/train_u0_from_init.py --kin-source causal_routes --kin-version v2 --stats-version 2 --variant kin_can_noK_clip8_ln (unchanged train_cwm3.py recipe from the pretrained trunk; head at the seed-k init) vs the scratch survivor U0_kin_can_noK_s<k>_kcv2_clip8_ln on identical causal-v2 rows (eval/paired_compare_joint.py).

1. AIDE self-supervised pretraining (val split; MAE in the pack's crop-normalised units, mean over the 5 horizons and 11 channels)

seed	best epoch	val Laplace NLL	MAE model	MAE copy-K0	MAE constant	MAE model / copy-K0	trunk sha256
0	18	-0.7165	0.0302	0.0318	0.0682	0.951	948f2d084187dcad
1	15	-0.7143	0.0303	0.0318	0.0682	0.953	e6823c6d741f0b75
2	17	-0.7217	0.0302	0.0318	0.0682	0.950	816b45f4f2ad1e3a

Per-horizon MAE (seed 0, model / copy-K0): +0.2 s 0.0185 / 0.0188, +0.4 s 0.0268 / 0.0276, +0.6 s 0.0318 / 0.0333, +0.8 s 0.0355 / 0.0378, +1.0 s 0.0384 / 0.0414

2. Fine-tuned U0 (AIDE trunk) vs scratch survivor — test onset PR-AUC@3s, paired device-cluster bootstrap on identical causal-v2 rows (A = AIDE-init, B = scratch)

regime	event	A s0/s1/s2	B s0/s1/s2	Δ s0 [CI]	Δ s1 [CI]	Δ s2 [CI]	mean Δ ± sd	seeds Δ ≥ +0.02	verdict
engaged	hand_leave	0.360 / 0.341 / 0.365	0.341 / 0.339 / 0.365	+0.018 [-0.005,+0.036]	+0.002 [-0.016,+0.028]	+0.000 [-0.014,+0.016]	0.007 ± 0.010	0/3	INCONCLUSIVE
engaged	hand_return	0.677 / 0.687 / 0.684	0.696 / 0.669 / 0.671	-0.019 [-0.035,+0.009]	+0.018 [-0.013,+0.054]	+0.014 [-0.006,+0.034]	0.004 ± 0.020	0/3	INCONCLUSIVE
engaged	head_turn_onset	0.251 / 0.244 / 0.254	0.259 / 0.257 / 0.255	-0.008 [-0.018,+0.006]	-0.013 [-0.019,-0.003]	-0.001 [-0.017,+0.010]	-0.007 ± 0.006	0/3	INCONCLUSIVE
manual	hand_leave	0.436 / 0.418 / 0.432	0.424 / 0.419 / 0.436	+0.012 [-0.002,+0.019]	-0.001 [-0.014,+0.006]	-0.003 [-0.010,+0.003]	0.002 ± 0.008	0/3	INCONCLUSIVE
manual	hand_return	0.816 / 0.823 / 0.794	0.818 / 0.824 / 0.806	-0.002 [-0.024,+0.012]	-0.000 [-0.018,+0.008]	-0.012 [-0.031,-0.003]	-0.005 ± 0.006	0/3	INCONCLUSIVE
manual	head_turn_onset	0.441 / 0.427 / 0.423	0.448 / 0.439 / 0.446	-0.007 [-0.011,+0.001]	-0.012 [-0.019,+0.008]	-0.023 [-0.030,-0.008]	-0.014 ± 0.008	0/3	INCONCLUSIVE

Other paired cells (A − B, mean over seeds; lower is better for NLL / MAE cells, labels inverted accordingly): engaged hazard_nll|hand_leave -0.000 ± 0.005 INCONCLUSIVE; engaged hazard_nll|hand_return -0.000 ± 0.009 INCONCLUSIVE; engaged hazard_nll|head_turn_onset 0.001 ± 0.003 INCONCLUSIVE; manual hazard_nll|hand_leave 0.002 ± 0.006 INCONCLUSIVE; manual hazard_nll|hand_return -0.004 ± 0.006 INCONCLUSIVE; manual hazard_nll|head_turn_onset 0.003 ± 0.006 INCONCLUSIVE

3. Fine-tuned U0 (AIDE trunk) vs the k0 current-state heuristic (causal v2 K0), same rows

regime	event	A (AIDE-init)	k0	mean Δ ± sd	verdict
engaged	hand_leave	0.355 ± 0.013	0.244 ± 0.000	0.111 ± 0.013	BEATS
engaged	hand_return	0.683 ± 0.005	0.621 ± 0.000	0.062 ± 0.005	INCONCLUSIVE
engaged	head_turn_onset	0.250 ± 0.005	0.238 ± 0.000	0.012 ± 0.005	INCONCLUSIVE
manual	hand_leave	0.429 ± 0.010	0.271 ± 0.000	0.157 ± 0.010	BEATS
manual	hand_return	0.811 ± 0.015	0.691 ± 0.000	0.120 ± 0.015	BEATS
manual	head_turn_onset	0.430 ± 0.009	0.323 ± 0.000	0.108 ± 0.009	BEATS

4. Pre-registered acceptance (engaged hand_leave AND head_turn_onset: Δ ≥ +0.02 on ≥ 2/3 seeds AND kill_checks.beats BEATS; no manual driver cell WORSE)

• engaged hand_leave: seeds with Δ ≥ +0.02: 0/3; beats verdict INCONCLUSIVE; deltas +0.018, +0.002, +0.000 -> not met
• engaged head_turn_onset: seeds with Δ ≥ +0.02: 0/3; beats verdict INCONCLUSIVE; deltas -0.008, -0.013, -0.001 -> not met
• manual driver cells WORSE: none

Verdict: FAIL — no transfer (pre-registered wording).

5. K5 shifted-twin / identity check (eval/shift_twin_c1.py, causal-v2 populations; A = AIDE-init, B = scratch)

regime	event	statistic	A	B	Δ s0/s1/s2	mean Δ	verdict / flag	note
manual	hand_leave	twin_pair_auroc@3s	0.682 ± 0.002	0.681 ± 0.003	+0.006 / -0.002 / -0.002	+0.001	NOT_WORSE	causal v2 populations; same route, dt 30-60 s,
manual	hand_return	latent_probe_prauc_centred	0.778 ± 0.007	0.782 ± 0.012	-0.002 / +0.011 / -0.022	-0.004	ok	causal v2 populations; per-device mean-centred latents
manual	hand_return	latent_probe_prauc	0.827 ± 0.019	0.828 ± 0.016	+0.004 / +0.027 / -0.035	-0.001	—	causal v2 populations; raw latents
manual	head_turn_onset	latent_probe_prauc_centred	0.449 ± 0.014	0.453 ± 0.004	-0.003 / -0.019 / +0.010	-0.004	ok	causal v2 populations; per-device mean-centred latents
manual	hand_leave	latent_probe_prauc_centred	0.489 ± 0.003	0.482 ± 0.004	+0.002 / +0.011 / +0.005	+0.006	ok	causal v2 populations; per-device mean-centred latents
manual	hand_leave	latent_probe_prauc	0.542 ± 0.002	0.538 ± 0.006	+0.001 / +0.011 / -0.001	+0.004	—	causal v2 populations; raw latents
manual	head_turn_onset	latent_probe_prauc	0.481 ± 0.014	0.485 ± 0.005	-0.002 / -0.017 / +0.007	-0.004	—	causal v2 populations; raw latents
manual	hand_return	random_pair_auroc@3s	0.756 ± 0.009	0.761 ± 0.012	-0.009 / +0.009 / -0.016	-0.005	—	causal v2 populations; same positives vs random negatives (i
manual	hand_return	twin_pair_auroc@3s	0.613 ± 0.004	0.620 ± 0.005	+0.001 / -0.010 / -0.011	-0.007	INCONCLUSIVE	causal v2 populations; same route, dt 30-60 s,
manual	head_turn_onset	random_pair_auroc@3s	0.766 ± 0.006	0.781 ± 0.004	-0.013 / -0.013 / -0.019	-0.015	—	causal v2 populations; same positives vs random negatives (i
manual	head_turn_onset	twin_pair_auroc@3s	0.631 ± 0.004	0.633 ± 0.003	+0.002 / +0.001 / -0.010	-0.003	NOT_WORSE	causal v2 populations; same route, dt 30-60 s,
manual	hand_leave	random_pair_auroc@3s	0.843 ± 0.002	0.842 ± 0.003	+0.002 / +0.005 / -0.001	+0.002	—	causal v2 populations; same positives vs random negatives (i
manual	all	identity_probe_acc	0.598 ± 0.001	0.612 ± 0.009	-0.004 / -0.015 / -0.023	-0.014	ok	causal v2 populations; majority-class rate 0.2522; GroupKFol
engaged	hand_leave	latent_probe_prauc_centred	0.307 ± 0.012	0.305 ± 0.010	-0.011 / +0.017 / -0.000	+0.002	ok	causal v2 populations; per-device mean-centred latents
engaged	hand_return	latent_probe_prauc_centred	0.582 ± 0.015	0.573 ± 0.014	-0.012 / +0.042 / -0.002	+0.010	ok	causal v2 populations; per-device mean-centred latents
engaged	hand_return	latent_probe_prauc	0.664 ± 0.031	0.674 ± 0.011	+0.018 / -0.020 / -0.028	-0.010	—	causal v2 populations; raw latents
engaged	head_turn_onset	latent_probe_prauc_centred	0.272 ± 0.015	0.283 ± 0.005	-0.005 / -0.030 / +0.004	-0.010	ok	causal v2 populations; per-device mean-centred latents
engaged	head_turn_onset	latent_probe_prauc	0.305 ± 0.013	0.311 ± 0.006	+0.003 / -0.028 / +0.008	-0.006	—	causal v2 populations; raw latents
engaged	hand_leave	latent_probe_prauc	0.331 ± 0.007	0.329 ± 0.016	-0.007 / +0.025 / -0.011	+0.002	—	causal v2 populations; raw latents
engaged	head_turn_onset	twin_pair_auroc@3s	0.606 ± 0.002	0.606 ± 0.003	+0.005 / -0.003 / -0.002	-0.000	NOT_WORSE	causal v2 populations; same route, dt 30-60 s,
engaged	hand_return	random_pair_auroc@3s	0.792 ± 0.011	0.786 ± 0.012	+0.003 / +0.006 / +0.008	+0.006	—	causal v2 populations; same positives vs random negatives (i
engaged	hand_return	twin_pair_auroc@3s	0.650 ± 0.013	0.640 ± 0.003	-0.006 / +0.013 / +0.025	+0.010	INCONCLUSIVE	causal v2 populations; same route, dt 30-60 s,
engaged	head_turn_onset	random_pair_auroc@3s	0.654 ± 0.010	0.658 ± 0.006	-0.016 / -0.006 / +0.011	-0.004	—	causal v2 populations; same positives vs random negatives (i
engaged	hand_leave	random_pair_auroc@3s	0.823 ± 0.005	0.822 ± 0.004	-0.006 / +0.001 / +0.007	+0.001	—	causal v2 populations; same positives vs random negatives (i
engaged	hand_leave	twin_pair_auroc@3s	0.718 ± 0.004	0.719 ± 0.001	-0.003 / +0.004 / -0.006	-0.001	NOT_WORSE	causal v2 populations; same route, dt 30-60 s,
engaged	all	identity_probe_acc	0.720 ± 0.009	0.735 ± 0.012	-0.011 / -0.022 / -0.011	-0.015	ok	causal v2 populations; majority-class rate 0.1402; GroupKFol

Reading: twin_pair_auroc = P(score(positive) > score(same-route twin negative, Δt 30–60 s, matched vEgo)); identity-only models give 0.5; a gain that appears on random negatives but not on twins rides on identity. identity_probe_acc = device-id accuracy of a linear probe on the frozen 128-d GRU state (flag if A − B > 0.05). latent_probe_prauc[_centred] = linear probe latent -> 3 s onset with / without per-device mean-centring.

6. Provenance

• U0_kin_can_noK_s0_kcv2_clip8_ln_aidepre: trunk U0_aidepre_trunk_s0/trunk.pt sha256 948f2d084187dcad (epoch 18, source train/pretrain_u_trunk_aide.py); loaded 7 tensors (gru, in_ln, mode), head kept at seed-0 init; flags {'kin_source': 'causal_routes', 'kin_version': 'v2', 'stats_version': '2'}; best epoch 9, val 0.1483, kin_source causal_routes v2, stats v2, dms_fields_used []
• U0_kin_can_noK_s1_kcv2_clip8_ln_aidepre: trunk U0_aidepre_trunk_s1/trunk.pt sha256 e6823c6d741f0b75 (epoch 15, source train/pretrain_u_trunk_aide.py); loaded 7 tensors (gru, in_ln, mode), head kept at seed-1 init; flags {'kin_source': 'causal_routes', 'kin_version': 'v2', 'stats_version': '2'}; best epoch 10, val 0.1479, kin_source causal_routes v2, stats v2, dms_fields_used []
• U0_kin_can_noK_s2_kcv2_clip8_ln_aidepre: trunk U0_aidepre_trunk_s2/trunk.pt sha256 816b45f4f2ad1e3a (epoch 17, source train/pretrain_u_trunk_aide.py); loaded 7 tensors (gru, in_ln, mode), head kept at seed-2 init; flags {'kin_source': 'causal_routes', 'kin_version': 'v2', 'stats_version': '2'}; best epoch 7, val 0.1491, kin_source causal_routes v2, stats v2, dms_fields_used []

Deviations of the AIDE channel build from the BATON build (from aide/results/cache/aide_kin11_summary.json):

• references: per-clip causal expanding median over the clip's own valid frames (j <= k), N_MIN=5 (BATON: per-route over moving frames, device pool from earlier routes, N_MIN=25); head_turn_dev / *_rest_dist are deviations from the clip's running median
• camera: AIDE incar 1920x1080 driver-facing camera; absolute x/y offsets differ from the BATON device camera (standardise per channel)
• grid: nearest-frame 10 Hz resampling of 15 fps video (opendriver adapter), no interpolation
• states: hands_on / head_turned computed with the BATON thresholds for description only (not BATON events)
• moving_gate: none (AIDE has no vehicle speed)
• coverage: 2898 clips, valid steps 0.741 (first second 0.492), clips without a valid step 108; hands_on prevalence 0.974, head_turned 0.074 (descriptive, clip-median references).