K-v2 programme — CockpitWM-K improvement, pre-registered 2026-09-14 (protocol C23) — val n = 405 (selection), test n = 609 (one evaluation per run), paired per-clip bootstrap (10000)
Generated 20260914T234626Z by aide/eval/report_kv2_program.py from pred_{val,test}.npz / metrics_{val,test}.json / run json / aide/results/kv2/*.json only (locked protocol: driver first, grid floor, val HM top 40 (threshold 83.33 px, NEVER a ranking statistic), test HM top 60 (threshold 72.05 px), PCK @ 96 px; every run selected on val All-MPJPE of the EMA weights only, test evaluated once by the harness). Δ = paired per-clip difference (negative = better for errors, positive = better for PCK), CI = clip-bootstrap percentile interval. Verdict rule (as arch_attribution.md): BEATS / WORSE = Δ CI excl. 0 on ≥ ceil(2n/3) seeds (5 seeds: 4) with the same sign on every seed; NOT_WORSE = all CIs cover 0 and |Δ| < 0.25; else INCONCLUSIVE. Confirmatory cells need all 5 seeds; otherwise "pending".
0. Pre-registration (verbatim decisions row) and timestamp check
- S0 UTC: 2026-09-14T22:51:07Z; earliest programme run
t_start_utc: 2026-09-14T23:08:22Z (21 programme runs finished of 188 possible names); runs started at or before S0: none.
| 2026-09-14 | K-v2 programme — S0 pre-registration (CockpitWM-K improvement; new model file aide/models/cwm3d_v2.py, --model cwm3d_v2; approved plan <local-notes>/plans/elegant-dazzling-stallman.md Part A; row written at S0 UTC 2026-09-14T22:51:07Z BEFORE any K-v2 / MB-grid run; the same UTC is in aide/results/kv2/S0_UTC.txt and protocol C23). Honesty clause: the K-v2 design (hierarchical part trunk kept; joint tokens + 2 spatial blocks + attention pooling; lean shared read-out; onset-hazard head gating the anchor mixture) was motivated by the TEST diagnostics of arch_attribution.md §3.2 (K vs MotionBERT-kp: right hand +0.71, body +0.28, h=2–4 +0.2–0.3, HM h=3/4 +2.4/+2.0) → therefore ALL K-v2 configuration choices are fixed on VAL only, and MotionBERT-kp (MB) receives the same val selection budget (S3, 10 val points each). Locked protocol unchanged: val-only selection on All-MPJPE of the EMA weights; the official test split is evaluated exactly once per run by the existing harness (append-only results.csv); no flip TTA; ensembles only as separately labelled rows; Stage-A recipe verbatim (--lr 3e-4 --wd 0.05 --warmup 10 --epochs 200 --bs 32 --ema 0.999 --clip 1.0 --patience 40 --aug-window 1 --noise-px 2.0 --joint-drop 0.1 --translate 20 --scale 0.05 --flip-p 0; K loss defaults euc + 0.1 NLL + 0.05 bone + 0.01 jerk; tensor-output models add --loss euc); paired per-clip clip-bootstrap 10,000 resamples seed 0 (aide/eval/metrics.paired_diff); verdict rule of arch_attribution.md: BEATS / WORSE = Δ CI excl. 0 on ≥ ceil(2n/3) seeds (5 seeds: 4/5) with the same sign on every seed; NOT_WORSE = all CIs cover 0 and |Δ| < 0.25; else INCONCLUSIVE. Every decision below is computed mechanically by aide/eval/kv2_gate.py from pred_val.npz + harness.load_split("val") (+ metrics.hm_mask on val) and written as UTC-stamped json in aide/results/kv2/ (s1_decision.json, s1b_<cfg>_decision.json, s3_decision.json, pretrain_check_<cfg>.json); downstream runs may start only after the corresponding json exists. *S1 val screen (seed 42, n = 405)**: 9 configs vs the existing cwm3d_kp_s42_v2 (val All 55.367): cwm3d_v2_s42_v2 (full), cwm3d_v2_nospatial_s42_v2 (--no-v2-spatial), _khead (--no-lean-head, K's per-part read-out), _nohaz (--no-hazard), _hazaux (--haz-mod 0: hazard BCE on, gating off), _plain (--no-physics --w-nll 0), _q060 / _q090 (--haz-quantile 0.60 / 0.90), _lam03 (--haz-weight 0.3); statistic = paired per-clip Δ val All (config − K); SURVIVE iff mean Δ ≤ −0.30 px (≈ 1.4 val seed-sd); at most 2 survivors (ranked by mean Δ; candidates within 0.10 px of each other are ties → the one with FEWER components wins, components = enabled optional modules of {spatial blocks, lean head, hazard BCE, hazard gating, physics+NLL} — full / q060 / q090 / lam03 = 5, nospatial / khead / hazaux / plain = 4, nohaz = 3; remaining ties by name); HM guard (never a selection statistic): KILL iff Δ val HM (40 clips) mean > +3.0 px AND its CI excludes 0. The 40-clip val HM subset is never used for ranking (level CI width ≈ 37 px vs 8.3 for val All; seed sd 0.93 vs 0.21; val HM threshold 83.3 px ≠ test 72.05 px). Test numbers of S1 runs are produced once by the trainer as usual and are RECORDED, never used for any decision. S1b: each survivor at seed 123 vs cwm3d_kp_s123_v2; CONTINUE to seeds 456 / 789 / 1011 iff Δ val(123) < 0 AND the two-seed mean ≤ −0.25 px; otherwise stop and report "val-killed at seed 123". **S2 (5 seeds 42 / 123 / 456 / 789 / 1011), without and with BATON trunk-only init**: K-v2's tokenizer / trunk tensors differ from K's → own pretraining cwm3d_v2_batonpre_pretrain_s123 (pretrain_cwm3d_baton.py --model cwm3d_v2 --seed 123: same corpus sha256 465dac1cf0a8265a, bs 256, 6 epochs, held-out-route selection, --zeroshot 0 = no zero-shot test evaluation) UNLESS kv2_gate.py pretrain-check <cfg> (dry train_kp.load_init_ckpt(parts="trunk") of the survivor model on CPU against aide/results/runs/cwm3d_kp_batonpre_pretrain_s123/ckpt_best.pt) reports n_skipped_shape = 0 AND zero missing TRUNK_PREFIXES tensors, in which case K's checkpoint is reused; fine-tunes cwm3d_v2_<cfg>_batonpre_trunkp123_s<k>_v2 (--init-parts trunk); K's own +pre arm is completed to 5 seeds with cwm3d_kp_batonpre_trunkp123_s{789,1011}_v2 (recipe and init of the existing s42/123/456 runs verbatim). **S3 MB fairness arm**: (i) same-budget val grid at seed 42: motionbert_kp_{d192, L4, lr1e-4, lr1e-3, wd0.01, wd0.1, do0.1, d192_L4, d192_L4_do0.1}_s42_v2 (--model motionbert_kp[_d192][_L4] name-encoded width / depth, --lr, --wd, --dropout flags; --loss euc) + the existing default motionbert_kp_s42_v2 = 10 val points = S1's 10; **MB-best = argmin val All** (ties within 0.10 px → fewer parameters); if MB-best ≠ default it is run at the 5 seeds; (ii) MB-best + BATON trunk-only init at 5 seeds: motionbert_kp_batonpre_trunkp123_s{789,1011}_v2 are run UNCONDITIONALLY (completing the attribution table); if MB-best changes lr / wd / dropout only, the existing motionbert_kp_batonpre_pretrain_s123 checkpoint is reused (motionbert_kp_<cfg>_batonpre_trunkp123_s<k>_v2); if MB-best changes d or L, a new pretraining motionbert_kp_<tag>_batonpre_pretrain_s123 (same corpus / schedule) precedes 5 fine-tunes. **K-v2(best)** = the S2 survivor with the lowest mean VAL All over its finished seeds (ties 0.10 → fewer components); with one survivor it is that one. **PRIMARY (two co-primary endpoints, test All-MPJPE, 5 seeds, both sides in the same pretraining state, beats rule above)**: P1 = K-v2(best) − MB-best (no pretraining); P2 = K-v2(best)+pre − MB-best+pre. **ADOPTION**: A1 = K-v2(best) − cwm3d_kp_s<k>_v2 BEATS on All over 5 seeds → K-v2 becomes the main-table CockpitWM-K row (K's row itself unchanged), otherwise K-v2 is an ablation row; A2 = K-v2(best)+pre − K+pre (cwm3d_kp_batonpre_trunkp123_s<k>_v2) BEATS → the "+ BATON trunk" row. **SECONDARY (descriptive with CIs, never an adoption basis)**: d-n, HM, PCK@0.05, HM h=5, per-part, per-horizon, AUC of the hazard probability haz_P against HM membership. **Paper-wording consequence table** (mirrors the 2026-09-14 attribution row): P1 ∧ P2 BEATS → an architecture advantage under the keypoint formulation may be written FOR K-v2 (K's own attribution sentence unchanged); P1 BEATS ∧ ¬P2 → "advantage without pretraining that the BATON trunk equalises", attribution sentence kept; ¬P1 ∧ P2 → "advantage conditional on the pretrained trunk", attribution sentence kept; neither → wording unchanged verbatim, K-v2 = ablation family + gen_tables.py NEGATIVES entry; any WORSE is reported as is. **Multiplicity**: ≤ 9 (S1) + 18 (S2) + 2 (K+pre seeds) + 9 (S3 i) + 4 (MB-best seeds) + 7 (S3 ii) ≈ 49 new single test evaluations, each required by the protocol, none used for selection; confirmatory claims are exactly P1, P2, A1, A2. **Already published numbers untouched**; all results only in the new table aide/results/tables/kv2_program.md (+ .json) written by aide/eval/report_kv2_program.py (reads only pred_{val,test}.npz, metrics_{val,test}.json, run json; append-only; machine-readable line **ADOPTION A1: ADOPTED\|NOT ADOPTED**). **Execution**: aide/train/run_kv2_program.sh lanes A (S1 → gate s1), B (S3 i → gate s3 → MB-best seeds / +pre), C (motionbert_kp_batonpre_trunkp123_s{789,1011}_v2, cwm3d_kp_batonpre_trunkp123_s{789,1011}_v2), D (pretrain-check → v2 pretraining if needed → S1b → gate s1b → seeds ± pre); every launch passes the project gate (free VRAM ≥ need + 4 GB, MemAvailable ≥ 28 GB, `pgrep -fc "python.(train_cwm3|train_kp|pretrain)"< 4, OMP/MKL 4;setsid nohup bash; a run whose run json exists is never re-run; pids in aide/results/kv2/pids.txt); --no-lock --max-minutes 45, --notes "kv2_program;;cfg="; --refs zvshrink=zv_shrink_val,simlpe=simlpe_kp_s<k or 42>_v2,kp=cwm3d_kp_s_v2,mb=motionbert_kp_s_v2(+seedinit=_s_v2for +pre runs; refs are evaluation-only). Lanes A / D start only afteraide/models/cwm3d_v2.pyexists andpytest -q aide/tests/test_cwm3d_v2.pypasses. Honest expectation (recorded): HM likely improves, All likely still ties MB; every result enters the table regardless. | plan Part A §A.2–A.4;aide/eval/kv2_gate.py, aide/eval/report_kv2_program.py, aide/train/run_kv2_program.sh, aide/tests/test_kv2.py, aide/models/motionbert_kp.py::build(name variants),aide/docs/protocol.mdC23,aide/results/kv2/S0_UTC.txt; motivation aide/results/tables/arch_attribution.md§3.2 | Every K-v2 / MB-grid run json must havet_start_utclater than 2026-09-14T22:51:07Z (checked byreport_kv2_program.py` §0); gate json must predate the downstream launches; "S1 outcome" and final "K-v2 outcome" rows + C23 outcome follow; paper text changes only after the outcome, per the consequence table |
1. S1 val screen (seed 42; statistic = paired per-clip Δ val All vs cwm3d_kp_s42_v2; SURVIVE iff mean Δ ≤ −0.30 px; ≤ 2 survivors, ties 0.10 → fewer components; HM guard KILL iff Δ val HM > +3.0 with CI excl. 0 — the val HM column is SECONDARY)
- Decision source:
aide/results/kv2/s1_decision.json written 2026-09-14T23:45:48Z (complete = True). Rule: paired per-clip Δ val All (config − cwm3d_kp_s42_v2), n = 405, 10000 clip-bootstrap seed 0; SURVIVE iff mean Δ ≤ -0.3 px; ≤ 2 survivors ranked by mean Δ, ties within 0.1 px -> fewer components (full/q060/q090/lam03 = 5, nospatial/khead/hazaux/plain = 4, nohaz = 3); HM guard (never a ranking statistic): KILL iff Δ val HM (40 clips) mean > +3.0 AND CI excl. 0.
| cfg |
run |
flags |
params |
best ep / epochs / stop / GPU min |
val All |
Δ val All [CI] |
Δ val HM [CI] (secondary) |
guard |
S1 status |
test All (recorded, not used) |
| K (baseline) |
cwm3d_kp_s42_v2 |
— |
5771449 |
58 / 99 / early_stop(patience=40) / 0.6 |
55.367 |
0 |
0 |
— |
— |
48.767 |
| full |
cwm3d_v2_s42_v2 |
(full) |
2736282 |
75 / 116 / early_stop(patience=40) / 2.25 |
55.170 |
-0.197 [-0.798,+0.387] |
+0.60 [-2.54,+3.45] |
ok |
KILL |
48.696 |
| nospatial |
cwm3d_v2_nospatial_s42_v2 |
--no-v2-spatial |
2666906 |
87 / 128 / early_stop(patience=40) / 1.69 |
54.788 |
-0.579 [-1.275,+0.058] |
-0.64 [-4.19,+1.99] |
ok |
SURVIVE |
48.267 |
| khead |
cwm3d_v2_khead_s42_v2 |
--no-lean-head |
5091964 |
52 / 93 / early_stop(patience=40) / 4.75 |
55.482 |
+0.115 [-0.442,+0.662] |
+1.06 [-1.16,+3.09] |
ok |
KILL |
48.642 |
| nohaz |
cwm3d_v2_nohaz_s42_v2 |
--no-hazard |
2702199 |
106 / 147 / early_stop(patience=40) / 5.71 |
55.325 |
-0.042 [-0.873,+0.723] |
+2.94 [-0.38,+6.86] |
ok |
KILL |
48.648 |
| hazaux |
cwm3d_v2_hazaux_s42_v2 |
--haz-mod 0 |
2736267 |
87 / 128 / early_stop(patience=40) / 4.51 |
55.140 |
-0.227 [-0.888,+0.393] |
+2.41 [+0.04,+4.95] |
ok |
KILL |
48.592 |
| plain |
cwm3d_v2_plain_s42_v2 |
--no-physics --w-nll 0 |
2736282 |
65 / 106 / early_stop(patience=40) / 3.69 |
55.218 |
-0.149 [-0.722,+0.397] |
+1.73 [-0.72,+4.18] |
ok |
KILL |
48.728 |
| q060 |
cwm3d_v2_q060_s42_v2 |
--haz-quantile 0.60 |
2736282 |
76 / 117 / early_stop(patience=40) / 4.32 |
55.320 |
-0.047 [-0.674,+0.606] |
+3.26 [+0.21,+7.50] |
KILL |
KILL |
48.704 |
| q090 |
cwm3d_v2_q090_s42_v2 |
--haz-quantile 0.90 |
2736282 |
71 / 112 / early_stop(patience=40) / 4.19 |
55.128 |
-0.239 [-0.774,+0.291] |
+1.33 [-1.27,+4.02] |
ok |
KILL |
48.380 |
| lam03 |
cwm3d_v2_lam03_s42_v2 |
--haz-weight 0.3 |
2736282 |
87 / 128 / early_stop(patience=40) / 5.51 |
55.019 |
-0.348 [-1.018,+0.310] |
+1.92 [-0.73,+5.20] |
ok |
SURVIVE |
48.512 |
S1 survivors (decision): nospatial, lam03.
1.1 S1b (survivors at seed 123 vs cwm3d_kp_s123_v2; CONTINUE iff Δ val(123) < 0 and 2-seed mean ≤ −0.25)
| cfg |
seed-123 run |
val All |
Δ val All s42 [CI] |
Δ val All s123 [CI] |
2-seed mean |
decision |
source |
| nospatial |
cwm3d_v2_nospatial_s123_v2 |
pending |
-0.579 [-1.275,+0.058] |
pending |
pending |
PENDING |
live |
| lam03 |
cwm3d_v2_lam03_s123_v2 |
pending |
-0.348 [-1.018,+0.310] |
pending |
pending |
PENDING |
live |
S1b: continuing to 5 seeds: none yet.
2. S2 — 5 seeds (42 / 123 / 456 / 789 / 1011), TEST (one evaluation per run), K-v2 survivors without / with BATON trunk-only init
- Pretraining
cwm3d_v2_batonpre_pretrain_s123: not run / not finished; pretrain-check: —.
(no K-v2 config has reached S2 yet)
3. S3 — MotionBERT-kp fairness arm (same val budget: 9 new grid points + existing default = 10; MB-best = argmin val All, ties 0.10 → fewer params)
- Decision source:
aide/results/kv2/s3_decision.json written 2026-09-14T23:08:40Z (complete = False). Rule: MB-best = argmin val All over the 10 grid points (9 new + existing default); ties within 0.1 px -> fewer parameters.
| cfg |
run |
model |
flags |
params |
best ep / epochs / stop / GPU min |
val All |
Δ val All vs default [CI] |
test All (recorded, not used) |
S3 |
| default |
motionbert_kp_s42_v2 |
motionbert_kp |
(default) |
813088 |
81 / 122 / early_stop(patience=40) / 4.6 |
55.550 |
+0.000 [+0.000,+0.000] |
48.279 |
MB-best |
| d192 |
motionbert_kp_d192_s42_v2 |
motionbert_kp_d192 |
(default) |
— |
pending |
pending |
pending |
— |
PENDING |
| L4 |
motionbert_kp_L4_s42_v2 |
motionbert_kp_L4 |
(default) |
— |
pending |
pending |
pending |
— |
PENDING |
| lr1e-4 |
motionbert_kp_lr1e-4_s42_v2 |
motionbert_kp |
--lr 1e-4 |
— |
pending |
pending |
pending |
— |
PENDING |
| lr1e-3 |
motionbert_kp_lr1e-3_s42_v2 |
motionbert_kp |
--lr 1e-3 |
— |
pending |
pending |
pending |
— |
PENDING |
| wd0.01 |
motionbert_kp_wd0.01_s42_v2 |
motionbert_kp |
--wd 0.01 |
— |
pending |
pending |
pending |
— |
PENDING |
| wd0.1 |
motionbert_kp_wd0.1_s42_v2 |
motionbert_kp |
--wd 0.1 |
— |
pending |
pending |
pending |
— |
PENDING |
| do0.1 |
motionbert_kp_do0.1_s42_v2 |
motionbert_kp |
--dropout 0.1 |
— |
pending |
pending |
pending |
— |
PENDING |
| d192_L4 |
motionbert_kp_d192_L4_s42_v2 |
motionbert_kp_d192_L4 |
(default) |
— |
pending |
pending |
pending |
— |
PENDING |
| d192_L4_do0.1 |
motionbert_kp_d192_L4_do0.1_s42_v2 |
motionbert_kp_d192_L4 |
--dropout 0.1 |
— |
pending |
pending |
pending |
— |
PENDING |
MB-best = default (provisional — S3 incomplete or no decision json; model motionbert_kp, flags []; changes d/L: False → existing motionbert_kp_batonpre_pretrain_s123 checkpoint reused).
3.1 MB-best and MB-best + BATON trunk-only init at 5 seeds; K + pre completion (test levels)
| seed |
arm |
run |
params |
best ep / epochs / stop |
GPU min |
val All |
All [CI] |
d-n [CI] |
HM [CI] |
PCK@0.05 [CI] |
HM h=5 [CI] |
| 42 |
MB-best |
motionbert_kp_s42_v2 |
813088 |
81 / 122 / early_stop(patience=40) |
4.6 |
55.550 |
48.28 [44.53, 52.37] |
2.192 [2.021, 2.377] |
130.34 [108.20, 158.53] |
87.17 [86.04, 88.28] |
163.95 [119.85, 223.86] |
| 42 |
MB-best + pre |
motionbert_kp_batonpre_trunkp123_s42_v2 |
813088 |
37 / 78 / early_stop(patience=40) |
1.94 |
54.528 |
47.74 [44.07, 51.78] |
2.167 [2.000, 2.350] |
128.03 [105.81, 156.47] |
87.46 [86.37, 88.54] |
159.20 [115.09, 219.85] |
| 42 |
K + pre |
cwm3d_kp_batonpre_trunkp123_s42_v2 |
5771449 |
31 / 72 / early_stop(patience=40) |
0.48 |
55.003 |
47.98 [44.28, 52.04] |
2.178 [2.010, 2.363] |
128.43 [106.25, 157.11] |
87.21 [86.10, 88.30] |
159.35 [115.56, 219.46] |
| 123 |
MB-best |
motionbert_kp_s123_v2 |
813088 |
58 / 99 / early_stop(patience=40) |
4.36 |
55.627 |
48.38 [44.65, 52.46] |
2.196 [2.027, 2.381] |
129.16 [106.70, 157.58] |
87.14 [86.02, 88.27] |
162.59 [118.87, 222.08] |
| 123 |
MB-best + pre |
motionbert_kp_batonpre_trunkp123_s123_v2 |
813088 |
38 / 79 / early_stop(patience=40) |
2.16 |
54.720 |
47.53 [43.87, 51.55] |
2.158 [1.992, 2.340] |
126.11 [103.96, 154.69] |
87.40 [86.32, 88.46] |
156.47 [113.00, 215.99] |
| 123 |
K + pre |
cwm3d_kp_batonpre_trunkp123_s123_v2 |
5771449 |
33 / 74 / early_stop(patience=40) |
0.88 |
54.828 |
48.04 [44.33, 52.11] |
2.181 [2.012, 2.365] |
129.13 [106.80, 157.67] |
87.20 [86.07, 88.32] |
160.11 [116.11, 220.04] |
| 456 |
MB-best |
motionbert_kp_s456_v2 |
813088 |
90 / 131 / early_stop(patience=40) |
5.17 |
55.166 |
48.48 [44.74, 52.58] |
2.201 [2.031, 2.387] |
130.23 [107.78, 158.89] |
87.14 [86.00, 88.25] |
162.24 [116.93, 223.50] |
| 456 |
MB-best + pre |
motionbert_kp_batonpre_trunkp123_s456_v2 |
813088 |
44 / 85 / early_stop(patience=40) |
1.68 |
54.797 |
47.91 [44.22, 51.93] |
2.175 [2.008, 2.357] |
127.59 [105.28, 156.14] |
87.39 [86.28, 88.47] |
158.92 [114.76, 219.74] |
| 456 |
K + pre |
cwm3d_kp_batonpre_trunkp123_s456_v2 |
5771449 |
36 / 77 / early_stop(patience=40) |
0.91 |
54.670 |
48.01 [44.31, 52.12] |
2.179 [2.011, 2.366] |
128.51 [106.16, 157.20] |
87.16 [86.02, 88.27] |
159.02 [115.09, 219.40] |
| 789 |
MB-best |
motionbert_kp_s789_v2 |
813088 |
77 / 118 / early_stop(patience=40) |
6.14 |
55.134 |
48.46 [44.71, 52.57] |
2.200 [2.030, 2.386] |
130.66 [108.19, 159.30] |
87.27 [86.16, 88.36] |
163.04 [118.30, 223.77] |
| 789 |
MB-best + pre |
motionbert_kp_batonpre_trunkp123_s789_v2 |
813088 |
37 / 78 / early_stop(patience=40) |
3.45 |
54.882 |
47.55 [43.88, 51.56] |
2.159 [1.992, 2.341] |
127.71 [105.39, 156.61] |
87.40 [86.30, 88.48] |
159.11 [114.88, 219.30] |
| 789 |
K + pre |
cwm3d_kp_batonpre_trunkp123_s789_v2 |
5771449 |
35 / 76 / early_stop(patience=40) |
1.22 |
54.590 |
48.12 [44.44, 52.21] |
2.184 [2.017, 2.370] |
129.32 [107.17, 157.78] |
87.06 [85.94, 88.18] |
161.39 [117.32, 221.77] |
| 1011 |
MB-best |
motionbert_kp_s1011_v2 |
813088 |
81 / 122 / early_stop(patience=40) |
5.69 |
55.506 |
48.68 [44.92, 52.82] |
2.210 [2.039, 2.398] |
131.49 [108.92, 160.13] |
86.97 [85.81, 88.12] |
166.22 [121.36, 227.01] |
| 1011 |
MB-best + pre |
motionbert_kp_batonpre_trunkp123_s1011_v2 |
813088 |
47 / 88 / early_stop(patience=40) |
3.98 |
54.868 |
47.85 [44.16, 51.87] |
2.172 [2.005, 2.355] |
129.47 [107.17, 157.92] |
87.40 [86.30, 88.49] |
160.75 [116.16, 221.59] |
| 1011 |
K + pre |
cwm3d_kp_batonpre_trunkp123_s1011_v2 |
5771449 |
34 / 75 / early_stop(patience=40) |
1.14 |
54.609 |
48.11 [44.40, 52.21] |
2.184 [2.016, 2.370] |
128.73 [106.48, 157.39] |
87.03 [85.91, 88.16] |
159.72 [115.63, 220.35] |
Δ = MB-best[default]+pre − MB-best[default], per seed (descriptive; n = 5)
| cell |
A s42 |
A s123 |
A s456 |
A s789 |
A s1011 |
B s42 |
B s123 |
B s456 |
B s789 |
B s1011 |
Δ s42 [CI] |
Δ s123 [CI] |
Δ s456 [CI] |
Δ s789 [CI] |
Δ s1011 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
47.74 |
47.53 |
47.91 |
47.55 |
47.85 |
48.28 |
48.38 |
48.48 |
48.46 |
48.68 |
-0.54 [-1.04,-0.05] |
-0.85 [-1.32,-0.37] |
-0.57 [-1.17,-0.00] |
-0.90 [-1.35,-0.46] |
-0.83 [-1.43,-0.26] |
-0.74 ± 0.17 |
BEATS |
| d-n |
2.167 |
2.158 |
2.175 |
2.159 |
2.172 |
2.192 |
2.196 |
2.201 |
2.200 |
2.210 |
-0.024 [-0.047,-0.002] |
-0.039 [-0.060,-0.017] |
-0.026 [-0.053,-0.000] |
-0.041 [-0.061,-0.021] |
-0.038 [-0.065,-0.012] |
-0.034 ± 0.008 |
BEATS |
| HM-MPJPE |
128.03 |
126.11 |
127.59 |
127.71 |
129.47 |
130.34 |
129.16 |
130.23 |
130.66 |
131.49 |
-2.31 [-4.57,-0.07] |
-3.06 [-5.14,-1.09] |
-2.64 [-4.65,-0.62] |
-2.95 [-4.97,-1.00] |
-2.02 [-4.53,+0.34] |
-2.60 ± 0.43 |
BEATS |
| PCK@0.05 |
87.46 |
87.40 |
87.39 |
87.40 |
87.40 |
87.17 |
87.14 |
87.14 |
87.27 |
86.97 |
+0.29 [+0.03,+0.56] |
+0.26 [-0.02,+0.54] |
+0.25 [-0.09,+0.60] |
+0.13 [-0.08,+0.34] |
+0.43 [+0.07,+0.82] |
+0.27 ± 0.11 |
INCONCLUSIVE |
| HM h=5 |
159.20 |
156.47 |
158.92 |
159.11 |
160.75 |
163.95 |
162.59 |
162.24 |
163.04 |
166.22 |
-4.75 [-8.82,-0.70] |
-6.12 [-10.44,-2.09] |
-3.32 [-7.53,+0.84] |
-3.94 [-7.78,-0.05] |
-5.47 [-10.09,-1.44] |
-4.72 ± 1.13 |
BEATS |
Δ = K+pre − K, per seed (descriptive; n = 5)
| cell |
A s42 |
A s123 |
A s456 |
A s789 |
A s1011 |
B s42 |
B s123 |
B s456 |
B s789 |
B s1011 |
Δ s42 [CI] |
Δ s123 [CI] |
Δ s456 [CI] |
Δ s789 [CI] |
Δ s1011 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
47.98 |
48.04 |
48.01 |
48.12 |
48.11 |
48.77 |
48.66 |
48.46 |
48.54 |
48.57 |
-0.79 [-1.29,-0.30] |
-0.62 [-1.17,-0.11] |
-0.45 [-0.89,-0.02] |
-0.43 [-0.88,+0.03] |
-0.46 [-0.89,-0.02] |
-0.55 ± 0.15 |
BEATS |
| d-n |
2.178 |
2.181 |
2.179 |
2.184 |
2.184 |
2.214 |
2.209 |
2.200 |
2.204 |
2.205 |
-0.036 [-0.059,-0.014] |
-0.028 [-0.053,-0.005] |
-0.021 [-0.041,-0.001] |
-0.019 [-0.040,+0.001] |
-0.021 [-0.041,-0.001] |
-0.025 ± 0.007 |
BEATS |
| HM-MPJPE |
128.43 |
129.13 |
128.51 |
129.32 |
128.73 |
132.49 |
131.05 |
131.37 |
131.73 |
131.22 |
-4.06 [-6.57,-1.74] |
-1.92 [-4.14,+0.22] |
-2.86 [-5.09,-0.73] |
-2.41 [-4.43,-0.48] |
-2.49 [-4.79,-0.27] |
-2.75 ± 0.81 |
BEATS |
| PCK@0.05 |
87.21 |
87.20 |
87.16 |
87.06 |
87.03 |
87.00 |
86.99 |
87.13 |
87.00 |
87.00 |
+0.20 [-0.05,+0.45] |
+0.22 [-0.07,+0.52] |
+0.03 [-0.20,+0.26] |
+0.06 [-0.16,+0.28] |
+0.03 [-0.20,+0.26] |
+0.11 ± 0.09 |
NOT_WORSE |
| HM h=5 |
159.35 |
160.11 |
159.02 |
161.39 |
159.72 |
164.19 |
163.36 |
162.55 |
163.04 |
163.70 |
-4.85 [-8.95,-1.00] |
-3.25 [-6.97,+0.55] |
-3.53 [-7.22,+0.08] |
-1.65 [-5.11,+1.98] |
-3.98 [-8.22,+0.07] |
-3.45 ± 1.17 |
INCONCLUSIVE |
Δ = K − MB-best[default], per seed (descriptive; n = 5)
| cell |
A s42 |
A s123 |
A s456 |
A s789 |
A s1011 |
B s42 |
B s123 |
B s456 |
B s789 |
B s1011 |
Δ s42 [CI] |
Δ s123 [CI] |
Δ s456 [CI] |
Δ s789 [CI] |
Δ s1011 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
48.77 |
48.66 |
48.46 |
48.54 |
48.57 |
48.28 |
48.38 |
48.48 |
48.46 |
48.68 |
+0.49 [-0.06,+1.05] |
+0.28 [-0.15,+0.75] |
-0.02 [-0.58,+0.49] |
+0.09 [-0.36,+0.53] |
-0.10 [-0.55,+0.32] |
+0.15 ± 0.24 |
INCONCLUSIVE |
| d-n |
2.214 |
2.209 |
2.200 |
2.204 |
2.205 |
2.192 |
2.196 |
2.201 |
2.200 |
2.210 |
+0.022 [-0.003,+0.048] |
+0.013 [-0.007,+0.034] |
-0.001 [-0.026,+0.022] |
+0.004 [-0.016,+0.024] |
-0.005 [-0.025,+0.015] |
+0.007 ± 0.011 |
NOT_WORSE |
| HM-MPJPE |
132.49 |
131.05 |
131.37 |
131.73 |
131.22 |
130.34 |
129.16 |
130.23 |
130.66 |
131.49 |
+2.15 [-0.84,+5.26] |
+1.89 [+0.02,+3.82] |
+1.14 [-1.06,+3.41] |
+1.07 [-1.32,+3.50] |
-0.27 [-2.17,+1.62] |
+1.19 ± 0.94 |
INCONCLUSIVE |
| PCK@0.05 |
87.00 |
86.99 |
87.13 |
87.00 |
87.00 |
87.17 |
87.14 |
87.14 |
87.27 |
86.97 |
-0.16 [-0.43,+0.11] |
-0.16 [-0.45,+0.11] |
-0.01 [-0.32,+0.33] |
-0.27 [-0.53,-0.02] |
+0.03 [-0.21,+0.29] |
-0.11 ± 0.12 |
INCONCLUSIVE |
| HM h=5 |
164.19 |
163.36 |
162.55 |
163.04 |
163.70 |
163.95 |
162.59 |
162.24 |
163.04 |
166.22 |
+0.24 [-5.20,+5.53] |
+0.77 [-2.56,+4.05] |
+0.30 [-4.32,+4.79] |
-0.00 [-4.68,+4.17] |
-2.51 [-6.02,+0.77] |
-0.24 ± 1.30 |
INCONCLUSIVE |
Δ = K+pre − MB-best[default]+pre, per seed (descriptive; n = 5)
| cell |
A s42 |
A s123 |
A s456 |
A s789 |
A s1011 |
B s42 |
B s123 |
B s456 |
B s789 |
B s1011 |
Δ s42 [CI] |
Δ s123 [CI] |
Δ s456 [CI] |
Δ s789 [CI] |
Δ s1011 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
47.98 |
48.04 |
48.01 |
48.12 |
48.11 |
47.74 |
47.53 |
47.91 |
47.55 |
47.85 |
+0.24 [-0.19,+0.68] |
+0.51 [+0.05,+0.98] |
+0.10 [-0.38,+0.61] |
+0.56 [+0.09,+1.07] |
+0.27 [-0.31,+0.86] |
+0.33 ± 0.19 |
INCONCLUSIVE |
| d-n |
2.178 |
2.181 |
2.179 |
2.184 |
2.184 |
2.167 |
2.158 |
2.175 |
2.159 |
2.172 |
+0.011 [-0.009,+0.031] |
+0.023 [+0.002,+0.045] |
+0.004 [-0.017,+0.028] |
+0.026 [+0.004,+0.048] |
+0.012 [-0.014,+0.039] |
+0.015 ± 0.009 |
INCONCLUSIVE |
| HM-MPJPE |
128.43 |
129.13 |
128.51 |
129.32 |
128.73 |
128.03 |
126.11 |
127.59 |
127.71 |
129.47 |
+0.40 [-2.08,+2.97] |
+3.02 [+0.64,+5.72] |
+0.91 [-1.54,+3.59] |
+1.61 [-0.58,+3.89] |
-0.74 [-3.90,+2.55] |
+1.04 ± 1.40 |
INCONCLUSIVE |
| PCK@0.05 |
87.21 |
87.20 |
87.16 |
87.06 |
87.03 |
87.46 |
87.40 |
87.39 |
87.40 |
87.40 |
-0.25 [-0.49,-0.02] |
-0.20 [-0.47,+0.07] |
-0.23 [-0.53,+0.05] |
-0.34 [-0.63,-0.06] |
-0.37 [-0.71,-0.05] |
-0.28 ± 0.07 |
INCONCLUSIVE |
| HM h=5 |
159.35 |
160.11 |
159.02 |
161.39 |
159.72 |
159.20 |
156.47 |
158.92 |
159.11 |
160.75 |
+0.15 [-4.30,+4.39] |
+3.64 [-0.80,+8.36] |
+0.09 [-5.04,+4.98] |
+2.28 [-2.04,+6.39] |
-1.03 [-6.37,+3.96] |
+1.03 ± 1.89 |
INCONCLUSIVE |
4. Primary endpoints and adoption (test All-MPJPE, 5 seeds, both sides in the same pretraining state; BEATS rule, 4/5)
- K-v2(best) = pending (lowest mean VAL All over finished seeds among S1b-continued survivors; ties 0.10 → fewer components): no continued survivor with finished seeds.
| endpoint |
n seeds |
Δ All per seed [CI] |
mean Δ ± sd |
verdict |
secondary (HM / PCK@0.05 / HM h=5 / d-n verdicts) |
| P1 |
0 |
pending (no K-v2(best) yet) |
— |
PENDING |
— |
| P2 |
0 |
pending (no K-v2(best) yet) |
— |
PENDING |
— |
| A1 |
0 |
pending (no K-v2(best) yet) |
— |
PENDING |
— |
| A2 |
0 |
pending (no K-v2(best) yet) |
— |
PENDING |
— |
4.1 Paper-wording consequence table (pre-registered; the realised row is marked)
| case |
wording consequence |
realised |
| P1 ∧ P2 |
an architecture advantage under the keypoint formulation may be written FOR K-v2 (K's own attribution sentence unchanged) |
pending |
| P1 ∧ ¬P2 |
"advantage without pretraining that the BATON trunk equalises"; attribution sentence kept |
pending |
| ¬P1 ∧ P2 |
"advantage conditional on the pretrained trunk"; attribution sentence kept |
pending |
| neither |
wording unchanged verbatim; K-v2 = ablation family + gen_tables.py NEGATIVES entry; any WORSE reported as is |
pending |
ADOPTION A1: NOT ADOPTED (pending — no K-v2(best) yet; NOT ADOPTED is the conservative default until 5 seeds are final)
ADOPTION A2: NOT ADOPTED (pending — no K-v2(best) yet)
- Consequence-table row: pending (P1 / P2 not final). Any WORSE verdict above is reported as is.
5. GPU minutes of the programme's new runs (from run json; the GPU was shared with a DMS feature-extraction job and other lanes, so minutes are wall-clock upper bounds)
| run |
GPU min |
| cwm3d_kp_batonpre_trunkp123_s1011_v2 |
1.14 |
| cwm3d_kp_batonpre_trunkp123_s789_v2 |
1.22 |
| cwm3d_v2_hazaux_s42_v2 |
4.51 |
| cwm3d_v2_khead_s42_v2 |
4.75 |
| cwm3d_v2_lam03_s42_v2 |
5.51 |
| cwm3d_v2_nohaz_s42_v2 |
5.71 |
| cwm3d_v2_nospatial_s42_v2 |
1.69 |
| cwm3d_v2_plain_s42_v2 |
3.69 |
| cwm3d_v2_q060_s42_v2 |
4.32 |
| cwm3d_v2_q090_s42_v2 |
4.19 |
| cwm3d_v2_s42_v2 |
2.25 |
| motionbert_kp_L4_s42_v2 |
4.71 |
| motionbert_kp_batonpre_trunkp123_s1011_v2 |
3.98 |
| motionbert_kp_batonpre_trunkp123_s789_v2 |
3.45 |
| motionbert_kp_d192_L4_s42_v2 |
5.18 |
| motionbert_kp_d192_s42_v2 |
6.14 |
| motionbert_kp_do0.1_s42_v2 |
3.89 |
| motionbert_kp_lr1e-3_s42_v2 |
2.75 |
| motionbert_kp_lr1e-4_s42_v2 |
3.18 |
| motionbert_kp_wd0.01_s42_v2 |
3.15 |
| motionbert_kp_wd0.1_s42_v2 |
2.83 |
| total (21 runs) |
78.2 |
6. Provenance
- Inputs:
pred_val.npz / pred_test.npz / metrics_val.json / metrics_test.json / run_<name>_<utc>.json under /home/USER/cwm3/aide/results/runs; gate decisions /home/USER/cwm3/aide/results/kv2: s1_decision.json, s1_decision_20260914T234548Z.json, s3_decision.json, s3_decision_20260914T230840Z.json; pre-registration row from /home/USER/cwm3/docs/decisions.md (S0 UTC 2026-09-14T22:51:07Z).
- Comparison runs
cwm3d_kp_s<k>_v2, cwm3d_kp_batonpre_trunkp123_s{42,123,456}_v2, motionbert_kp_s<k>_v2, motionbert_kp_batonpre_trunkp123_s{42,123,456}_v2 are the EXISTING runs (never retrained); every programme run is a new run dir.
- Selection: val All-MPJPE of the EMA weights inside the trainer; S1 / S1b / S3 decisions from pred_val.npz only (
aide/eval/kv2_gate.py); one test evaluation per run by the harness; no test number was used for any decision.
- Launcher
aide/train/run_kv2_program.sh (gate: free VRAM ≥ need + 4 GB, MemAvailable ≥ 28 GB, < 4 trainers, OMP/MKL 4; finished runs skipped); pids aide/results/kv2/pids.txt.
- Pending runs referenced above: 11.
- Trainer code hashes (
code_sha256 of each finished programme run's run json; distinct values = code changed between launches, e.g. the train_kp.py:630 guard fix of 2026-09-14T23:11:53Z, behaviour-neutral for dict-output models — see aide/results/kv2/NOTES.md): e313c0ff89746293 (10 runs: motionbert_kp_L4_s42_v2, motionbert_kp_batonpre_trunkp123_s1011_v2, motionbert_kp_batonpre_trunkp123_s789_v2, motionbert_kp_d192_L4_s42_v2, motionbert_kp_d192_s42_v2, motionbert_kp_do0.1_s42_v2, motionbert_kp_lr1e-3_s42_v2, motionbert_kp_lr1e-4_s42_v2, motionbert_kp_wd0.01_s42_v2, motionbert_kp_wd0.1_s42_v2); 8caba4a8efccd311 (8 runs: cwm3d_v2_hazaux_s42_v2, cwm3d_v2_khead_s42_v2, cwm3d_v2_lam03_s42_v2, cwm3d_v2_nohaz_s42_v2, cwm3d_v2_nospatial_s42_v2, cwm3d_v2_plain_s42_v2, cwm3d_v2_q060_s42_v2, cwm3d_v2_q090_s42_v2); 22306515fd4a42a0 (2 runs: cwm3d_kp_batonpre_trunkp123_s1011_v2, cwm3d_kp_batonpre_trunkp123_s789_v2); 652c4ba49ba937f2 (1 runs: cwm3d_v2_s42_v2).
K-v2 programme — CockpitWM-K improvement, pre-registered 2026-09-14 (protocol C23) — val n = 405 (selection), test n = 609 (one evaluation per run), paired per-clip bootstrap (10000)
Generated 20260914T234723Z by aide/eval/report_kv2_program.py from pred_{val,test}.npz / metrics_{val,test}.json / run json / aide/results/kv2/*.json only (locked protocol: driver first, grid floor, val HM top 40 (threshold 83.33 px, NEVER a ranking statistic), test HM top 60 (threshold 72.05 px), PCK @ 96 px; every run selected on val All-MPJPE of the EMA weights only, test evaluated once by the harness). Δ = paired per-clip difference (negative = better for errors, positive = better for PCK), CI = clip-bootstrap percentile interval. Verdict rule (as arch_attribution.md): BEATS / WORSE = Δ CI excl. 0 on ≥ ceil(2n/3) seeds (5 seeds: 4) with the same sign on every seed; NOT_WORSE = all CIs cover 0 and |Δ| < 0.25; else INCONCLUSIVE. Confirmatory cells need all 5 seeds; otherwise "pending".
0. Pre-registration (verbatim decisions row) and timestamp check
- S0 UTC: 2026-09-14T22:51:07Z; earliest programme run
t_start_utc: 2026-09-14T23:08:22Z (21 programme runs finished of 188 possible names); runs started at or before S0: none.
| 2026-09-14 | K-v2 programme — S0 pre-registration (CockpitWM-K improvement; new model file aide/models/cwm3d_v2.py, --model cwm3d_v2; approved plan <local-notes>/plans/elegant-dazzling-stallman.md Part A; row written at S0 UTC 2026-09-14T22:51:07Z BEFORE any K-v2 / MB-grid run; the same UTC is in aide/results/kv2/S0_UTC.txt and protocol C23). Honesty clause: the K-v2 design (hierarchical part trunk kept; joint tokens + 2 spatial blocks + attention pooling; lean shared read-out; onset-hazard head gating the anchor mixture) was motivated by the TEST diagnostics of arch_attribution.md §3.2 (K vs MotionBERT-kp: right hand +0.71, body +0.28, h=2–4 +0.2–0.3, HM h=3/4 +2.4/+2.0) → therefore ALL K-v2 configuration choices are fixed on VAL only, and MotionBERT-kp (MB) receives the same val selection budget (S3, 10 val points each). Locked protocol unchanged: val-only selection on All-MPJPE of the EMA weights; the official test split is evaluated exactly once per run by the existing harness (append-only results.csv); no flip TTA; ensembles only as separately labelled rows; Stage-A recipe verbatim (--lr 3e-4 --wd 0.05 --warmup 10 --epochs 200 --bs 32 --ema 0.999 --clip 1.0 --patience 40 --aug-window 1 --noise-px 2.0 --joint-drop 0.1 --translate 20 --scale 0.05 --flip-p 0; K loss defaults euc + 0.1 NLL + 0.05 bone + 0.01 jerk; tensor-output models add --loss euc); paired per-clip clip-bootstrap 10,000 resamples seed 0 (aide/eval/metrics.paired_diff); verdict rule of arch_attribution.md: BEATS / WORSE = Δ CI excl. 0 on ≥ ceil(2n/3) seeds (5 seeds: 4/5) with the same sign on every seed; NOT_WORSE = all CIs cover 0 and |Δ| < 0.25; else INCONCLUSIVE. Every decision below is computed mechanically by aide/eval/kv2_gate.py from pred_val.npz + harness.load_split("val") (+ metrics.hm_mask on val) and written as UTC-stamped json in aide/results/kv2/ (s1_decision.json, s1b_<cfg>_decision.json, s3_decision.json, pretrain_check_<cfg>.json); downstream runs may start only after the corresponding json exists. *S1 val screen (seed 42, n = 405)**: 9 configs vs the existing cwm3d_kp_s42_v2 (val All 55.367): cwm3d_v2_s42_v2 (full), cwm3d_v2_nospatial_s42_v2 (--no-v2-spatial), _khead (--no-lean-head, K's per-part read-out), _nohaz (--no-hazard), _hazaux (--haz-mod 0: hazard BCE on, gating off), _plain (--no-physics --w-nll 0), _q060 / _q090 (--haz-quantile 0.60 / 0.90), _lam03 (--haz-weight 0.3); statistic = paired per-clip Δ val All (config − K); SURVIVE iff mean Δ ≤ −0.30 px (≈ 1.4 val seed-sd); at most 2 survivors (ranked by mean Δ; candidates within 0.10 px of each other are ties → the one with FEWER components wins, components = enabled optional modules of {spatial blocks, lean head, hazard BCE, hazard gating, physics+NLL} — full / q060 / q090 / lam03 = 5, nospatial / khead / hazaux / plain = 4, nohaz = 3; remaining ties by name); HM guard (never a selection statistic): KILL iff Δ val HM (40 clips) mean > +3.0 px AND its CI excludes 0. The 40-clip val HM subset is never used for ranking (level CI width ≈ 37 px vs 8.3 for val All; seed sd 0.93 vs 0.21; val HM threshold 83.3 px ≠ test 72.05 px). Test numbers of S1 runs are produced once by the trainer as usual and are RECORDED, never used for any decision. S1b: each survivor at seed 123 vs cwm3d_kp_s123_v2; CONTINUE to seeds 456 / 789 / 1011 iff Δ val(123) < 0 AND the two-seed mean ≤ −0.25 px; otherwise stop and report "val-killed at seed 123". **S2 (5 seeds 42 / 123 / 456 / 789 / 1011), without and with BATON trunk-only init**: K-v2's tokenizer / trunk tensors differ from K's → own pretraining cwm3d_v2_batonpre_pretrain_s123 (pretrain_cwm3d_baton.py --model cwm3d_v2 --seed 123: same corpus sha256 465dac1cf0a8265a, bs 256, 6 epochs, held-out-route selection, --zeroshot 0 = no zero-shot test evaluation) UNLESS kv2_gate.py pretrain-check <cfg> (dry train_kp.load_init_ckpt(parts="trunk") of the survivor model on CPU against aide/results/runs/cwm3d_kp_batonpre_pretrain_s123/ckpt_best.pt) reports n_skipped_shape = 0 AND zero missing TRUNK_PREFIXES tensors, in which case K's checkpoint is reused; fine-tunes cwm3d_v2_<cfg>_batonpre_trunkp123_s<k>_v2 (--init-parts trunk); K's own +pre arm is completed to 5 seeds with cwm3d_kp_batonpre_trunkp123_s{789,1011}_v2 (recipe and init of the existing s42/123/456 runs verbatim). **S3 MB fairness arm**: (i) same-budget val grid at seed 42: motionbert_kp_{d192, L4, lr1e-4, lr1e-3, wd0.01, wd0.1, do0.1, d192_L4, d192_L4_do0.1}_s42_v2 (--model motionbert_kp[_d192][_L4] name-encoded width / depth, --lr, --wd, --dropout flags; --loss euc) + the existing default motionbert_kp_s42_v2 = 10 val points = S1's 10; **MB-best = argmin val All** (ties within 0.10 px → fewer parameters); if MB-best ≠ default it is run at the 5 seeds; (ii) MB-best + BATON trunk-only init at 5 seeds: motionbert_kp_batonpre_trunkp123_s{789,1011}_v2 are run UNCONDITIONALLY (completing the attribution table); if MB-best changes lr / wd / dropout only, the existing motionbert_kp_batonpre_pretrain_s123 checkpoint is reused (motionbert_kp_<cfg>_batonpre_trunkp123_s<k>_v2); if MB-best changes d or L, a new pretraining motionbert_kp_<tag>_batonpre_pretrain_s123 (same corpus / schedule) precedes 5 fine-tunes. **K-v2(best)** = the S2 survivor with the lowest mean VAL All over its finished seeds (ties 0.10 → fewer components); with one survivor it is that one. **PRIMARY (two co-primary endpoints, test All-MPJPE, 5 seeds, both sides in the same pretraining state, beats rule above)**: P1 = K-v2(best) − MB-best (no pretraining); P2 = K-v2(best)+pre − MB-best+pre. **ADOPTION**: A1 = K-v2(best) − cwm3d_kp_s<k>_v2 BEATS on All over 5 seeds → K-v2 becomes the main-table CockpitWM-K row (K's row itself unchanged), otherwise K-v2 is an ablation row; A2 = K-v2(best)+pre − K+pre (cwm3d_kp_batonpre_trunkp123_s<k>_v2) BEATS → the "+ BATON trunk" row. **SECONDARY (descriptive with CIs, never an adoption basis)**: d-n, HM, PCK@0.05, HM h=5, per-part, per-horizon, AUC of the hazard probability haz_P against HM membership. **Paper-wording consequence table** (mirrors the 2026-09-14 attribution row): P1 ∧ P2 BEATS → an architecture advantage under the keypoint formulation may be written FOR K-v2 (K's own attribution sentence unchanged); P1 BEATS ∧ ¬P2 → "advantage without pretraining that the BATON trunk equalises", attribution sentence kept; ¬P1 ∧ P2 → "advantage conditional on the pretrained trunk", attribution sentence kept; neither → wording unchanged verbatim, K-v2 = ablation family + gen_tables.py NEGATIVES entry; any WORSE is reported as is. **Multiplicity**: ≤ 9 (S1) + 18 (S2) + 2 (K+pre seeds) + 9 (S3 i) + 4 (MB-best seeds) + 7 (S3 ii) ≈ 49 new single test evaluations, each required by the protocol, none used for selection; confirmatory claims are exactly P1, P2, A1, A2. **Already published numbers untouched**; all results only in the new table aide/results/tables/kv2_program.md (+ .json) written by aide/eval/report_kv2_program.py (reads only pred_{val,test}.npz, metrics_{val,test}.json, run json; append-only; machine-readable line **ADOPTION A1: ADOPTED\|NOT ADOPTED**). **Execution**: aide/train/run_kv2_program.sh lanes A (S1 → gate s1), B (S3 i → gate s3 → MB-best seeds / +pre), C (motionbert_kp_batonpre_trunkp123_s{789,1011}_v2, cwm3d_kp_batonpre_trunkp123_s{789,1011}_v2), D (pretrain-check → v2 pretraining if needed → S1b → gate s1b → seeds ± pre); every launch passes the project gate (free VRAM ≥ need + 4 GB, MemAvailable ≥ 28 GB, `pgrep -fc "python.(train_cwm3|train_kp|pretrain)"< 4, OMP/MKL 4;setsid nohup bash; a run whose run json exists is never re-run; pids in aide/results/kv2/pids.txt); --no-lock --max-minutes 45, --notes "kv2_program;;cfg="; --refs zvshrink=zv_shrink_val,simlpe=simlpe_kp_s<k or 42>_v2,kp=cwm3d_kp_s_v2,mb=motionbert_kp_s_v2(+seedinit=_s_v2for +pre runs; refs are evaluation-only). Lanes A / D start only afteraide/models/cwm3d_v2.pyexists andpytest -q aide/tests/test_cwm3d_v2.pypasses. Honest expectation (recorded): HM likely improves, All likely still ties MB; every result enters the table regardless. | plan Part A §A.2–A.4;aide/eval/kv2_gate.py, aide/eval/report_kv2_program.py, aide/train/run_kv2_program.sh, aide/tests/test_kv2.py, aide/models/motionbert_kp.py::build(name variants),aide/docs/protocol.mdC23,aide/results/kv2/S0_UTC.txt; motivation aide/results/tables/arch_attribution.md§3.2 | Every K-v2 / MB-grid run json must havet_start_utclater than 2026-09-14T22:51:07Z (checked byreport_kv2_program.py` §0); gate json must predate the downstream launches; "S1 outcome" and final "K-v2 outcome" rows + C23 outcome follow; paper text changes only after the outcome, per the consequence table |
1. S1 val screen (seed 42; statistic = paired per-clip Δ val All vs cwm3d_kp_s42_v2; SURVIVE iff mean Δ ≤ −0.30 px; ≤ 2 survivors, ties 0.10 → fewer components; HM guard KILL iff Δ val HM > +3.0 with CI excl. 0 — the val HM column is SECONDARY)
- Decision source:
aide/results/kv2/s1_decision.json written 2026-09-14T23:45:48Z (complete = True). Rule: paired per-clip Δ val All (config − cwm3d_kp_s42_v2), n = 405, 10000 clip-bootstrap seed 0; SURVIVE iff mean Δ ≤ -0.3 px; ≤ 2 survivors ranked by mean Δ, ties within 0.1 px -> fewer components (full/q060/q090/lam03 = 5, nospatial/khead/hazaux/plain = 4, nohaz = 3); HM guard (never a ranking statistic): KILL iff Δ val HM (40 clips) mean > +3.0 AND CI excl. 0.
| cfg |
run |
flags |
params |
best ep / epochs / stop / GPU min |
val All |
Δ val All [CI] |
Δ val HM [CI] (secondary) |
guard |
S1 status |
test All (recorded, not used) |
| K (baseline) |
cwm3d_kp_s42_v2 |
— |
5771449 |
58 / 99 / early_stop(patience=40) / 0.6 |
55.367 |
0 |
0 |
— |
— |
48.767 |
| full |
cwm3d_v2_s42_v2 |
(full) |
2736282 |
75 / 116 / early_stop(patience=40) / 2.25 |
55.170 |
-0.197 [-0.798,+0.387] |
+0.60 [-2.54,+3.45] |
ok |
KILL |
48.696 |
| nospatial |
cwm3d_v2_nospatial_s42_v2 |
--no-v2-spatial |
2666906 |
87 / 128 / early_stop(patience=40) / 1.69 |
54.788 |
-0.579 [-1.275,+0.058] |
-0.64 [-4.19,+1.99] |
ok |
SURVIVE |
48.267 |
| khead |
cwm3d_v2_khead_s42_v2 |
--no-lean-head |
5091964 |
52 / 93 / early_stop(patience=40) / 4.75 |
55.482 |
+0.115 [-0.442,+0.662] |
+1.06 [-1.16,+3.09] |
ok |
KILL |
48.642 |
| nohaz |
cwm3d_v2_nohaz_s42_v2 |
--no-hazard |
2702199 |
106 / 147 / early_stop(patience=40) / 5.71 |
55.325 |
-0.042 [-0.873,+0.723] |
+2.94 [-0.38,+6.86] |
ok |
KILL |
48.648 |
| hazaux |
cwm3d_v2_hazaux_s42_v2 |
--haz-mod 0 |
2736267 |
87 / 128 / early_stop(patience=40) / 4.51 |
55.140 |
-0.227 [-0.888,+0.393] |
+2.41 [+0.04,+4.95] |
ok |
KILL |
48.592 |
| plain |
cwm3d_v2_plain_s42_v2 |
--no-physics --w-nll 0 |
2736282 |
65 / 106 / early_stop(patience=40) / 3.69 |
55.218 |
-0.149 [-0.722,+0.397] |
+1.73 [-0.72,+4.18] |
ok |
KILL |
48.728 |
| q060 |
cwm3d_v2_q060_s42_v2 |
--haz-quantile 0.60 |
2736282 |
76 / 117 / early_stop(patience=40) / 4.32 |
55.320 |
-0.047 [-0.674,+0.606] |
+3.26 [+0.21,+7.50] |
KILL |
KILL |
48.704 |
| q090 |
cwm3d_v2_q090_s42_v2 |
--haz-quantile 0.90 |
2736282 |
71 / 112 / early_stop(patience=40) / 4.19 |
55.128 |
-0.239 [-0.774,+0.291] |
+1.33 [-1.27,+4.02] |
ok |
KILL |
48.380 |
| lam03 |
cwm3d_v2_lam03_s42_v2 |
--haz-weight 0.3 |
2736282 |
87 / 128 / early_stop(patience=40) / 5.51 |
55.019 |
-0.348 [-1.018,+0.310] |
+1.92 [-0.73,+5.20] |
ok |
SURVIVE |
48.512 |
S1 survivors (decision): nospatial, lam03.
1.1 S1b (survivors at seed 123 vs cwm3d_kp_s123_v2; CONTINUE iff Δ val(123) < 0 and 2-seed mean ≤ −0.25)
| cfg |
seed-123 run |
val All |
Δ val All s42 [CI] |
Δ val All s123 [CI] |
2-seed mean |
decision |
source |
| nospatial |
cwm3d_v2_nospatial_s123_v2 |
pending |
-0.579 [-1.275,+0.058] |
pending |
pending |
PENDING |
live |
| lam03 |
cwm3d_v2_lam03_s123_v2 |
pending |
-0.348 [-1.018,+0.310] |
pending |
pending |
PENDING |
live |
S1b: continuing to 5 seeds: none yet.
2. S2 — 5 seeds (42 / 123 / 456 / 789 / 1011), TEST (one evaluation per run), K-v2 survivors without / with BATON trunk-only init
- Pretraining
cwm3d_v2_batonpre_pretrain_s123: not run / not finished; pretrain-check: —.
(no K-v2 config has reached S2 yet)
3. S3 — MotionBERT-kp fairness arm (same val budget: 9 new grid points + existing default = 10; MB-best = argmin val All, ties 0.10 → fewer params)
- Decision source: no complete
s3_decision.json yet — LIVE recomputation (not a decision). Rule: MB-best = argmin val All over the 10 grid points (9 new + existing default); ties within 0.1 px -> fewer parameters.
| cfg |
run |
model |
flags |
params |
best ep / epochs / stop / GPU min |
val All |
Δ val All vs default [CI] |
test All (recorded, not used) |
S3 |
| default |
motionbert_kp_s42_v2 |
motionbert_kp |
(default) |
813088 |
81 / 122 / early_stop(patience=40) / 4.6 |
55.550 |
+0.000 [+0.000,+0.000] |
48.279 |
|
| d192 |
motionbert_kp_d192_s42_v2 |
motionbert_kp_d192 |
(default) |
1809440 |
64 / 105 / early_stop(patience=40) / 6.14 |
55.559 |
+0.009 [-0.573,+0.575] |
48.595 |
|
| L4 |
motionbert_kp_L4_s42_v2 |
motionbert_kp_L4 |
(default) |
1077664 |
82 / 123 / early_stop(patience=40) / 4.71 |
55.433 |
-0.117 [-0.572,+0.313] |
48.488 |
|
| lr1e-4 |
motionbert_kp_lr1e-4_s42_v2 |
motionbert_kp |
--lr 1e-4 |
813088 |
79 / 120 / early_stop(patience=40) / 3.18 |
55.310 |
-0.240 [-0.773,+0.247] |
48.493 |
|
| lr1e-3 |
motionbert_kp_lr1e-3_s42_v2 |
motionbert_kp |
--lr 1e-3 |
813088 |
60 / 101 / early_stop(patience=40) / 2.75 |
55.281 |
-0.269 [-0.732,+0.131] |
48.394 |
|
| wd0.01 |
motionbert_kp_wd0.01_s42_v2 |
motionbert_kp |
--wd 0.01 |
813088 |
73 / 114 / early_stop(patience=40) / 3.15 |
55.366 |
-0.184 [-0.659,+0.199] |
48.342 |
|
| wd0.1 |
motionbert_kp_wd0.1_s42_v2 |
motionbert_kp |
--wd 0.1 |
813088 |
64 / 105 / early_stop(patience=40) / 2.83 |
55.603 |
+0.053 [-0.324,+0.440] |
48.424 |
|
| do0.1 |
motionbert_kp_do0.1_s42_v2 |
motionbert_kp |
--dropout 0.1 |
813088 |
95 / 136 / early_stop(patience=40) / 3.89 |
55.215 |
-0.335 [-0.888,+0.129] |
48.389 |
MB-best |
| d192_L4 |
motionbert_kp_d192_L4_s42_v2 |
motionbert_kp_d192_L4 |
(default) |
2402912 |
67 / 108 / early_stop(patience=40) / 5.18 |
55.399 |
-0.151 [-0.632,+0.325] |
48.418 |
|
| d192_L4_do0.1 |
motionbert_kp_d192_L4_do0.1_s42_v2 |
motionbert_kp_d192_L4 |
--dropout 0.1 |
— |
pending |
pending |
pending |
— |
PENDING |
MB-best = do0.1 (provisional — S3 incomplete or no decision json; model motionbert_kp, flags ['--dropout', '0.1']; changes d/L: False → existing motionbert_kp_batonpre_pretrain_s123 checkpoint reused).
3.1 MB-best and MB-best + BATON trunk-only init at 5 seeds; K + pre completion (test levels)
| seed |
arm |
run |
params |
best ep / epochs / stop |
GPU min |
val All |
All [CI] |
d-n [CI] |
HM [CI] |
PCK@0.05 [CI] |
HM h=5 [CI] |
| 42 |
MB default |
motionbert_kp_s42_v2 |
813088 |
81 / 122 / early_stop(patience=40) |
4.6 |
55.550 |
48.28 [44.53, 52.37] |
2.192 [2.021, 2.377] |
130.34 [108.20, 158.53] |
87.17 [86.04, 88.28] |
163.95 [119.85, 223.86] |
| 42 |
MB default + pre |
motionbert_kp_batonpre_trunkp123_s42_v2 |
813088 |
37 / 78 / early_stop(patience=40) |
1.94 |
54.528 |
47.74 [44.07, 51.78] |
2.167 [2.000, 2.350] |
128.03 [105.81, 156.47] |
87.46 [86.37, 88.54] |
159.20 [115.09, 219.85] |
| 42 |
MB-best |
motionbert_kp_do0.1_s42_v2 |
813088 |
95 / 136 / early_stop(patience=40) |
3.89 |
55.215 |
48.39 [44.59, 52.49] |
2.197 [2.024, 2.383] |
130.48 [107.84, 158.95] |
87.26 [86.15, 88.36] |
163.78 [118.78, 224.72] |
| 42 |
MB-best + pre |
motionbert_kp_do0.1_batonpre_trunkp123_s42_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 42 |
K + pre |
cwm3d_kp_batonpre_trunkp123_s42_v2 |
5771449 |
31 / 72 / early_stop(patience=40) |
0.48 |
55.003 |
47.98 [44.28, 52.04] |
2.178 [2.010, 2.363] |
128.43 [106.25, 157.11] |
87.21 [86.10, 88.30] |
159.35 [115.56, 219.46] |
| 123 |
MB default |
motionbert_kp_s123_v2 |
813088 |
58 / 99 / early_stop(patience=40) |
4.36 |
55.627 |
48.38 [44.65, 52.46] |
2.196 [2.027, 2.381] |
129.16 [106.70, 157.58] |
87.14 [86.02, 88.27] |
162.59 [118.87, 222.08] |
| 123 |
MB default + pre |
motionbert_kp_batonpre_trunkp123_s123_v2 |
813088 |
38 / 79 / early_stop(patience=40) |
2.16 |
54.720 |
47.53 [43.87, 51.55] |
2.158 [1.992, 2.340] |
126.11 [103.96, 154.69] |
87.40 [86.32, 88.46] |
156.47 [113.00, 215.99] |
| 123 |
MB-best |
motionbert_kp_do0.1_s123_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 123 |
MB-best + pre |
motionbert_kp_do0.1_batonpre_trunkp123_s123_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 123 |
K + pre |
cwm3d_kp_batonpre_trunkp123_s123_v2 |
5771449 |
33 / 74 / early_stop(patience=40) |
0.88 |
54.828 |
48.04 [44.33, 52.11] |
2.181 [2.012, 2.365] |
129.13 [106.80, 157.67] |
87.20 [86.07, 88.32] |
160.11 [116.11, 220.04] |
| 456 |
MB default |
motionbert_kp_s456_v2 |
813088 |
90 / 131 / early_stop(patience=40) |
5.17 |
55.166 |
48.48 [44.74, 52.58] |
2.201 [2.031, 2.387] |
130.23 [107.78, 158.89] |
87.14 [86.00, 88.25] |
162.24 [116.93, 223.50] |
| 456 |
MB default + pre |
motionbert_kp_batonpre_trunkp123_s456_v2 |
813088 |
44 / 85 / early_stop(patience=40) |
1.68 |
54.797 |
47.91 [44.22, 51.93] |
2.175 [2.008, 2.357] |
127.59 [105.28, 156.14] |
87.39 [86.28, 88.47] |
158.92 [114.76, 219.74] |
| 456 |
MB-best |
motionbert_kp_do0.1_s456_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 456 |
MB-best + pre |
motionbert_kp_do0.1_batonpre_trunkp123_s456_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 456 |
K + pre |
cwm3d_kp_batonpre_trunkp123_s456_v2 |
5771449 |
36 / 77 / early_stop(patience=40) |
0.91 |
54.670 |
48.01 [44.31, 52.12] |
2.179 [2.011, 2.366] |
128.51 [106.16, 157.20] |
87.16 [86.02, 88.27] |
159.02 [115.09, 219.40] |
| 789 |
MB default |
motionbert_kp_s789_v2 |
813088 |
77 / 118 / early_stop(patience=40) |
6.14 |
55.134 |
48.46 [44.71, 52.57] |
2.200 [2.030, 2.386] |
130.66 [108.19, 159.30] |
87.27 [86.16, 88.36] |
163.04 [118.30, 223.77] |
| 789 |
MB default + pre |
motionbert_kp_batonpre_trunkp123_s789_v2 |
813088 |
37 / 78 / early_stop(patience=40) |
3.45 |
54.882 |
47.55 [43.88, 51.56] |
2.159 [1.992, 2.341] |
127.71 [105.39, 156.61] |
87.40 [86.30, 88.48] |
159.11 [114.88, 219.30] |
| 789 |
MB-best |
motionbert_kp_do0.1_s789_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 789 |
MB-best + pre |
motionbert_kp_do0.1_batonpre_trunkp123_s789_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 789 |
K + pre |
cwm3d_kp_batonpre_trunkp123_s789_v2 |
5771449 |
35 / 76 / early_stop(patience=40) |
1.22 |
54.590 |
48.12 [44.44, 52.21] |
2.184 [2.017, 2.370] |
129.32 [107.17, 157.78] |
87.06 [85.94, 88.18] |
161.39 [117.32, 221.77] |
| 1011 |
MB default |
motionbert_kp_s1011_v2 |
813088 |
81 / 122 / early_stop(patience=40) |
5.69 |
55.506 |
48.68 [44.92, 52.82] |
2.210 [2.039, 2.398] |
131.49 [108.92, 160.13] |
86.97 [85.81, 88.12] |
166.22 [121.36, 227.01] |
| 1011 |
MB default + pre |
motionbert_kp_batonpre_trunkp123_s1011_v2 |
813088 |
47 / 88 / early_stop(patience=40) |
3.98 |
54.868 |
47.85 [44.16, 51.87] |
2.172 [2.005, 2.355] |
129.47 [107.17, 157.92] |
87.40 [86.30, 88.49] |
160.75 [116.16, 221.59] |
| 1011 |
MB-best |
motionbert_kp_do0.1_s1011_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 1011 |
MB-best + pre |
motionbert_kp_do0.1_batonpre_trunkp123_s1011_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 1011 |
K + pre |
cwm3d_kp_batonpre_trunkp123_s1011_v2 |
5771449 |
34 / 75 / early_stop(patience=40) |
1.14 |
54.609 |
48.11 [44.40, 52.21] |
2.184 [2.016, 2.370] |
128.73 [106.48, 157.39] |
87.03 [85.91, 88.16] |
159.72 [115.63, 220.35] |
Δ = MB-best[do0.1]+pre − MB-best[do0.1], per seed (descriptive; n = 0)
(pending)
Δ = K+pre − K, per seed (descriptive; n = 5)
| cell |
A s42 |
A s123 |
A s456 |
A s789 |
A s1011 |
B s42 |
B s123 |
B s456 |
B s789 |
B s1011 |
Δ s42 [CI] |
Δ s123 [CI] |
Δ s456 [CI] |
Δ s789 [CI] |
Δ s1011 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
47.98 |
48.04 |
48.01 |
48.12 |
48.11 |
48.77 |
48.66 |
48.46 |
48.54 |
48.57 |
-0.79 [-1.29,-0.30] |
-0.62 [-1.17,-0.11] |
-0.45 [-0.89,-0.02] |
-0.43 [-0.88,+0.03] |
-0.46 [-0.89,-0.02] |
-0.55 ± 0.15 |
BEATS |
| d-n |
2.178 |
2.181 |
2.179 |
2.184 |
2.184 |
2.214 |
2.209 |
2.200 |
2.204 |
2.205 |
-0.036 [-0.059,-0.014] |
-0.028 [-0.053,-0.005] |
-0.021 [-0.041,-0.001] |
-0.019 [-0.040,+0.001] |
-0.021 [-0.041,-0.001] |
-0.025 ± 0.007 |
BEATS |
| HM-MPJPE |
128.43 |
129.13 |
128.51 |
129.32 |
128.73 |
132.49 |
131.05 |
131.37 |
131.73 |
131.22 |
-4.06 [-6.57,-1.74] |
-1.92 [-4.14,+0.22] |
-2.86 [-5.09,-0.73] |
-2.41 [-4.43,-0.48] |
-2.49 [-4.79,-0.27] |
-2.75 ± 0.81 |
BEATS |
| PCK@0.05 |
87.21 |
87.20 |
87.16 |
87.06 |
87.03 |
87.00 |
86.99 |
87.13 |
87.00 |
87.00 |
+0.20 [-0.05,+0.45] |
+0.22 [-0.07,+0.52] |
+0.03 [-0.20,+0.26] |
+0.06 [-0.16,+0.28] |
+0.03 [-0.20,+0.26] |
+0.11 ± 0.09 |
NOT_WORSE |
| HM h=5 |
159.35 |
160.11 |
159.02 |
161.39 |
159.72 |
164.19 |
163.36 |
162.55 |
163.04 |
163.70 |
-4.85 [-8.95,-1.00] |
-3.25 [-6.97,+0.55] |
-3.53 [-7.22,+0.08] |
-1.65 [-5.11,+1.98] |
-3.98 [-8.22,+0.07] |
-3.45 ± 1.17 |
INCONCLUSIVE |
Δ = K − MB-best[do0.1], per seed (descriptive; n = 1)
| cell |
A s42 |
B s42 |
Δ s42 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
48.77 |
48.39 |
+0.38 [-0.21,+0.96] |
+0.38 ± nan |
INCONCLUSIVE |
| d-n |
2.214 |
2.197 |
+0.017 [-0.010,+0.044] |
+0.017 ± nan |
NOT_WORSE |
| HM-MPJPE |
132.49 |
130.48 |
+2.01 [-0.71,+4.80] |
+2.01 ± nan |
INCONCLUSIVE |
| PCK@0.05 |
87.00 |
87.26 |
-0.25 [-0.54,+0.01] |
-0.25 ± nan |
INCONCLUSIVE |
| HM h=5 |
164.19 |
163.78 |
+0.41 [-4.94,+5.58] |
+0.41 ± nan |
INCONCLUSIVE |
Δ = K+pre − MB-best[do0.1]+pre, per seed (descriptive; n = 0)
(pending)
4. Primary endpoints and adoption (test All-MPJPE, 5 seeds, both sides in the same pretraining state; BEATS rule, 4/5)
- K-v2(best) = pending (lowest mean VAL All over finished seeds among S1b-continued survivors; ties 0.10 → fewer components): no continued survivor with finished seeds.
| endpoint |
n seeds |
Δ All per seed [CI] |
mean Δ ± sd |
verdict |
secondary (HM / PCK@0.05 / HM h=5 / d-n verdicts) |
| P1 |
0 |
pending (no K-v2(best) yet) |
— |
PENDING |
— |
| P2 |
0 |
pending (no K-v2(best) yet) |
— |
PENDING |
— |
| A1 |
0 |
pending (no K-v2(best) yet) |
— |
PENDING |
— |
| A2 |
0 |
pending (no K-v2(best) yet) |
— |
PENDING |
— |
4.1 Paper-wording consequence table (pre-registered; the realised row is marked)
| case |
wording consequence |
realised |
| P1 ∧ P2 |
an architecture advantage under the keypoint formulation may be written FOR K-v2 (K's own attribution sentence unchanged) |
pending |
| P1 ∧ ¬P2 |
"advantage without pretraining that the BATON trunk equalises"; attribution sentence kept |
pending |
| ¬P1 ∧ P2 |
"advantage conditional on the pretrained trunk"; attribution sentence kept |
pending |
| neither |
wording unchanged verbatim; K-v2 = ablation family + gen_tables.py NEGATIVES entry; any WORSE reported as is |
pending |
ADOPTION A1: NOT ADOPTED (pending — no K-v2(best) yet; NOT ADOPTED is the conservative default until 5 seeds are final)
ADOPTION A2: NOT ADOPTED (pending — no K-v2(best) yet)
- Consequence-table row: pending (P1 / P2 not final). Any WORSE verdict above is reported as is.
5. GPU minutes of the programme's new runs (from run json; the GPU was shared with a DMS feature-extraction job and other lanes, so minutes are wall-clock upper bounds)
| run |
GPU min |
| cwm3d_kp_batonpre_trunkp123_s1011_v2 |
1.14 |
| cwm3d_kp_batonpre_trunkp123_s789_v2 |
1.22 |
| cwm3d_v2_hazaux_s42_v2 |
4.51 |
| cwm3d_v2_khead_s42_v2 |
4.75 |
| cwm3d_v2_lam03_s42_v2 |
5.51 |
| cwm3d_v2_nohaz_s42_v2 |
5.71 |
| cwm3d_v2_nospatial_s42_v2 |
1.69 |
| cwm3d_v2_plain_s42_v2 |
3.69 |
| cwm3d_v2_q060_s42_v2 |
4.32 |
| cwm3d_v2_q090_s42_v2 |
4.19 |
| cwm3d_v2_s42_v2 |
2.25 |
| motionbert_kp_L4_s42_v2 |
4.71 |
| motionbert_kp_batonpre_trunkp123_s1011_v2 |
3.98 |
| motionbert_kp_batonpre_trunkp123_s789_v2 |
3.45 |
| motionbert_kp_d192_L4_s42_v2 |
5.18 |
| motionbert_kp_d192_s42_v2 |
6.14 |
| motionbert_kp_do0.1_s42_v2 |
3.89 |
| motionbert_kp_lr1e-3_s42_v2 |
2.75 |
| motionbert_kp_lr1e-4_s42_v2 |
3.18 |
| motionbert_kp_wd0.01_s42_v2 |
3.15 |
| motionbert_kp_wd0.1_s42_v2 |
2.83 |
| total (21 runs) |
78.2 |
6. Provenance
- Inputs:
pred_val.npz / pred_test.npz / metrics_val.json / metrics_test.json / run_<name>_<utc>.json under /home/USER/cwm3/aide/results/runs; gate decisions /home/USER/cwm3/aide/results/kv2: s1_decision.json, s1_decision_20260914T234548Z.json, s3_decision.json, s3_decision_20260914T230840Z.json; pre-registration row from /home/USER/cwm3/docs/decisions.md (S0 UTC 2026-09-14T22:51:07Z).
- Comparison runs
cwm3d_kp_s<k>_v2, cwm3d_kp_batonpre_trunkp123_s{42,123,456}_v2, motionbert_kp_s<k>_v2, motionbert_kp_batonpre_trunkp123_s{42,123,456}_v2 are the EXISTING runs (never retrained); every programme run is a new run dir.
- Selection: val All-MPJPE of the EMA weights inside the trainer; S1 / S1b / S3 decisions from pred_val.npz only (
aide/eval/kv2_gate.py); one test evaluation per run by the harness; no test number was used for any decision.
- Launcher
aide/train/run_kv2_program.sh (gate: free VRAM ≥ need + 4 GB, MemAvailable ≥ 28 GB, < 4 trainers, OMP/MKL 4; finished runs skipped); pids aide/results/kv2/pids.txt.
- Pending runs referenced above: 3. This block is append-only; earlier blocks are never edited.
- Trainer code hashes (
code_sha256 of each finished programme run's run json; distinct values = code changed between launches, e.g. the train_kp.py:630 guard fix of 2026-09-14T23:11:53Z, behaviour-neutral for dict-output models — see aide/results/kv2/NOTES.md): e313c0ff89746293 (10 runs: motionbert_kp_L4_s42_v2, motionbert_kp_batonpre_trunkp123_s1011_v2, motionbert_kp_batonpre_trunkp123_s789_v2, motionbert_kp_d192_L4_s42_v2, motionbert_kp_d192_s42_v2, motionbert_kp_do0.1_s42_v2, motionbert_kp_lr1e-3_s42_v2, motionbert_kp_lr1e-4_s42_v2, motionbert_kp_wd0.01_s42_v2, motionbert_kp_wd0.1_s42_v2); 8caba4a8efccd311 (8 runs: cwm3d_v2_hazaux_s42_v2, cwm3d_v2_khead_s42_v2, cwm3d_v2_lam03_s42_v2, cwm3d_v2_nohaz_s42_v2, cwm3d_v2_nospatial_s42_v2, cwm3d_v2_plain_s42_v2, cwm3d_v2_q060_s42_v2, cwm3d_v2_q090_s42_v2); 22306515fd4a42a0 (2 runs: cwm3d_kp_batonpre_trunkp123_s1011_v2, cwm3d_kp_batonpre_trunkp123_s789_v2); 652c4ba49ba937f2 (1 runs: cwm3d_v2_s42_v2).
K-v2 programme — CockpitWM-K improvement, pre-registered 2026-09-14 (protocol C23) — val n = 405 (selection), test n = 609 (one evaluation per run), paired per-clip bootstrap (10000)
Generated 20260915T012920Z by aide/eval/report_kv2_program.py from pred_{val,test}.npz / metrics_{val,test}.json / run json / aide/results/kv2/*.json only (locked protocol: driver first, grid floor, val HM top 40 (threshold 83.33 px, NEVER a ranking statistic), test HM top 60 (threshold 72.05 px), PCK @ 96 px; every run selected on val All-MPJPE of the EMA weights only, test evaluated once by the harness). Δ = paired per-clip difference (negative = better for errors, positive = better for PCK), CI = clip-bootstrap percentile interval. Verdict rule (as arch_attribution.md): BEATS / WORSE = Δ CI excl. 0 on ≥ ceil(2n/3) seeds (5 seeds: 4) with the same sign on every seed; NOT_WORSE = all CIs cover 0 and |Δ| < 0.25; else INCONCLUSIVE. Confirmatory cells need all 5 seeds; otherwise "pending".
0. Pre-registration (verbatim decisions row) and timestamp check
- S0 UTC: 2026-09-14T22:51:07Z; earliest programme run
t_start_utc: 2026-09-14T23:08:22Z (42 programme runs finished of 189 possible names); runs started at or before S0: none.
| 2026-09-14 | K-v2 programme — S0 pre-registration (CockpitWM-K improvement; new model file aide/models/cwm3d_v2.py, --model cwm3d_v2; approved plan <local-notes>/plans/elegant-dazzling-stallman.md Part A; row written at S0 UTC 2026-09-14T22:51:07Z BEFORE any K-v2 / MB-grid run; the same UTC is in aide/results/kv2/S0_UTC.txt and protocol C23). Honesty clause: the K-v2 design (hierarchical part trunk kept; joint tokens + 2 spatial blocks + attention pooling; lean shared read-out; onset-hazard head gating the anchor mixture) was motivated by the TEST diagnostics of arch_attribution.md §3.2 (K vs MotionBERT-kp: right hand +0.71, body +0.28, h=2–4 +0.2–0.3, HM h=3/4 +2.4/+2.0) → therefore ALL K-v2 configuration choices are fixed on VAL only, and MotionBERT-kp (MB) receives the same val selection budget (S3, 10 val points each). Locked protocol unchanged: val-only selection on All-MPJPE of the EMA weights; the official test split is evaluated exactly once per run by the existing harness (append-only results.csv); no flip TTA; ensembles only as separately labelled rows; Stage-A recipe verbatim (--lr 3e-4 --wd 0.05 --warmup 10 --epochs 200 --bs 32 --ema 0.999 --clip 1.0 --patience 40 --aug-window 1 --noise-px 2.0 --joint-drop 0.1 --translate 20 --scale 0.05 --flip-p 0; K loss defaults euc + 0.1 NLL + 0.05 bone + 0.01 jerk; tensor-output models add --loss euc); paired per-clip clip-bootstrap 10,000 resamples seed 0 (aide/eval/metrics.paired_diff); verdict rule of arch_attribution.md: BEATS / WORSE = Δ CI excl. 0 on ≥ ceil(2n/3) seeds (5 seeds: 4/5) with the same sign on every seed; NOT_WORSE = all CIs cover 0 and |Δ| < 0.25; else INCONCLUSIVE. Every decision below is computed mechanically by aide/eval/kv2_gate.py from pred_val.npz + harness.load_split("val") (+ metrics.hm_mask on val) and written as UTC-stamped json in aide/results/kv2/ (s1_decision.json, s1b_<cfg>_decision.json, s3_decision.json, pretrain_check_<cfg>.json); downstream runs may start only after the corresponding json exists. *S1 val screen (seed 42, n = 405)**: 9 configs vs the existing cwm3d_kp_s42_v2 (val All 55.367): cwm3d_v2_s42_v2 (full), cwm3d_v2_nospatial_s42_v2 (--no-v2-spatial), _khead (--no-lean-head, K's per-part read-out), _nohaz (--no-hazard), _hazaux (--haz-mod 0: hazard BCE on, gating off), _plain (--no-physics --w-nll 0), _q060 / _q090 (--haz-quantile 0.60 / 0.90), _lam03 (--haz-weight 0.3); statistic = paired per-clip Δ val All (config − K); SURVIVE iff mean Δ ≤ −0.30 px (≈ 1.4 val seed-sd); at most 2 survivors (ranked by mean Δ; candidates within 0.10 px of each other are ties → the one with FEWER components wins, components = enabled optional modules of {spatial blocks, lean head, hazard BCE, hazard gating, physics+NLL} — full / q060 / q090 / lam03 = 5, nospatial / khead / hazaux / plain = 4, nohaz = 3; remaining ties by name); HM guard (never a selection statistic): KILL iff Δ val HM (40 clips) mean > +3.0 px AND its CI excludes 0. The 40-clip val HM subset is never used for ranking (level CI width ≈ 37 px vs 8.3 for val All; seed sd 0.93 vs 0.21; val HM threshold 83.3 px ≠ test 72.05 px). Test numbers of S1 runs are produced once by the trainer as usual and are RECORDED, never used for any decision. S1b: each survivor at seed 123 vs cwm3d_kp_s123_v2; CONTINUE to seeds 456 / 789 / 1011 iff Δ val(123) < 0 AND the two-seed mean ≤ −0.25 px; otherwise stop and report "val-killed at seed 123". **S2 (5 seeds 42 / 123 / 456 / 789 / 1011), without and with BATON trunk-only init**: K-v2's tokenizer / trunk tensors differ from K's → own pretraining cwm3d_v2_batonpre_pretrain_s123 (pretrain_cwm3d_baton.py --model cwm3d_v2 --seed 123: same corpus sha256 465dac1cf0a8265a, bs 256, 6 epochs, held-out-route selection, --zeroshot 0 = no zero-shot test evaluation) UNLESS kv2_gate.py pretrain-check <cfg> (dry train_kp.load_init_ckpt(parts="trunk") of the survivor model on CPU against aide/results/runs/cwm3d_kp_batonpre_pretrain_s123/ckpt_best.pt) reports n_skipped_shape = 0 AND zero missing TRUNK_PREFIXES tensors, in which case K's checkpoint is reused; fine-tunes cwm3d_v2_<cfg>_batonpre_trunkp123_s<k>_v2 (--init-parts trunk); K's own +pre arm is completed to 5 seeds with cwm3d_kp_batonpre_trunkp123_s{789,1011}_v2 (recipe and init of the existing s42/123/456 runs verbatim). **S3 MB fairness arm**: (i) same-budget val grid at seed 42: motionbert_kp_{d192, L4, lr1e-4, lr1e-3, wd0.01, wd0.1, do0.1, d192_L4, d192_L4_do0.1}_s42_v2 (--model motionbert_kp[_d192][_L4] name-encoded width / depth, --lr, --wd, --dropout flags; --loss euc) + the existing default motionbert_kp_s42_v2 = 10 val points = S1's 10; **MB-best = argmin val All** (ties within 0.10 px → fewer parameters); if MB-best ≠ default it is run at the 5 seeds; (ii) MB-best + BATON trunk-only init at 5 seeds: motionbert_kp_batonpre_trunkp123_s{789,1011}_v2 are run UNCONDITIONALLY (completing the attribution table); if MB-best changes lr / wd / dropout only, the existing motionbert_kp_batonpre_pretrain_s123 checkpoint is reused (motionbert_kp_<cfg>_batonpre_trunkp123_s<k>_v2); if MB-best changes d or L, a new pretraining motionbert_kp_<tag>_batonpre_pretrain_s123 (same corpus / schedule) precedes 5 fine-tunes. **K-v2(best)** = the S2 survivor with the lowest mean VAL All over its finished seeds (ties 0.10 → fewer components); with one survivor it is that one. **PRIMARY (two co-primary endpoints, test All-MPJPE, 5 seeds, both sides in the same pretraining state, beats rule above)**: P1 = K-v2(best) − MB-best (no pretraining); P2 = K-v2(best)+pre − MB-best+pre. **ADOPTION**: A1 = K-v2(best) − cwm3d_kp_s<k>_v2 BEATS on All over 5 seeds → K-v2 becomes the main-table CockpitWM-K row (K's row itself unchanged), otherwise K-v2 is an ablation row; A2 = K-v2(best)+pre − K+pre (cwm3d_kp_batonpre_trunkp123_s<k>_v2) BEATS → the "+ BATON trunk" row. **SECONDARY (descriptive with CIs, never an adoption basis)**: d-n, HM, PCK@0.05, HM h=5, per-part, per-horizon, AUC of the hazard probability haz_P against HM membership. **Paper-wording consequence table** (mirrors the 2026-09-14 attribution row): P1 ∧ P2 BEATS → an architecture advantage under the keypoint formulation may be written FOR K-v2 (K's own attribution sentence unchanged); P1 BEATS ∧ ¬P2 → "advantage without pretraining that the BATON trunk equalises", attribution sentence kept; ¬P1 ∧ P2 → "advantage conditional on the pretrained trunk", attribution sentence kept; neither → wording unchanged verbatim, K-v2 = ablation family + gen_tables.py NEGATIVES entry; any WORSE is reported as is. **Multiplicity**: ≤ 9 (S1) + 18 (S2) + 2 (K+pre seeds) + 9 (S3 i) + 4 (MB-best seeds) + 7 (S3 ii) ≈ 49 new single test evaluations, each required by the protocol, none used for selection; confirmatory claims are exactly P1, P2, A1, A2. **Already published numbers untouched**; all results only in the new table aide/results/tables/kv2_program.md (+ .json) written by aide/eval/report_kv2_program.py (reads only pred_{val,test}.npz, metrics_{val,test}.json, run json; append-only; machine-readable line **ADOPTION A1: ADOPTED\|NOT ADOPTED**). **Execution**: aide/train/run_kv2_program.sh lanes A (S1 → gate s1), B (S3 i → gate s3 → MB-best seeds / +pre), C (motionbert_kp_batonpre_trunkp123_s{789,1011}_v2, cwm3d_kp_batonpre_trunkp123_s{789,1011}_v2), D (pretrain-check → v2 pretraining if needed → S1b → gate s1b → seeds ± pre); every launch passes the project gate (free VRAM ≥ need + 4 GB, MemAvailable ≥ 28 GB, `pgrep -fc "python.(train_cwm3|train_kp|pretrain)"< 4, OMP/MKL 4;setsid nohup bash; a run whose run json exists is never re-run; pids in aide/results/kv2/pids.txt); --no-lock --max-minutes 45, --notes "kv2_program;;cfg="; --refs zvshrink=zv_shrink_val,simlpe=simlpe_kp_s<k or 42>_v2,kp=cwm3d_kp_s_v2,mb=motionbert_kp_s_v2(+seedinit=_s_v2for +pre runs; refs are evaluation-only). Lanes A / D start only afteraide/models/cwm3d_v2.pyexists andpytest -q aide/tests/test_cwm3d_v2.pypasses. Honest expectation (recorded): HM likely improves, All likely still ties MB; every result enters the table regardless. | plan Part A §A.2–A.4;aide/eval/kv2_gate.py, aide/eval/report_kv2_program.py, aide/train/run_kv2_program.sh, aide/tests/test_kv2.py, aide/models/motionbert_kp.py::build(name variants),aide/docs/protocol.mdC23,aide/results/kv2/S0_UTC.txt; motivation aide/results/tables/arch_attribution.md§3.2 | Every K-v2 / MB-grid run json must havet_start_utclater than 2026-09-14T22:51:07Z (checked byreport_kv2_program.py` §0); gate json must predate the downstream launches; "S1 outcome" and final "K-v2 outcome" rows + C23 outcome follow; paper text changes only after the outcome, per the consequence table |
1. S1 val screen (seed 42; statistic = paired per-clip Δ val All vs cwm3d_kp_s42_v2; SURVIVE iff mean Δ ≤ −0.30 px; ≤ 2 survivors, ties 0.10 → fewer components; HM guard KILL iff Δ val HM > +3.0 with CI excl. 0 — the val HM column is SECONDARY)
- Decision source:
aide/results/kv2/s1_decision.json written 2026-09-14T23:45:48Z (complete = True). Rule: paired per-clip Δ val All (config − cwm3d_kp_s42_v2), n = 405, 10000 clip-bootstrap seed 0; SURVIVE iff mean Δ ≤ -0.3 px; ≤ 2 survivors ranked by mean Δ, ties within 0.1 px -> fewer components (full/q060/q090/lam03 = 5, nospatial/khead/hazaux/plain = 4, nohaz = 3); HM guard (never a ranking statistic): KILL iff Δ val HM (40 clips) mean > +3.0 AND CI excl. 0.
| cfg |
run |
flags |
params |
best ep / epochs / stop / GPU min |
val All |
Δ val All [CI] |
Δ val HM [CI] (secondary) |
guard |
S1 status |
test All (recorded, not used) |
| K (baseline) |
cwm3d_kp_s42_v2 |
— |
5771449 |
58 / 99 / early_stop(patience=40) / 0.6 |
55.367 |
0 |
0 |
— |
— |
48.767 |
| full |
cwm3d_v2_s42_v2 |
(full) |
2736282 |
75 / 116 / early_stop(patience=40) / 2.25 |
55.170 |
-0.197 [-0.798,+0.387] |
+0.60 [-2.54,+3.45] |
ok |
KILL |
48.696 |
| nospatial |
cwm3d_v2_nospatial_s42_v2 |
--no-v2-spatial |
2666906 |
87 / 128 / early_stop(patience=40) / 1.69 |
54.788 |
-0.579 [-1.275,+0.058] |
-0.64 [-4.19,+1.99] |
ok |
SURVIVE |
48.267 |
| khead |
cwm3d_v2_khead_s42_v2 |
--no-lean-head |
5091964 |
52 / 93 / early_stop(patience=40) / 4.75 |
55.482 |
+0.115 [-0.442,+0.662] |
+1.06 [-1.16,+3.09] |
ok |
KILL |
48.642 |
| nohaz |
cwm3d_v2_nohaz_s42_v2 |
--no-hazard |
2702199 |
106 / 147 / early_stop(patience=40) / 5.71 |
55.325 |
-0.042 [-0.873,+0.723] |
+2.94 [-0.38,+6.86] |
ok |
KILL |
48.648 |
| hazaux |
cwm3d_v2_hazaux_s42_v2 |
--haz-mod 0 |
2736267 |
87 / 128 / early_stop(patience=40) / 4.51 |
55.140 |
-0.227 [-0.888,+0.393] |
+2.41 [+0.04,+4.95] |
ok |
KILL |
48.592 |
| plain |
cwm3d_v2_plain_s42_v2 |
--no-physics --w-nll 0 |
2736282 |
65 / 106 / early_stop(patience=40) / 3.69 |
55.218 |
-0.149 [-0.722,+0.397] |
+1.73 [-0.72,+4.18] |
ok |
KILL |
48.728 |
| q060 |
cwm3d_v2_q060_s42_v2 |
--haz-quantile 0.60 |
2736282 |
76 / 117 / early_stop(patience=40) / 4.32 |
55.320 |
-0.047 [-0.674,+0.606] |
+3.26 [+0.21,+7.50] |
KILL |
KILL |
48.704 |
| q090 |
cwm3d_v2_q090_s42_v2 |
--haz-quantile 0.90 |
2736282 |
71 / 112 / early_stop(patience=40) / 4.19 |
55.128 |
-0.239 [-0.774,+0.291] |
+1.33 [-1.27,+4.02] |
ok |
KILL |
48.380 |
| lam03 |
cwm3d_v2_lam03_s42_v2 |
--haz-weight 0.3 |
2736282 |
87 / 128 / early_stop(patience=40) / 5.51 |
55.019 |
-0.348 [-1.018,+0.310] |
+1.92 [-0.73,+5.20] |
ok |
SURVIVE |
48.512 |
S1 survivors (decision): nospatial, lam03.
1.1 S1b (survivors at seed 123 vs cwm3d_kp_s123_v2; CONTINUE iff Δ val(123) < 0 and 2-seed mean ≤ −0.25)
| cfg |
seed-123 run |
val All |
Δ val All s42 [CI] |
Δ val All s123 [CI] |
2-seed mean |
decision |
source |
| nospatial |
cwm3d_v2_nospatial_s123_v2 |
54.801 |
-0.579 [-1.275,+0.058] |
-0.566 [-1.121,-0.038] |
-0.573 |
CONTINUE |
decision json 2026-09-15T00:00:38Z |
| lam03 |
cwm3d_v2_lam03_s123_v2 |
55.297 |
-0.348 [-1.018,+0.310] |
-0.070 [-0.678,+0.534] |
-0.209 |
KILLED |
decision json 2026-09-15T00:00:39Z |
S1b: continuing to 5 seeds: nospatial.
2. S2 — 5 seeds (42 / 123 / 456 / 789 / 1011), TEST (one evaluation per run), K-v2 survivors without / with BATON trunk-only init
- Pretraining
cwm3d_v2_batonpre_pretrain_s123 (architecture-matched trunk for lam03): not run / not finished.
- Pretraining
cwm3d_v2_nospatial_batonpre_pretrain_s123 (architecture-matched trunk for nospatial): finished (6 epochs, best epoch 5, held-out All 27.18, 10.72 GPU-min, corpus sha256 465dac1cf0a8265a, command aide/train/pretrain_cwm3d_baton.py --model cwm3d_v2 --no-v2-spatial --seed 123 --name cwm3d_v2_nospatial_batonpre_pretrain_s123 --zeroshot 0).
- pretrain-check
nospatial: vs K checkpoint → NEED_V2_PRETRAIN (loaded 47 / 54 trunk tensors, shape-skipped 0, missing 7); vs own pretraining → trunk_complete = True (loaded 54 / 54, shape-skipped 0, missing 0).
- pretrain-check
lam03: vs K checkpoint → no json (loaded — / — trunk tensors, shape-skipped —, missing —); vs own pretraining → no json.
2.1 nospatial (--no-v2-spatial)
| seed |
arm |
run |
params |
best ep / epochs / stop |
GPU min |
val All |
All [CI] |
d-n [CI] |
HM [CI] |
PCK@0.05 [CI] |
HM h=5 [CI] |
| 42 |
K-v2 |
cwm3d_v2_nospatial_s42_v2 |
2666906 |
87 / 128 / early_stop(patience=40) |
1.69 |
54.788 |
48.27 [44.54, 52.40] |
2.191 [2.022, 2.379] |
128.74 [106.37, 157.12] |
87.10 [85.94, 88.24] |
159.70 [116.06, 218.87] |
| 42 |
K-v2 + pre |
cwm3d_v2_nospatial_batonpre_trunkp123_s42_v2 |
2666906 |
44 / 85 / early_stop(patience=40) |
1.21 |
54.689 |
47.98 [44.26, 52.04] |
2.178 [2.009, 2.363] |
128.58 [106.36, 156.90] |
87.20 [86.08, 88.32] |
159.20 [115.08, 219.21] |
| 123 |
K-v2 |
cwm3d_v2_nospatial_s123_v2 |
2666906 |
84 / 125 / early_stop(patience=40) |
3.71 |
54.801 |
48.53 [44.75, 52.63] |
2.203 [2.031, 2.389] |
129.11 [106.68, 157.62] |
86.95 [85.78, 88.12] |
161.40 [117.80, 220.27] |
| 123 |
K-v2 + pre |
cwm3d_v2_nospatial_batonpre_trunkp123_s123_v2 |
2666906 |
42 / 83 / early_stop(patience=40) |
1.09 |
54.321 |
47.95 [44.22, 52.05] |
2.177 [2.007, 2.363] |
129.25 [106.85, 157.68] |
87.29 [86.15, 88.40] |
160.85 [116.52, 221.12] |
| 456 |
K-v2 |
cwm3d_v2_nospatial_s456_v2 |
2666906 |
91 / 132 / early_stop(patience=40) |
3.67 |
55.150 |
48.53 [44.76, 52.59] |
2.203 [2.032, 2.388] |
129.92 [107.92, 158.13] |
87.01 [85.86, 88.16] |
160.73 [117.34, 219.34] |
| 456 |
K-v2 + pre |
cwm3d_v2_nospatial_batonpre_trunkp123_s456_v2 |
2666906 |
44 / 85 / early_stop(patience=40) |
1.12 |
54.352 |
48.04 [44.34, 52.10] |
2.181 [2.013, 2.365] |
128.45 [106.26, 156.89] |
87.09 [85.95, 88.21] |
159.24 [115.23, 219.46] |
| 789 |
K-v2 |
cwm3d_v2_nospatial_s789_v2 |
2666906 |
80 / 121 / early_stop(patience=40) |
1.65 |
55.048 |
48.44 [44.65, 52.56] |
2.199 [2.027, 2.386] |
130.96 [108.49, 159.39] |
86.96 [85.82, 88.10] |
161.72 [117.96, 220.66] |
| 789 |
K-v2 + pre |
cwm3d_v2_nospatial_batonpre_trunkp123_s789_v2 |
2666906 |
40 / 81 / early_stop(patience=40) |
1.08 |
54.710 |
47.81 [44.12, 51.85] |
2.170 [2.003, 2.354] |
128.48 [106.16, 157.11] |
87.25 [86.14, 88.36] |
159.53 [115.32, 219.61] |
| 1011 |
K-v2 |
cwm3d_v2_nospatial_s1011_v2 |
2666906 |
81 / 122 / early_stop(patience=40) |
1.58 |
54.744 |
48.30 [44.53, 52.43] |
2.192 [2.021, 2.380] |
131.03 [108.45, 159.36] |
87.09 [85.94, 88.23] |
164.10 [120.54, 222.22] |
| 1011 |
K-v2 + pre |
cwm3d_v2_nospatial_batonpre_trunkp123_s1011_v2 |
2666906 |
44 / 85 / early_stop(patience=40) |
1.12 |
54.559 |
48.10 [44.37, 52.18] |
2.183 [2.014, 2.369] |
129.87 [107.84, 158.26] |
87.28 [86.16, 88.39] |
161.58 [118.12, 220.56] |
Δ = K-v2 − K, per seed (n = 5 seed pairs finished; n_req = ceil(2n/3) = 4)
| cell |
A s42 |
A s123 |
A s456 |
A s789 |
A s1011 |
B s42 |
B s123 |
B s456 |
B s789 |
B s1011 |
Δ s42 [CI] |
Δ s123 [CI] |
Δ s456 [CI] |
Δ s789 [CI] |
Δ s1011 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
48.27 |
48.53 |
48.53 |
48.44 |
48.30 |
48.77 |
48.66 |
48.46 |
48.54 |
48.57 |
-0.50 [-1.00,-0.01] |
-0.13 [-0.69,+0.33] |
+0.07 [-0.42,+0.59] |
-0.10 [-0.44,+0.24] |
-0.28 [-0.60,+0.05] |
-0.19 ± 0.21 |
INCONCLUSIVE |
| d-n |
2.191 |
2.203 |
2.203 |
2.199 |
2.192 |
2.214 |
2.209 |
2.200 |
2.204 |
2.205 |
-0.023 [-0.045,-0.001] |
-0.006 [-0.031,+0.015] |
+0.003 [-0.019,+0.027] |
-0.005 [-0.020,+0.011] |
-0.013 [-0.027,+0.002] |
-0.009 ± 0.010 |
INCONCLUSIVE |
| HM-MPJPE |
128.74 |
129.11 |
129.92 |
130.96 |
131.03 |
132.49 |
131.05 |
131.37 |
131.73 |
131.22 |
-3.75 [-5.86,-1.76] |
-1.94 [-3.80,-0.25] |
-1.45 [-3.86,+0.83] |
-0.77 [-2.46,+1.04] |
-0.19 [-2.01,+1.73] |
-1.62 ± 1.36 |
INCONCLUSIVE |
| PCK@0.05 |
87.10 |
86.95 |
87.01 |
86.96 |
87.09 |
87.00 |
86.99 |
87.13 |
87.00 |
87.00 |
+0.10 [-0.17,+0.36] |
-0.04 [-0.34,+0.28] |
-0.11 [-0.42,+0.16] |
-0.04 [-0.26,+0.17] |
+0.09 [-0.10,+0.29] |
-0.00 ± 0.09 |
NOT_WORSE |
| HM h=5 |
159.70 |
161.40 |
160.73 |
161.72 |
164.10 |
164.19 |
163.36 |
162.55 |
163.04 |
163.70 |
-4.49 [-9.02,-0.43] |
-1.96 [-4.68,+0.59] |
-1.82 [-6.12,+2.26] |
-1.33 [-4.21,+1.76] |
+0.39 [-2.73,+3.48] |
-1.84 ± 1.75 |
INCONCLUSIVE |
| All body |
70.64 |
70.83 |
70.85 |
70.96 |
70.73 |
71.39 |
71.29 |
71.02 |
71.05 |
71.03 |
-0.75 [-1.33,-0.20] |
-0.45 [-1.15,+0.14] |
-0.18 [-0.74,+0.48] |
-0.09 [-0.57,+0.39] |
-0.30 [-0.78,+0.17] |
-0.35 ± 0.26 |
INCONCLUSIVE |
| All face |
25.58 |
25.80 |
25.58 |
25.75 |
25.76 |
25.64 |
25.86 |
25.57 |
25.62 |
25.75 |
-0.07 [-0.60,+0.46] |
-0.06 [-0.60,+0.47] |
+0.01 [-0.55,+0.59] |
+0.14 [-0.28,+0.58] |
+0.01 [-0.37,+0.39] |
+0.01 ± 0.08 |
NOT_WORSE |
| All lhand |
64.79 |
65.01 |
65.17 |
64.71 |
64.37 |
65.67 |
65.11 |
65.03 |
65.05 |
65.30 |
-0.89 [-2.10,+0.31] |
-0.10 [-1.17,+0.93] |
+0.14 [-0.88,+1.21] |
-0.34 [-1.20,+0.51] |
-0.93 [-1.84,-0.06] |
-0.42 ± 0.47 |
INCONCLUSIVE |
| All rhand |
77.51 |
78.01 |
78.57 |
77.77 |
77.40 |
78.72 |
78.03 |
78.09 |
78.41 |
77.94 |
-1.22 [-2.47,+0.02] |
-0.02 [-1.20,+1.11] |
+0.48 [-0.86,+1.86] |
-0.65 [-1.58,+0.31] |
-0.54 [-1.30,+0.23] |
-0.39 ± 0.65 |
INCONCLUSIVE |
| All h=1 |
34.88 |
34.80 |
34.89 |
34.86 |
34.77 |
35.08 |
35.03 |
35.07 |
35.00 |
35.14 |
-0.20 [-0.82,+0.48] |
-0.24 [-0.81,+0.21] |
-0.18 [-0.57,+0.20] |
-0.15 [-0.54,+0.27] |
-0.37 [-0.88,+0.04] |
-0.23 ± 0.09 |
INCONCLUSIVE |
| All h=2 |
44.18 |
44.31 |
44.52 |
44.47 |
44.25 |
44.84 |
44.66 |
44.50 |
44.55 |
44.60 |
-0.66 [-1.44,+0.01] |
-0.35 [-0.97,+0.16] |
+0.02 [-0.48,+0.53] |
-0.09 [-0.52,+0.34] |
-0.36 [-0.71,-0.00] |
-0.29 ± 0.27 |
INCONCLUSIVE |
| All h=3 |
51.58 |
51.88 |
51.93 |
51.77 |
51.55 |
52.24 |
52.04 |
51.87 |
51.93 |
51.85 |
-0.66 [-1.39,+0.01] |
-0.16 [-0.81,+0.44] |
+0.06 [-0.62,+0.79] |
-0.16 [-0.65,+0.32] |
-0.30 [-0.76,+0.17] |
-0.24 ± 0.27 |
INCONCLUSIVE |
| All h=4 |
53.26 |
53.69 |
53.54 |
53.55 |
53.40 |
53.91 |
53.76 |
53.50 |
53.68 |
53.60 |
-0.65 [-1.36,+0.02] |
-0.07 [-0.73,+0.55] |
+0.04 [-0.70,+0.83] |
-0.13 [-0.64,+0.39] |
-0.20 [-0.65,+0.26] |
-0.20 ± 0.26 |
INCONCLUSIVE |
| All h=5 |
57.43 |
57.96 |
57.79 |
57.58 |
57.50 |
57.77 |
57.81 |
57.38 |
57.55 |
57.67 |
-0.34 [-1.05,+0.33] |
+0.15 [-0.46,+0.75] |
+0.41 [-0.37,+1.28] |
+0.02 [-0.51,+0.52] |
-0.17 [-0.62,+0.29] |
+0.02 ± 0.29 |
INCONCLUSIVE |
Δ = K-v2+pre − K+pre, per seed (n = 5 seed pairs finished; n_req = ceil(2n/3) = 4)
| cell |
A s42 |
A s123 |
A s456 |
A s789 |
A s1011 |
B s42 |
B s123 |
B s456 |
B s789 |
B s1011 |
Δ s42 [CI] |
Δ s123 [CI] |
Δ s456 [CI] |
Δ s789 [CI] |
Δ s1011 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
47.98 |
47.95 |
48.04 |
47.81 |
48.10 |
47.98 |
48.04 |
48.01 |
48.12 |
48.11 |
+0.00 [-0.37,+0.36] |
-0.09 [-0.47,+0.27] |
+0.03 [-0.40,+0.45] |
-0.31 [-0.72,+0.07] |
-0.02 [-0.37,+0.32] |
-0.08 ± 0.14 |
INCONCLUSIVE |
| d-n |
2.178 |
2.177 |
2.181 |
2.170 |
2.183 |
2.178 |
2.181 |
2.179 |
2.184 |
2.184 |
+0.000 [-0.017,+0.016] |
-0.004 [-0.022,+0.012] |
+0.001 [-0.018,+0.020] |
-0.014 [-0.033,+0.003] |
-0.001 [-0.017,+0.015] |
-0.003 ± 0.006 |
NOT_WORSE |
| HM-MPJPE |
128.58 |
129.25 |
128.45 |
128.48 |
129.87 |
128.43 |
129.13 |
128.51 |
129.32 |
128.73 |
+0.15 [-2.22,+2.25] |
+0.12 [-1.63,+1.72] |
-0.06 [-2.48,+1.92] |
-0.84 [-2.64,+0.85] |
+1.14 [-0.67,+2.94] |
+0.10 ± 0.71 |
INCONCLUSIVE |
| PCK@0.05 |
87.20 |
87.29 |
87.09 |
87.25 |
87.28 |
87.21 |
87.20 |
87.16 |
87.06 |
87.03 |
-0.00 [-0.21,+0.22] |
+0.08 [-0.10,+0.28] |
-0.07 [-0.34,+0.23] |
+0.19 [-0.06,+0.46] |
+0.24 [+0.02,+0.50] |
+0.09 ± 0.13 |
INCONCLUSIVE |
| HM h=5 |
159.20 |
160.85 |
159.24 |
159.53 |
161.58 |
159.35 |
160.11 |
159.02 |
161.39 |
159.72 |
-0.15 [-4.17,+3.78] |
+0.74 [-2.12,+3.40] |
+0.22 [-4.10,+4.02] |
-1.86 [-5.12,+1.28] |
+1.86 [-2.13,+5.96] |
+0.16 ± 1.36 |
INCONCLUSIVE |
| All body |
70.48 |
70.17 |
70.36 |
70.27 |
70.26 |
70.64 |
70.68 |
70.78 |
70.48 |
70.37 |
-0.16 [-0.58,+0.27] |
-0.50 [-1.04,-0.02] |
-0.41 [-0.89,+0.05] |
-0.21 [-0.68,+0.24] |
-0.11 [-0.49,+0.28] |
-0.28 ± 0.17 |
INCONCLUSIVE |
| All face |
25.45 |
25.64 |
25.52 |
25.28 |
25.66 |
25.35 |
25.40 |
25.28 |
25.47 |
25.37 |
+0.09 [-0.33,+0.49] |
+0.24 [-0.13,+0.60] |
+0.24 [-0.29,+0.76] |
-0.19 [-0.67,+0.23] |
+0.28 [-0.11,+0.65] |
+0.13 ± 0.20 |
INCONCLUSIVE |
| All lhand |
63.92 |
63.86 |
63.41 |
63.85 |
63.42 |
64.02 |
63.91 |
64.19 |
64.16 |
64.39 |
-0.10 [-1.13,+0.95] |
-0.04 [-1.02,+0.97] |
-0.78 [-1.81,+0.23] |
-0.30 [-1.39,+0.77] |
-0.98 [-2.02,+0.04] |
-0.44 ± 0.42 |
INCONCLUSIVE |
| All rhand |
77.14 |
76.77 |
77.93 |
76.92 |
78.00 |
77.14 |
77.47 |
77.21 |
77.73 |
77.91 |
+0.00 [-0.98,+0.98] |
-0.70 [-1.73,+0.30] |
+0.72 [-0.29,+1.74] |
-0.81 [-1.85,+0.20] |
+0.09 [-0.96,+1.14] |
-0.14 ± 0.63 |
INCONCLUSIVE |
| All h=1 |
34.69 |
34.59 |
34.78 |
34.56 |
34.79 |
34.53 |
34.34 |
34.69 |
34.56 |
34.62 |
+0.15 [-0.23,+0.53] |
+0.25 [-0.14,+0.66] |
+0.10 [-0.31,+0.48] |
+0.00 [-0.36,+0.34] |
+0.17 [-0.18,+0.52] |
+0.13 ± 0.09 |
NOT_WORSE |
| All h=2 |
43.96 |
44.01 |
44.06 |
43.76 |
44.04 |
44.11 |
44.08 |
44.13 |
44.10 |
44.20 |
-0.15 [-0.62,+0.33] |
-0.07 [-0.58,+0.40] |
-0.06 [-0.59,+0.44] |
-0.35 [-0.79,+0.09] |
-0.15 [-0.60,+0.28] |
-0.16 ± 0.12 |
INCONCLUSIVE |
| All h=3 |
51.22 |
51.14 |
51.28 |
51.05 |
51.43 |
51.31 |
51.50 |
51.36 |
51.42 |
51.54 |
-0.09 [-0.55,+0.37] |
-0.36 [-0.88,+0.11] |
-0.08 [-0.64,+0.46] |
-0.36 [-0.87,+0.12] |
-0.10 [-0.59,+0.37] |
-0.20 ± 0.15 |
INCONCLUSIVE |
| All h=4 |
53.06 |
52.95 |
53.05 |
52.88 |
53.12 |
53.02 |
53.19 |
53.05 |
53.22 |
53.22 |
+0.03 [-0.50,+0.56] |
-0.24 [-0.81,+0.28] |
-0.01 [-0.62,+0.60] |
-0.34 [-0.94,+0.23] |
-0.09 [-0.66,+0.47] |
-0.13 ± 0.16 |
INCONCLUSIVE |
| All h=5 |
56.97 |
57.07 |
57.01 |
56.79 |
57.09 |
56.92 |
57.10 |
56.81 |
57.28 |
56.99 |
+0.05 [-0.51,+0.62] |
-0.03 [-0.60,+0.50] |
+0.20 [-0.44,+0.83] |
-0.49 [-1.15,+0.12] |
+0.10 [-0.50,+0.70] |
-0.04 ± 0.27 |
INCONCLUSIVE |
Δ = K-v2 − MB-best[do0.1], per seed (n = 5 seed pairs finished; n_req = ceil(2n/3) = 4)
| cell |
A s42 |
A s123 |
A s456 |
A s789 |
A s1011 |
B s42 |
B s123 |
B s456 |
B s789 |
B s1011 |
Δ s42 [CI] |
Δ s123 [CI] |
Δ s456 [CI] |
Δ s789 [CI] |
Δ s1011 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
48.27 |
48.53 |
48.53 |
48.44 |
48.30 |
48.39 |
48.41 |
48.21 |
48.37 |
48.92 |
-0.12 [-0.75,+0.49] |
+0.12 [-0.31,+0.54] |
+0.32 [-0.26,+0.92] |
+0.07 [-0.43,+0.59] |
-0.62 [-1.24,-0.03] |
-0.05 ± 0.36 |
INCONCLUSIVE |
| d-n |
2.191 |
2.203 |
2.203 |
2.199 |
2.192 |
2.197 |
2.197 |
2.188 |
2.196 |
2.221 |
-0.006 [-0.034,+0.022] |
+0.005 [-0.014,+0.025] |
+0.015 [-0.012,+0.042] |
+0.003 [-0.020,+0.027] |
-0.028 [-0.056,-0.001] |
-0.002 ± 0.016 |
INCONCLUSIVE |
| HM-MPJPE |
128.74 |
129.11 |
129.92 |
130.96 |
131.03 |
130.48 |
129.51 |
128.99 |
130.78 |
130.61 |
-1.74 [-3.99,+0.38] |
-0.40 [-2.37,+1.65] |
+0.93 [-1.89,+3.74] |
+0.18 [-2.95,+3.59] |
+0.43 [-2.06,+2.97] |
-0.12 ± 1.02 |
INCONCLUSIVE |
| PCK@0.05 |
87.10 |
86.95 |
87.01 |
86.96 |
87.09 |
87.26 |
87.11 |
87.30 |
87.20 |
86.90 |
-0.16 [-0.51,+0.16] |
-0.16 [-0.47,+0.12] |
-0.28 [-0.62,+0.04] |
-0.24 [-0.56,+0.06] |
+0.20 [-0.08,+0.48] |
-0.13 ± 0.19 |
INCONCLUSIVE |
| HM h=5 |
159.70 |
161.40 |
160.73 |
161.72 |
164.10 |
163.78 |
162.75 |
162.24 |
164.33 |
163.65 |
-4.08 [-8.62,+0.25] |
-1.34 [-5.07,+2.43] |
-1.51 [-6.34,+3.12] |
-2.62 [-7.44,+2.14] |
+0.44 [-4.18,+4.84] |
-1.82 ± 1.67 |
INCONCLUSIVE |
| All body |
70.64 |
70.83 |
70.85 |
70.96 |
70.73 |
70.88 |
70.94 |
70.85 |
70.68 |
71.10 |
-0.23 [-0.86,+0.38] |
-0.10 [-0.79,+0.53] |
+0.00 [-0.73,+0.75] |
+0.28 [-0.33,+0.89] |
-0.37 [-1.03,+0.28] |
-0.09 ± 0.25 |
INCONCLUSIVE |
| All face |
25.58 |
25.80 |
25.58 |
25.75 |
25.76 |
25.39 |
25.70 |
25.53 |
25.45 |
25.92 |
+0.19 [-0.43,+0.87] |
+0.10 [-0.33,+0.57] |
+0.06 [-0.50,+0.66] |
+0.31 [-0.32,+1.00] |
-0.16 [-0.77,+0.47] |
+0.10 ± 0.17 |
INCONCLUSIVE |
| All lhand |
64.79 |
65.01 |
65.17 |
64.71 |
64.37 |
65.53 |
65.27 |
65.04 |
65.36 |
66.48 |
-0.74 [-2.22,+0.64] |
-0.26 [-1.39,+0.82] |
+0.13 [-1.25,+1.42] |
-0.65 [-2.00,+0.58] |
-2.11 [-4.05,-0.42] |
-0.73 ± 0.85 |
INCONCLUSIVE |
| All rhand |
77.51 |
78.01 |
78.57 |
77.77 |
77.40 |
77.89 |
77.19 |
76.79 |
77.97 |
78.36 |
-0.38 [-1.90,+1.02] |
+0.81 [-0.36,+1.98] |
+1.78 [+0.27,+3.40] |
-0.20 [-1.44,+1.03] |
-0.96 [-2.40,+0.47] |
+0.21 ± 1.09 |
INCONCLUSIVE |
| All h=1 |
34.88 |
34.80 |
34.89 |
34.86 |
34.77 |
34.97 |
34.99 |
34.94 |
35.20 |
35.23 |
-0.09 [-0.55,+0.36] |
-0.19 [-0.63,+0.19] |
-0.05 [-0.55,+0.42] |
-0.35 [-0.89,+0.19] |
-0.45 [-1.04,+0.10] |
-0.23 ± 0.17 |
INCONCLUSIVE |
| All h=2 |
44.18 |
44.31 |
44.52 |
44.47 |
44.25 |
44.28 |
44.37 |
44.14 |
44.29 |
44.95 |
-0.10 [-0.77,+0.56] |
-0.07 [-0.57,+0.42] |
+0.38 [-0.18,+0.94] |
+0.17 [-0.41,+0.80] |
-0.71 [-1.40,-0.04] |
-0.06 ± 0.41 |
INCONCLUSIVE |
| All h=3 |
51.58 |
51.88 |
51.93 |
51.77 |
51.55 |
51.69 |
51.85 |
51.49 |
51.80 |
52.39 |
-0.11 [-0.89,+0.66] |
+0.03 [-0.59,+0.64] |
+0.43 [-0.33,+1.23] |
-0.03 [-0.74,+0.70] |
-0.84 [-1.65,-0.03] |
-0.10 ± 0.46 |
INCONCLUSIVE |
| All h=4 |
53.26 |
53.69 |
53.54 |
53.55 |
53.40 |
53.38 |
53.41 |
53.00 |
53.10 |
53.86 |
-0.12 [-0.98,+0.70] |
+0.28 [-0.34,+0.90] |
+0.54 [-0.36,+1.46] |
+0.44 [-0.31,+1.22] |
-0.46 [-1.26,+0.32] |
+0.14 ± 0.42 |
INCONCLUSIVE |
| All h=5 |
57.43 |
57.96 |
57.79 |
57.58 |
57.50 |
57.63 |
57.43 |
57.46 |
57.45 |
58.16 |
-0.19 [-1.05,+0.65] |
+0.53 [-0.09,+1.15] |
+0.32 [-0.55,+1.27] |
+0.12 [-0.60,+0.84] |
-0.66 [-1.50,+0.16] |
+0.03 ± 0.46 |
INCONCLUSIVE |
Δ = K-v2+pre − MB-best[do0.1]+pre, per seed (n = 5 seed pairs finished; n_req = ceil(2n/3) = 4)
| cell |
A s42 |
A s123 |
A s456 |
A s789 |
A s1011 |
B s42 |
B s123 |
B s456 |
B s789 |
B s1011 |
Δ s42 [CI] |
Δ s123 [CI] |
Δ s456 [CI] |
Δ s789 [CI] |
Δ s1011 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
47.98 |
47.95 |
48.04 |
47.81 |
48.10 |
47.78 |
47.44 |
47.74 |
47.74 |
47.77 |
+0.20 [-0.34,+0.72] |
+0.51 [-0.00,+1.05] |
+0.30 [-0.21,+0.82] |
+0.07 [-0.42,+0.55] |
+0.33 [-0.24,+0.90] |
+0.28 ± 0.16 |
INCONCLUSIVE |
| d-n |
2.178 |
2.177 |
2.181 |
2.170 |
2.183 |
2.169 |
2.153 |
2.167 |
2.167 |
2.168 |
+0.009 [-0.015,+0.033] |
+0.023 [-0.000,+0.048] |
+0.014 [-0.010,+0.037] |
+0.003 [-0.019,+0.025] |
+0.015 [-0.011,+0.041] |
+0.013 ± 0.007 |
NOT_WORSE |
| HM-MPJPE |
128.58 |
129.25 |
128.45 |
128.48 |
129.87 |
128.16 |
127.07 |
127.75 |
127.88 |
128.48 |
+0.42 [-1.80,+2.63] |
+2.18 [-0.47,+5.11] |
+0.70 [-1.92,+3.42] |
+0.60 [-1.34,+2.58] |
+1.39 [-1.55,+4.38] |
+1.06 ± 0.73 |
INCONCLUSIVE |
| PCK@0.05 |
87.20 |
87.29 |
87.09 |
87.25 |
87.28 |
87.45 |
87.45 |
87.38 |
87.36 |
87.41 |
-0.24 [-0.54,+0.05] |
-0.17 [-0.47,+0.12] |
-0.29 [-0.61,+0.02] |
-0.11 [-0.37,+0.17] |
-0.13 [-0.45,+0.17] |
-0.19 ± 0.08 |
INCONCLUSIVE |
| HM h=5 |
159.20 |
160.85 |
159.24 |
159.53 |
161.58 |
159.24 |
157.09 |
158.75 |
158.93 |
159.83 |
-0.05 [-3.75,+3.54] |
+3.76 [-0.73,+8.20] |
+0.48 [-3.58,+4.53] |
+0.60 [-3.47,+4.54] |
+1.75 [-3.19,+6.91] |
+1.31 ± 1.52 |
INCONCLUSIVE |
| All body |
70.48 |
70.17 |
70.36 |
70.27 |
70.26 |
70.16 |
69.94 |
70.19 |
70.01 |
69.98 |
+0.32 [-0.28,+0.94] |
+0.24 [-0.39,+0.85] |
+0.18 [-0.47,+0.83] |
+0.26 [-0.36,+0.85] |
+0.28 [-0.39,+0.95] |
+0.25 ± 0.05 |
INCONCLUSIVE |
| All face |
25.45 |
25.64 |
25.52 |
25.28 |
25.66 |
25.44 |
25.26 |
25.40 |
25.38 |
25.28 |
+0.01 [-0.48,+0.47] |
+0.37 [-0.16,+0.94] |
+0.12 [-0.46,+0.73] |
-0.11 [-0.52,+0.29] |
+0.37 [-0.22,+0.99] |
+0.15 ± 0.22 |
INCONCLUSIVE |
| All lhand |
63.92 |
63.86 |
63.41 |
63.85 |
63.42 |
63.71 |
62.90 |
63.24 |
63.58 |
63.62 |
+0.21 [-1.11,+1.52] |
+0.96 [-0.26,+2.19] |
+0.17 [-1.19,+1.52] |
+0.27 [-0.97,+1.56] |
-0.20 [-1.61,+1.14] |
+0.28 ± 0.42 |
INCONCLUSIVE |
| All rhand |
77.14 |
76.77 |
77.93 |
76.92 |
78.00 |
76.50 |
75.91 |
76.78 |
76.69 |
77.22 |
+0.64 [-0.96,+2.24] |
+0.86 [-0.68,+2.38] |
+1.15 [-0.24,+2.51] |
+0.23 [-1.31,+1.71] |
+0.78 [-1.02,+2.52] |
+0.73 ± 0.34 |
INCONCLUSIVE |
| All h=1 |
34.69 |
34.59 |
34.78 |
34.56 |
34.79 |
34.53 |
34.00 |
34.63 |
34.52 |
34.48 |
+0.16 [-0.41,+0.69] |
+0.59 [+0.07,+1.10] |
+0.16 [-0.34,+0.63] |
+0.04 [-0.47,+0.51] |
+0.32 [-0.27,+0.87] |
+0.25 ± 0.21 |
INCONCLUSIVE |
| All h=2 |
43.96 |
44.01 |
44.06 |
43.76 |
44.04 |
43.93 |
43.58 |
43.88 |
43.96 |
43.95 |
+0.04 [-0.58,+0.67] |
+0.43 [-0.15,+1.02] |
+0.19 [-0.45,+0.83] |
-0.21 [-0.82,+0.38] |
+0.09 [-0.61,+0.79] |
+0.11 ± 0.23 |
INCONCLUSIVE |
| All h=3 |
51.22 |
51.14 |
51.28 |
51.05 |
51.43 |
51.06 |
51.02 |
51.34 |
51.20 |
51.42 |
+0.16 [-0.45,+0.79] |
+0.12 [-0.59,+0.85] |
-0.06 [-0.76,+0.67] |
-0.15 [-0.83,+0.53] |
+0.02 [-0.77,+0.80] |
+0.02 ± 0.13 |
NOT_WORSE |
| All h=4 |
53.06 |
52.95 |
53.05 |
52.88 |
53.12 |
52.52 |
52.25 |
52.24 |
52.36 |
52.38 |
+0.53 [-0.16,+1.23] |
+0.70 [+0.02,+1.38] |
+0.81 [+0.10,+1.53] |
+0.52 [-0.11,+1.15] |
+0.74 [-0.02,+1.54] |
+0.66 ± 0.13 |
INCONCLUSIVE |
| All h=5 |
56.97 |
57.07 |
57.01 |
56.79 |
57.09 |
56.88 |
56.33 |
56.61 |
56.62 |
56.61 |
+0.10 [-0.67,+0.84] |
+0.74 [-0.04,+1.52] |
+0.40 [-0.30,+1.11] |
+0.17 [-0.52,+0.84] |
+0.48 [-0.29,+1.26] |
+0.38 ± 0.26 |
INCONCLUSIVE |
Secondary: AUC of the hazard probability haz_P (max / mean over parts × horizons) for HM membership (descriptive)
| seed |
split |
AUC(max haz_P) |
AUC(mean haz_P) |
| 42 |
val |
0.812 |
0.832 |
| 42 |
test |
0.831 |
0.845 |
| 123 |
val |
0.793 |
0.796 |
| 123 |
test |
0.827 |
0.835 |
| 456 |
val |
0.779 |
0.785 |
| 456 |
test |
0.830 |
0.845 |
| 789 |
val |
0.796 |
0.801 |
| 789 |
test |
0.827 |
0.837 |
| 1011 |
val |
0.797 |
0.801 |
| 1011 |
test |
0.812 |
0.829 |
2.2 lam03 (--haz-weight 0.3)
| seed |
arm |
run |
params |
best ep / epochs / stop |
GPU min |
val All |
All [CI] |
d-n [CI] |
HM [CI] |
PCK@0.05 [CI] |
HM h=5 [CI] |
| 42 |
K-v2 |
cwm3d_v2_lam03_s42_v2 |
2736282 |
87 / 128 / early_stop(patience=40) |
5.51 |
55.019 |
48.51 [44.76, 52.63] |
2.202 [2.032, 2.389] |
130.39 [108.20, 158.69] |
87.00 [85.84, 88.14] |
161.80 [118.48, 220.73] |
| 42 |
K-v2 + pre |
cwm3d_v2_lam03_batonpre_trunkp123_s42_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 123 |
K-v2 |
cwm3d_v2_lam03_s123_v2 |
2736282 |
76 / 117 / early_stop(patience=40) |
4.47 |
55.297 |
48.53 [44.78, 52.64] |
2.203 [2.033, 2.389] |
130.48 [108.05, 158.75] |
87.07 [85.94, 88.19] |
162.10 [118.09, 221.88] |
| 123 |
K-v2 + pre |
cwm3d_v2_lam03_batonpre_trunkp123_s123_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 456 |
K-v2 |
cwm3d_v2_lam03_s456_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 456 |
K-v2 + pre |
cwm3d_v2_lam03_batonpre_trunkp123_s456_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 789 |
K-v2 |
cwm3d_v2_lam03_s789_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 789 |
K-v2 + pre |
cwm3d_v2_lam03_batonpre_trunkp123_s789_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 1011 |
K-v2 |
cwm3d_v2_lam03_s1011_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 1011 |
K-v2 + pre |
cwm3d_v2_lam03_batonpre_trunkp123_s1011_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
Δ = K-v2 − K, per seed (n = 2 seed pairs finished; n_req = ceil(2n/3) = 2)
| cell |
A s42 |
A s123 |
B s42 |
B s123 |
Δ s42 [CI] |
Δ s123 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
48.51 |
48.53 |
48.77 |
48.66 |
-0.26 [-0.72,+0.19] |
-0.13 [-0.71,+0.35] |
-0.19 ± 0.09 |
INCONCLUSIVE |
| d-n |
2.202 |
2.203 |
2.214 |
2.209 |
-0.012 [-0.032,+0.009] |
-0.006 [-0.032,+0.016] |
-0.009 ± 0.004 |
NOT_WORSE |
| HM-MPJPE |
130.39 |
130.48 |
132.49 |
131.05 |
-2.10 [-4.70,+0.41] |
-0.57 [-2.38,+1.33] |
-1.34 ± 1.09 |
INCONCLUSIVE |
| PCK@0.05 |
87.00 |
87.07 |
87.00 |
86.99 |
+0.00 [-0.20,+0.20] |
+0.08 [-0.18,+0.42] |
+0.04 ± 0.06 |
NOT_WORSE |
| HM h=5 |
161.80 |
162.10 |
164.19 |
163.36 |
-2.39 [-6.42,+1.45] |
-1.26 [-4.39,+1.79] |
-1.83 ± 0.80 |
INCONCLUSIVE |
| All body |
70.83 |
70.75 |
71.39 |
71.29 |
-0.56 [-1.16,+0.03] |
-0.54 [-1.23,+0.03] |
-0.55 ± 0.02 |
INCONCLUSIVE |
| All face |
25.54 |
25.61 |
25.64 |
25.86 |
-0.10 [-0.58,+0.34] |
-0.25 [-0.78,+0.20] |
-0.18 ± 0.10 |
INCONCLUSIVE |
| All lhand |
65.34 |
65.44 |
65.67 |
65.11 |
-0.33 [-1.62,+0.91] |
+0.34 [-0.79,+1.52] |
+0.00 ± 0.47 |
INCONCLUSIVE |
| All rhand |
78.43 |
78.34 |
78.72 |
78.03 |
-0.29 [-1.47,+0.96] |
+0.31 [-0.83,+1.43] |
+0.01 ± 0.42 |
INCONCLUSIVE |
| All h=1 |
34.87 |
35.11 |
35.08 |
35.03 |
-0.21 [-0.79,+0.37] |
+0.08 [-0.54,+0.59] |
-0.06 ± 0.20 |
NOT_WORSE |
| All h=2 |
44.63 |
44.53 |
44.84 |
44.66 |
-0.21 [-0.91,+0.42] |
-0.12 [-0.82,+0.46] |
-0.17 ± 0.06 |
NOT_WORSE |
| All h=3 |
51.92 |
51.97 |
52.24 |
52.04 |
-0.32 [-1.02,+0.33] |
-0.06 [-0.77,+0.56] |
-0.19 ± 0.18 |
INCONCLUSIVE |
| All h=4 |
53.62 |
53.56 |
53.91 |
53.76 |
-0.29 [-0.90,+0.31] |
-0.20 [-0.90,+0.44] |
-0.24 ± 0.06 |
INCONCLUSIVE |
| All h=5 |
57.52 |
57.48 |
57.77 |
57.81 |
-0.25 [-0.86,+0.33] |
-0.33 [-0.98,+0.26] |
-0.29 ± 0.06 |
INCONCLUSIVE |
Δ = K-v2+pre − K+pre, per seed (n = 0 seed pairs finished; n_req = ceil(2n/3) = —)
(pending — no finished seed pair)
Δ = K-v2 − MB-best[do0.1], per seed (n = 2 seed pairs finished; n_req = ceil(2n/3) = 2)
| cell |
A s42 |
A s123 |
B s42 |
B s123 |
Δ s42 [CI] |
Δ s123 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
48.51 |
48.53 |
48.39 |
48.41 |
+0.12 [-0.47,+0.68] |
+0.12 [-0.25,+0.48] |
+0.12 ± 0.00 |
NOT_WORSE |
| d-n |
2.202 |
2.203 |
2.197 |
2.197 |
+0.006 [-0.022,+0.031] |
+0.006 [-0.011,+0.022] |
+0.006 ± 0.000 |
NOT_WORSE |
| HM-MPJPE |
130.39 |
130.48 |
130.48 |
129.51 |
-0.09 [-2.76,+2.71] |
+0.98 [-0.50,+2.61] |
+0.44 ± 0.76 |
INCONCLUSIVE |
| PCK@0.05 |
87.00 |
87.07 |
87.26 |
87.11 |
-0.25 [-0.55,+0.02] |
-0.04 [-0.27,+0.21] |
-0.15 ± 0.15 |
INCONCLUSIVE |
| HM h=5 |
161.80 |
162.10 |
163.78 |
162.75 |
-1.98 [-7.13,+3.15] |
-0.65 [-3.66,+2.07] |
-1.31 ± 0.94 |
INCONCLUSIVE |
| All body |
70.83 |
70.75 |
70.88 |
70.94 |
-0.04 [-0.60,+0.49] |
-0.19 [-0.80,+0.35] |
-0.12 ± 0.10 |
NOT_WORSE |
| All face |
25.54 |
25.61 |
25.39 |
25.70 |
+0.15 [-0.52,+0.86] |
-0.09 [-0.52,+0.35] |
+0.03 ± 0.17 |
NOT_WORSE |
| All lhand |
65.34 |
65.44 |
65.53 |
65.27 |
-0.19 [-1.37,+0.87] |
+0.17 [-0.70,+1.05] |
-0.01 ± 0.26 |
NOT_WORSE |
| All rhand |
78.43 |
78.34 |
77.89 |
77.19 |
+0.54 [-0.66,+1.70] |
+1.15 [+0.10,+2.24] |
+0.85 ± 0.43 |
INCONCLUSIVE |
| All h=1 |
34.87 |
35.11 |
34.97 |
34.99 |
-0.10 [-0.61,+0.38] |
+0.13 [-0.34,+0.51] |
+0.02 ± 0.16 |
NOT_WORSE |
| All h=2 |
44.63 |
44.53 |
44.28 |
44.37 |
+0.35 [-0.32,+0.97] |
+0.16 [-0.34,+0.64] |
+0.25 ± 0.13 |
INCONCLUSIVE |
| All h=3 |
51.92 |
51.97 |
51.69 |
51.85 |
+0.23 [-0.55,+0.97] |
+0.13 [-0.40,+0.65] |
+0.18 ± 0.07 |
NOT_WORSE |
| All h=4 |
53.62 |
53.56 |
53.38 |
53.41 |
+0.24 [-0.61,+1.05] |
+0.16 [-0.33,+0.65] |
+0.20 ± 0.06 |
NOT_WORSE |
| All h=5 |
57.52 |
57.48 |
57.63 |
57.43 |
-0.11 [-0.96,+0.70] |
+0.04 [-0.41,+0.50] |
-0.03 ± 0.11 |
NOT_WORSE |
Δ = K-v2+pre − MB-best[do0.1]+pre, per seed (n = 0 seed pairs finished; n_req = ceil(2n/3) = —)
(pending — no finished seed pair)
Secondary: AUC of the hazard probability haz_P (max / mean over parts × horizons) for HM membership (descriptive)
| seed |
split |
AUC(max haz_P) |
AUC(mean haz_P) |
| 42 |
val |
0.800 |
0.798 |
| 42 |
test |
0.831 |
0.835 |
| 123 |
val |
0.792 |
0.800 |
| 123 |
test |
0.824 |
0.823 |
3. S3 — MotionBERT-kp fairness arm (same val budget: 9 new grid points + existing default = 10; MB-best = argmin val All, ties 0.10 → fewer params)
- Decision source:
aide/results/kv2/s3_decision.json written 2026-09-14T23:49:07Z (complete = True). Rule: MB-best = argmin val All over the 10 grid points (9 new + existing default); ties within 0.1 px -> fewer parameters.
| cfg |
run |
model |
flags |
params |
best ep / epochs / stop / GPU min |
val All |
Δ val All vs default [CI] |
test All (recorded, not used) |
S3 |
| default |
motionbert_kp_s42_v2 |
motionbert_kp |
(default) |
813088 |
81 / 122 / early_stop(patience=40) / 4.6 |
55.550 |
+0.000 [+0.000,+0.000] |
48.279 |
|
| d192 |
motionbert_kp_d192_s42_v2 |
motionbert_kp_d192 |
(default) |
1809440 |
64 / 105 / early_stop(patience=40) / 6.14 |
55.559 |
+0.009 [-0.573,+0.575] |
48.595 |
|
| L4 |
motionbert_kp_L4_s42_v2 |
motionbert_kp_L4 |
(default) |
1077664 |
82 / 123 / early_stop(patience=40) / 4.71 |
55.433 |
-0.117 [-0.572,+0.313] |
48.488 |
|
| lr1e-4 |
motionbert_kp_lr1e-4_s42_v2 |
motionbert_kp |
--lr 1e-4 |
813088 |
79 / 120 / early_stop(patience=40) / 3.18 |
55.310 |
-0.240 [-0.773,+0.247] |
48.493 |
|
| lr1e-3 |
motionbert_kp_lr1e-3_s42_v2 |
motionbert_kp |
--lr 1e-3 |
813088 |
60 / 101 / early_stop(patience=40) / 2.75 |
55.281 |
-0.269 [-0.732,+0.131] |
48.394 |
|
| wd0.01 |
motionbert_kp_wd0.01_s42_v2 |
motionbert_kp |
--wd 0.01 |
813088 |
73 / 114 / early_stop(patience=40) / 3.15 |
55.366 |
-0.184 [-0.659,+0.199] |
48.342 |
|
| wd0.1 |
motionbert_kp_wd0.1_s42_v2 |
motionbert_kp |
--wd 0.1 |
813088 |
64 / 105 / early_stop(patience=40) / 2.83 |
55.603 |
+0.053 [-0.324,+0.440] |
48.424 |
|
| do0.1 |
motionbert_kp_do0.1_s42_v2 |
motionbert_kp |
--dropout 0.1 |
813088 |
95 / 136 / early_stop(patience=40) / 3.89 |
55.215 |
-0.335 [-0.888,+0.129] |
48.389 |
MB-best |
| d192_L4 |
motionbert_kp_d192_L4_s42_v2 |
motionbert_kp_d192_L4 |
(default) |
2402912 |
67 / 108 / early_stop(patience=40) / 5.18 |
55.399 |
-0.151 [-0.632,+0.325] |
48.418 |
|
| d192_L4_do0.1 |
motionbert_kp_d192_L4_do0.1_s42_v2 |
motionbert_kp_d192_L4 |
--dropout 0.1 |
2402912 |
55 / 96 / early_stop(patience=40) / 3.82 |
55.425 |
-0.125 [-0.587,+0.327] |
48.477 |
|
MB-best = do0.1 (final; model motionbert_kp, flags ['--dropout', '0.1']; changes d/L: False → existing motionbert_kp_batonpre_pretrain_s123 checkpoint reused).
3.1 MB-best and MB-best + BATON trunk-only init at 5 seeds; K + pre completion (test levels)
| seed |
arm |
run |
params |
best ep / epochs / stop |
GPU min |
val All |
All [CI] |
d-n [CI] |
HM [CI] |
PCK@0.05 [CI] |
HM h=5 [CI] |
| 42 |
MB default |
motionbert_kp_s42_v2 |
813088 |
81 / 122 / early_stop(patience=40) |
4.6 |
55.550 |
48.28 [44.53, 52.37] |
2.192 [2.021, 2.377] |
130.34 [108.20, 158.53] |
87.17 [86.04, 88.28] |
163.95 [119.85, 223.86] |
| 42 |
MB default + pre |
motionbert_kp_batonpre_trunkp123_s42_v2 |
813088 |
37 / 78 / early_stop(patience=40) |
1.94 |
54.528 |
47.74 [44.07, 51.78] |
2.167 [2.000, 2.350] |
128.03 [105.81, 156.47] |
87.46 [86.37, 88.54] |
159.20 [115.09, 219.85] |
| 42 |
MB-best |
motionbert_kp_do0.1_s42_v2 |
813088 |
95 / 136 / early_stop(patience=40) |
3.89 |
55.215 |
48.39 [44.59, 52.49] |
2.197 [2.024, 2.383] |
130.48 [107.84, 158.95] |
87.26 [86.15, 88.36] |
163.78 [118.78, 224.72] |
| 42 |
MB-best + pre |
motionbert_kp_do0.1_batonpre_trunkp123_s42_v2 |
813088 |
41 / 82 / early_stop(patience=40) |
2.59 |
54.575 |
47.78 [44.10, 51.83] |
2.169 [2.002, 2.353] |
128.16 [105.87, 156.59] |
87.45 [86.35, 88.53] |
159.24 [114.93, 219.57] |
| 42 |
K + pre |
cwm3d_kp_batonpre_trunkp123_s42_v2 |
5771449 |
31 / 72 / early_stop(patience=40) |
0.48 |
55.003 |
47.98 [44.28, 52.04] |
2.178 [2.010, 2.363] |
128.43 [106.25, 157.11] |
87.21 [86.10, 88.30] |
159.35 [115.56, 219.46] |
| 123 |
MB default |
motionbert_kp_s123_v2 |
813088 |
58 / 99 / early_stop(patience=40) |
4.36 |
55.627 |
48.38 [44.65, 52.46] |
2.196 [2.027, 2.381] |
129.16 [106.70, 157.58] |
87.14 [86.02, 88.27] |
162.59 [118.87, 222.08] |
| 123 |
MB default + pre |
motionbert_kp_batonpre_trunkp123_s123_v2 |
813088 |
38 / 79 / early_stop(patience=40) |
2.16 |
54.720 |
47.53 [43.87, 51.55] |
2.158 [1.992, 2.340] |
126.11 [103.96, 154.69] |
87.40 [86.32, 88.46] |
156.47 [113.00, 215.99] |
| 123 |
MB-best |
motionbert_kp_do0.1_s123_v2 |
813088 |
54 / 95 / early_stop(patience=40) |
2.1 |
55.656 |
48.41 [44.69, 52.49] |
2.197 [2.029, 2.383] |
129.51 [107.22, 158.01] |
87.11 [85.97, 88.23] |
162.75 [118.82, 222.30] |
| 123 |
MB-best + pre |
motionbert_kp_do0.1_batonpre_trunkp123_s123_v2 |
813088 |
42 / 83 / early_stop(patience=40) |
2.68 |
54.666 |
47.44 [43.78, 51.45] |
2.153 [1.987, 2.336] |
127.07 [105.00, 155.58] |
87.45 [86.36, 88.52] |
157.09 [113.60, 216.60] |
| 123 |
K + pre |
cwm3d_kp_batonpre_trunkp123_s123_v2 |
5771449 |
33 / 74 / early_stop(patience=40) |
0.88 |
54.828 |
48.04 [44.33, 52.11] |
2.181 [2.012, 2.365] |
129.13 [106.80, 157.67] |
87.20 [86.07, 88.32] |
160.11 [116.11, 220.04] |
| 456 |
MB default |
motionbert_kp_s456_v2 |
813088 |
90 / 131 / early_stop(patience=40) |
5.17 |
55.166 |
48.48 [44.74, 52.58] |
2.201 [2.031, 2.387] |
130.23 [107.78, 158.89] |
87.14 [86.00, 88.25] |
162.24 [116.93, 223.50] |
| 456 |
MB default + pre |
motionbert_kp_batonpre_trunkp123_s456_v2 |
813088 |
44 / 85 / early_stop(patience=40) |
1.68 |
54.797 |
47.91 [44.22, 51.93] |
2.175 [2.008, 2.357] |
127.59 [105.28, 156.14] |
87.39 [86.28, 88.47] |
158.92 [114.76, 219.74] |
| 456 |
MB-best |
motionbert_kp_do0.1_s456_v2 |
813088 |
65 / 106 / early_stop(patience=40) |
2.5 |
55.356 |
48.21 [44.50, 52.28] |
2.188 [2.020, 2.373] |
128.99 [106.63, 157.76] |
87.30 [86.18, 88.39] |
162.24 [117.69, 222.93] |
| 456 |
MB-best + pre |
motionbert_kp_do0.1_batonpre_trunkp123_s456_v2 |
813088 |
43 / 84 / early_stop(patience=40) |
2.74 |
54.828 |
47.74 [44.07, 51.79] |
2.167 [2.001, 2.351] |
127.75 [105.54, 156.19] |
87.38 [86.27, 88.46] |
158.75 [114.66, 219.20] |
| 456 |
K + pre |
cwm3d_kp_batonpre_trunkp123_s456_v2 |
5771449 |
36 / 77 / early_stop(patience=40) |
0.91 |
54.670 |
48.01 [44.31, 52.12] |
2.179 [2.011, 2.366] |
128.51 [106.16, 157.20] |
87.16 [86.02, 88.27] |
159.02 [115.09, 219.40] |
| 789 |
MB default |
motionbert_kp_s789_v2 |
813088 |
77 / 118 / early_stop(patience=40) |
6.14 |
55.134 |
48.46 [44.71, 52.57] |
2.200 [2.030, 2.386] |
130.66 [108.19, 159.30] |
87.27 [86.16, 88.36] |
163.04 [118.30, 223.77] |
| 789 |
MB default + pre |
motionbert_kp_batonpre_trunkp123_s789_v2 |
813088 |
37 / 78 / early_stop(patience=40) |
3.45 |
54.882 |
47.55 [43.88, 51.56] |
2.159 [1.992, 2.341] |
127.71 [105.39, 156.61] |
87.40 [86.30, 88.48] |
159.11 [114.88, 219.30] |
| 789 |
MB-best |
motionbert_kp_do0.1_s789_v2 |
813088 |
89 / 130 / early_stop(patience=40) |
3.43 |
55.295 |
48.37 [44.60, 52.47] |
2.196 [2.025, 2.382] |
130.78 [108.51, 159.34] |
87.20 [86.08, 88.30] |
164.33 [119.77, 224.96] |
| 789 |
MB-best + pre |
motionbert_kp_do0.1_batonpre_trunkp123_s789_v2 |
813088 |
40 / 81 / early_stop(patience=40) |
2.5 |
54.786 |
47.74 [44.06, 51.78] |
2.167 [2.000, 2.350] |
127.88 [105.66, 156.44] |
87.36 [86.25, 88.45] |
158.93 [114.68, 219.29] |
| 789 |
K + pre |
cwm3d_kp_batonpre_trunkp123_s789_v2 |
5771449 |
35 / 76 / early_stop(patience=40) |
1.22 |
54.590 |
48.12 [44.44, 52.21] |
2.184 [2.017, 2.370] |
129.32 [107.17, 157.78] |
87.06 [85.94, 88.18] |
161.39 [117.32, 221.77] |
| 1011 |
MB default |
motionbert_kp_s1011_v2 |
813088 |
81 / 122 / early_stop(patience=40) |
5.69 |
55.506 |
48.68 [44.92, 52.82] |
2.210 [2.039, 2.398] |
131.49 [108.92, 160.13] |
86.97 [85.81, 88.12] |
166.22 [121.36, 227.01] |
| 1011 |
MB default + pre |
motionbert_kp_batonpre_trunkp123_s1011_v2 |
813088 |
47 / 88 / early_stop(patience=40) |
3.98 |
54.868 |
47.85 [44.16, 51.87] |
2.172 [2.005, 2.355] |
129.47 [107.17, 157.92] |
87.40 [86.30, 88.49] |
160.75 [116.16, 221.59] |
| 1011 |
MB-best |
motionbert_kp_do0.1_s1011_v2 |
813088 |
138 / 179 / early_stop(patience=40) |
4.66 |
55.253 |
48.92 [45.11, 53.00] |
2.221 [2.048, 2.406] |
130.61 [108.46, 158.67] |
86.90 [85.74, 88.04] |
163.65 [119.68, 223.28] |
| 1011 |
MB-best + pre |
motionbert_kp_do0.1_batonpre_trunkp123_s1011_v2 |
813088 |
59 / 100 / early_stop(patience=40) |
2.71 |
54.566 |
47.77 [44.07, 51.75] |
2.168 [2.001, 2.349] |
128.48 [106.50, 156.68] |
87.41 [86.31, 88.48] |
159.83 [115.53, 220.28] |
| 1011 |
K + pre |
cwm3d_kp_batonpre_trunkp123_s1011_v2 |
5771449 |
34 / 75 / early_stop(patience=40) |
1.14 |
54.609 |
48.11 [44.40, 52.21] |
2.184 [2.016, 2.370] |
128.73 [106.48, 157.39] |
87.03 [85.91, 88.16] |
159.72 [115.63, 220.35] |
Δ = MB-best[do0.1]+pre − MB-best[do0.1], per seed (descriptive; n = 5)
| cell |
A s42 |
A s123 |
A s456 |
A s789 |
A s1011 |
B s42 |
B s123 |
B s456 |
B s789 |
B s1011 |
Δ s42 [CI] |
Δ s123 [CI] |
Δ s456 [CI] |
Δ s789 [CI] |
Δ s1011 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
47.78 |
47.44 |
47.74 |
47.74 |
47.77 |
48.39 |
48.41 |
48.21 |
48.37 |
48.92 |
-0.61 [-1.17,-0.07] |
-0.97 [-1.50,-0.46] |
-0.47 [-0.96,+0.03] |
-0.63 [-1.10,-0.16] |
-1.15 [-1.82,-0.53] |
-0.77 ± 0.28 |
BEATS |
| d-n |
2.169 |
2.153 |
2.167 |
2.167 |
2.168 |
2.197 |
2.197 |
2.188 |
2.196 |
2.221 |
-0.028 [-0.053,-0.003] |
-0.044 [-0.068,-0.021] |
-0.021 [-0.044,+0.001] |
-0.029 [-0.050,-0.007] |
-0.052 [-0.082,-0.024] |
-0.035 ± 0.013 |
BEATS |
| HM-MPJPE |
128.16 |
127.07 |
127.75 |
127.88 |
128.48 |
130.48 |
129.51 |
128.99 |
130.78 |
130.61 |
-2.32 [-4.83,-0.06] |
-2.44 [-4.62,-0.37] |
-1.25 [-3.56,+0.87] |
-2.90 [-5.13,-0.88] |
-2.13 [-4.94,+0.62] |
-2.21 ± 0.61 |
INCONCLUSIVE |
| PCK@0.05 |
87.45 |
87.45 |
87.38 |
87.36 |
87.41 |
87.26 |
87.11 |
87.30 |
87.20 |
86.90 |
+0.19 [-0.09,+0.46] |
+0.35 [+0.08,+0.63] |
+0.08 [-0.16,+0.33] |
+0.15 [-0.06,+0.36] |
+0.51 [+0.19,+0.87] |
+0.26 ± 0.17 |
INCONCLUSIVE |
| HM h=5 |
159.24 |
157.09 |
158.75 |
158.93 |
159.83 |
163.78 |
162.75 |
162.24 |
164.33 |
163.65 |
-4.54 [-8.85,-0.30] |
-5.66 [-10.34,-1.22] |
-3.49 [-8.12,+0.52] |
-5.41 [-9.75,-1.54] |
-3.82 [-8.18,+0.25] |
-4.58 ± 0.95 |
INCONCLUSIVE |
Δ = K+pre − K, per seed (descriptive; n = 5)
| cell |
A s42 |
A s123 |
A s456 |
A s789 |
A s1011 |
B s42 |
B s123 |
B s456 |
B s789 |
B s1011 |
Δ s42 [CI] |
Δ s123 [CI] |
Δ s456 [CI] |
Δ s789 [CI] |
Δ s1011 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
47.98 |
48.04 |
48.01 |
48.12 |
48.11 |
48.77 |
48.66 |
48.46 |
48.54 |
48.57 |
-0.79 [-1.29,-0.30] |
-0.62 [-1.17,-0.11] |
-0.45 [-0.89,-0.02] |
-0.43 [-0.88,+0.03] |
-0.46 [-0.89,-0.02] |
-0.55 ± 0.15 |
BEATS |
| d-n |
2.178 |
2.181 |
2.179 |
2.184 |
2.184 |
2.214 |
2.209 |
2.200 |
2.204 |
2.205 |
-0.036 [-0.059,-0.014] |
-0.028 [-0.053,-0.005] |
-0.021 [-0.041,-0.001] |
-0.019 [-0.040,+0.001] |
-0.021 [-0.041,-0.001] |
-0.025 ± 0.007 |
BEATS |
| HM-MPJPE |
128.43 |
129.13 |
128.51 |
129.32 |
128.73 |
132.49 |
131.05 |
131.37 |
131.73 |
131.22 |
-4.06 [-6.57,-1.74] |
-1.92 [-4.14,+0.22] |
-2.86 [-5.09,-0.73] |
-2.41 [-4.43,-0.48] |
-2.49 [-4.79,-0.27] |
-2.75 ± 0.81 |
BEATS |
| PCK@0.05 |
87.21 |
87.20 |
87.16 |
87.06 |
87.03 |
87.00 |
86.99 |
87.13 |
87.00 |
87.00 |
+0.20 [-0.05,+0.45] |
+0.22 [-0.07,+0.52] |
+0.03 [-0.20,+0.26] |
+0.06 [-0.16,+0.28] |
+0.03 [-0.20,+0.26] |
+0.11 ± 0.09 |
NOT_WORSE |
| HM h=5 |
159.35 |
160.11 |
159.02 |
161.39 |
159.72 |
164.19 |
163.36 |
162.55 |
163.04 |
163.70 |
-4.85 [-8.95,-1.00] |
-3.25 [-6.97,+0.55] |
-3.53 [-7.22,+0.08] |
-1.65 [-5.11,+1.98] |
-3.98 [-8.22,+0.07] |
-3.45 ± 1.17 |
INCONCLUSIVE |
Δ = K − MB-best[do0.1], per seed (descriptive; n = 5)
| cell |
A s42 |
A s123 |
A s456 |
A s789 |
A s1011 |
B s42 |
B s123 |
B s456 |
B s789 |
B s1011 |
Δ s42 [CI] |
Δ s123 [CI] |
Δ s456 [CI] |
Δ s789 [CI] |
Δ s1011 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
48.77 |
48.66 |
48.46 |
48.54 |
48.57 |
48.39 |
48.41 |
48.21 |
48.37 |
48.92 |
+0.38 [-0.21,+0.96] |
+0.25 [-0.15,+0.69] |
+0.25 [-0.19,+0.69] |
+0.17 [-0.29,+0.64] |
-0.34 [-0.98,+0.25] |
+0.14 ± 0.28 |
INCONCLUSIVE |
| d-n |
2.214 |
2.209 |
2.200 |
2.204 |
2.205 |
2.197 |
2.197 |
2.188 |
2.196 |
2.221 |
+0.017 [-0.010,+0.044] |
+0.011 [-0.007,+0.031] |
+0.012 [-0.009,+0.032] |
+0.008 [-0.013,+0.029] |
-0.016 [-0.045,+0.011] |
+0.006 ± 0.013 |
NOT_WORSE |
| HM-MPJPE |
132.49 |
131.05 |
131.37 |
131.73 |
131.22 |
130.48 |
129.51 |
128.99 |
130.78 |
130.61 |
+2.01 [-0.71,+4.80] |
+1.54 [-0.56,+3.72] |
+2.38 [+0.08,+4.80] |
+0.95 [-1.79,+3.81] |
+0.62 [-1.60,+2.73] |
+1.50 ± 0.73 |
INCONCLUSIVE |
| PCK@0.05 |
87.00 |
86.99 |
87.13 |
87.00 |
87.00 |
87.26 |
87.11 |
87.30 |
87.20 |
86.90 |
-0.25 [-0.54,+0.01] |
-0.12 [-0.41,+0.14] |
-0.17 [-0.47,+0.12] |
-0.20 [-0.48,+0.06] |
+0.11 [-0.19,+0.41] |
-0.13 ± 0.14 |
INCONCLUSIVE |
| HM h=5 |
164.19 |
163.36 |
162.55 |
163.04 |
163.70 |
163.78 |
162.75 |
162.24 |
164.33 |
163.65 |
+0.41 [-4.94,+5.58] |
+0.61 [-3.26,+4.35] |
+0.30 [-3.97,+4.60] |
-1.29 [-6.68,+3.41] |
+0.05 [-4.50,+4.01] |
+0.02 ± 0.76 |
INCONCLUSIVE |
Δ = K+pre − MB-best[do0.1]+pre, per seed (descriptive; n = 5)
| cell |
A s42 |
A s123 |
A s456 |
A s789 |
A s1011 |
B s42 |
B s123 |
B s456 |
B s789 |
B s1011 |
Δ s42 [CI] |
Δ s123 [CI] |
Δ s456 [CI] |
Δ s789 [CI] |
Δ s1011 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
47.98 |
48.04 |
48.01 |
48.12 |
48.11 |
47.78 |
47.44 |
47.74 |
47.74 |
47.77 |
+0.20 [-0.29,+0.70] |
+0.60 [+0.11,+1.12] |
+0.27 [-0.24,+0.79] |
+0.38 [-0.12,+0.90] |
+0.35 [-0.24,+0.97] |
+0.36 ± 0.15 |
INCONCLUSIVE |
| d-n |
2.178 |
2.181 |
2.179 |
2.184 |
2.184 |
2.169 |
2.153 |
2.167 |
2.167 |
2.168 |
+0.009 [-0.013,+0.032] |
+0.027 [+0.005,+0.051] |
+0.012 [-0.011,+0.036] |
+0.017 [-0.005,+0.041] |
+0.016 [-0.011,+0.044] |
+0.016 ± 0.007 |
INCONCLUSIVE |
| HM-MPJPE |
128.43 |
129.13 |
128.51 |
129.32 |
128.73 |
128.16 |
127.07 |
127.75 |
127.88 |
128.48 |
+0.27 [-2.42,+3.03] |
+2.06 [-0.67,+4.96] |
+0.76 [-1.81,+3.62] |
+1.44 [-0.64,+3.70] |
+0.25 [-3.14,+3.84] |
+0.96 ± 0.78 |
INCONCLUSIVE |
| PCK@0.05 |
87.21 |
87.20 |
87.16 |
87.06 |
87.03 |
87.45 |
87.45 |
87.38 |
87.36 |
87.41 |
-0.24 [-0.50,+0.02] |
-0.25 [-0.53,+0.01] |
-0.22 [-0.54,+0.07] |
-0.29 [-0.58,-0.02] |
-0.38 [-0.71,-0.07] |
-0.28 ± 0.06 |
INCONCLUSIVE |
| HM h=5 |
159.35 |
160.11 |
159.02 |
161.39 |
159.72 |
159.24 |
157.09 |
158.75 |
158.93 |
159.83 |
+0.10 [-4.83,+4.85] |
+3.02 [-2.27,+8.37] |
+0.26 [-4.85,+5.29] |
+2.46 [-1.86,+6.95] |
-0.11 [-5.85,+5.56] |
+1.15 ± 1.47 |
INCONCLUSIVE |
4. Primary endpoints and adoption (test All-MPJPE, 5 seeds, both sides in the same pretraining state; BEATS rule, 4/5)
- K-v2(best) = nospatial (lowest mean VAL All over finished seeds among S1b-continued survivors; ties 0.10 → fewer components): nospatial val All 54.906 over 5 seeds, 4 components.
| endpoint |
n seeds |
Δ All per seed [CI] |
mean Δ ± sd |
verdict |
secondary (HM / PCK@0.05 / HM h=5 / d-n verdicts) |
| P1 = K-v2(nospatial) − MB-best[do0.1] |
5 |
s42 -0.12 [-0.75,+0.49]; s123 +0.12 [-0.31,+0.54]; s456 +0.32 [-0.26,+0.92]; s789 +0.07 [-0.43,+0.59]; s1011 -0.62 [-1.24,-0.03] |
-0.05 ± 0.36 |
INCONCLUSIVE |
d-n INCONCLUSIVE (-0.00); HM-MPJPE INCONCLUSIVE (-0.12); PCK@0.05 INCONCLUSIVE (-0.13); HM h=5 INCONCLUSIVE (-1.82) |
| P2 = K-v2(nospatial)+pre − MB-best[do0.1]+pre |
5 |
s42 +0.20 [-0.34,+0.72]; s123 +0.51 [-0.00,+1.05]; s456 +0.30 [-0.21,+0.82]; s789 +0.07 [-0.42,+0.55]; s1011 +0.33 [-0.24,+0.90] |
+0.28 ± 0.16 |
INCONCLUSIVE |
d-n NOT_WORSE (+0.01); HM-MPJPE INCONCLUSIVE (+1.06); PCK@0.05 INCONCLUSIVE (-0.19); HM h=5 INCONCLUSIVE (+1.31) |
| A1 = K-v2(nospatial) − K |
5 |
s42 -0.50 [-1.00,-0.01]; s123 -0.13 [-0.69,+0.33]; s456 +0.07 [-0.42,+0.59]; s789 -0.10 [-0.44,+0.24]; s1011 -0.28 [-0.60,+0.05] |
-0.19 ± 0.21 |
INCONCLUSIVE |
d-n INCONCLUSIVE (-0.01); HM-MPJPE INCONCLUSIVE (-1.62); PCK@0.05 NOT_WORSE (-0.00); HM h=5 INCONCLUSIVE (-1.84) |
| A2 = K-v2(nospatial)+pre − K+pre |
5 |
s42 +0.00 [-0.37,+0.36]; s123 -0.09 [-0.47,+0.27]; s456 +0.03 [-0.40,+0.45]; s789 -0.31 [-0.72,+0.07]; s1011 -0.02 [-0.37,+0.32] |
-0.08 ± 0.14 |
INCONCLUSIVE |
d-n NOT_WORSE (-0.00); HM-MPJPE INCONCLUSIVE (+0.10); PCK@0.05 INCONCLUSIVE (+0.09); HM h=5 INCONCLUSIVE (+0.16) |
4.1 Paper-wording consequence table (pre-registered; the realised row is marked)
| case |
wording consequence |
realised |
| P1 ∧ P2 |
an architecture advantage under the keypoint formulation may be written FOR K-v2 (K's own attribution sentence unchanged) |
|
| P1 ∧ ¬P2 |
"advantage without pretraining that the BATON trunk equalises"; attribution sentence kept |
|
| ¬P1 ∧ P2 |
"advantage conditional on the pretrained trunk"; attribution sentence kept |
|
| neither |
wording unchanged verbatim; K-v2 = ablation family + gen_tables.py NEGATIVES entry; any WORSE reported as is |
← REALISED |
ADOPTION A1: NOT ADOPTED
ADOPTION A2: NOT ADOPTED
- Consequence-table row: neither. Any WORSE verdict above is reported as is.
5. GPU minutes of the programme's new runs (from run json; the GPU was shared with a DMS feature-extraction job and other lanes, so minutes are wall-clock upper bounds)
| run |
GPU min |
| cwm3d_kp_batonpre_trunkp123_s1011_v2 |
1.14 |
| cwm3d_kp_batonpre_trunkp123_s789_v2 |
1.22 |
| cwm3d_v2_hazaux_s42_v2 |
4.51 |
| cwm3d_v2_khead_s42_v2 |
4.75 |
| cwm3d_v2_lam03_s123_v2 |
4.47 |
| cwm3d_v2_lam03_s42_v2 |
5.51 |
| cwm3d_v2_nohaz_s42_v2 |
5.71 |
| cwm3d_v2_nospatial_batonpre_pretrain_s123 |
10.72 |
| cwm3d_v2_nospatial_batonpre_trunkp123_s1011_v2 |
1.12 |
| cwm3d_v2_nospatial_batonpre_trunkp123_s123_v2 |
1.09 |
| cwm3d_v2_nospatial_batonpre_trunkp123_s42_v2 |
1.21 |
| cwm3d_v2_nospatial_batonpre_trunkp123_s456_v2 |
1.12 |
| cwm3d_v2_nospatial_batonpre_trunkp123_s789_v2 |
1.08 |
| cwm3d_v2_nospatial_s1011_v2 |
1.58 |
| cwm3d_v2_nospatial_s123_v2 |
3.71 |
| cwm3d_v2_nospatial_s42_v2 |
1.69 |
| cwm3d_v2_nospatial_s456_v2 |
3.67 |
| cwm3d_v2_nospatial_s789_v2 |
1.65 |
| cwm3d_v2_plain_s42_v2 |
3.69 |
| cwm3d_v2_q060_s42_v2 |
4.32 |
| cwm3d_v2_q090_s42_v2 |
4.19 |
| cwm3d_v2_s42_v2 |
2.25 |
| motionbert_kp_L4_s42_v2 |
4.71 |
| motionbert_kp_batonpre_trunkp123_s1011_v2 |
3.98 |
| motionbert_kp_batonpre_trunkp123_s789_v2 |
3.45 |
| motionbert_kp_d192_L4_do0.1_s42_v2 |
3.82 |
| motionbert_kp_d192_L4_s42_v2 |
5.18 |
| motionbert_kp_d192_s42_v2 |
6.14 |
| motionbert_kp_do0.1_batonpre_trunkp123_s1011_v2 |
2.71 |
| motionbert_kp_do0.1_batonpre_trunkp123_s123_v2 |
2.68 |
| motionbert_kp_do0.1_batonpre_trunkp123_s42_v2 |
2.59 |
| motionbert_kp_do0.1_batonpre_trunkp123_s456_v2 |
2.74 |
| motionbert_kp_do0.1_batonpre_trunkp123_s789_v2 |
2.50 |
| motionbert_kp_do0.1_s1011_v2 |
4.66 |
| motionbert_kp_do0.1_s123_v2 |
2.10 |
| motionbert_kp_do0.1_s42_v2 |
3.89 |
| motionbert_kp_do0.1_s456_v2 |
2.50 |
| motionbert_kp_do0.1_s789_v2 |
3.43 |
| motionbert_kp_lr1e-3_s42_v2 |
2.75 |
| motionbert_kp_lr1e-4_s42_v2 |
3.18 |
| motionbert_kp_wd0.01_s42_v2 |
3.15 |
| motionbert_kp_wd0.1_s42_v2 |
2.83 |
| total (42 runs) |
139.4 |
6. Provenance
- Inputs:
pred_val.npz / pred_test.npz / metrics_val.json / metrics_test.json / run_<name>_<utc>.json under /home/USER/cwm3/aide/results/runs; gate decisions /home/USER/cwm3/aide/results/kv2: pretrain_check_nospatial.json, pretrain_check_nospatial_20260915T000040Z.json, pretrain_check_nospatial_own.json, pretrain_check_nospatial_own_20260915T001217Z.json, s1_decision.json, s1_decision_20260914T234548Z.json, s1b_lam03_decision.json, s1b_lam03_decision_20260915T000039Z.json, s1b_nospatial_decision.json, s1b_nospatial_decision_20260915T000038Z.json, s3_decision.json, s3_decision_20260914T230840Z.json, s3_decision_20260914T234907Z.json; pre-registration row from /home/USER/cwm3/docs/decisions.md (S0 UTC 2026-09-14T22:51:07Z).
- Comparison runs
cwm3d_kp_s<k>_v2, cwm3d_kp_batonpre_trunkp123_s{42,123,456}_v2, motionbert_kp_s<k>_v2, motionbert_kp_batonpre_trunkp123_s{42,123,456}_v2 are the EXISTING runs (never retrained); every programme run is a new run dir.
- Selection: val All-MPJPE of the EMA weights inside the trainer; S1 / S1b / S3 decisions from pred_val.npz only (
aide/eval/kv2_gate.py); one test evaluation per run by the harness; no test number was used for any decision.
- Launcher
aide/train/run_kv2_program.sh (gate: free VRAM ≥ need + 4 GB, MemAvailable ≥ 28 GB, < 4 trainers, OMP/MKL 4; finished runs skipped); pids aide/results/kv2/pids.txt.
- Pending runs referenced above: 0. This block is append-only; earlier blocks are never edited.
- Trainer code hashes (
code_sha256 of each finished programme run's run json; distinct values = code changed between launches, e.g. the train_kp.py:630 guard fix of 2026-09-14T23:11:53Z, behaviour-neutral for dict-output models — see aide/results/kv2/NOTES.md): e313c0ff89746293 (20 runs: motionbert_kp_L4_s42_v2, motionbert_kp_batonpre_trunkp123_s1011_v2, motionbert_kp_batonpre_trunkp123_s789_v2, motionbert_kp_d192_L4_do0.1_s42_v2, motionbert_kp_d192_L4_s42_v2, motionbert_kp_d192_s42_v2, motionbert_kp_do0.1_batonpre_trunkp123_s1011_v2, motionbert_kp_do0.1_batonpre_trunkp123_s123_v2, motionbert_kp_do0.1_batonpre_trunkp123_s42_v2, motionbert_kp_do0.1_batonpre_trunkp123_s456_v2, motionbert_kp_do0.1_batonpre_trunkp123_s789_v2, motionbert_kp_do0.1_s1011_v2, motionbert_kp_do0.1_s123_v2, motionbert_kp_do0.1_s42_v2, motionbert_kp_do0.1_s456_v2, motionbert_kp_do0.1_s789_v2, motionbert_kp_lr1e-3_s42_v2, motionbert_kp_lr1e-4_s42_v2, motionbert_kp_wd0.01_s42_v2, motionbert_kp_wd0.1_s42_v2); 8caba4a8efccd311 (18 runs: cwm3d_v2_hazaux_s42_v2, cwm3d_v2_khead_s42_v2, cwm3d_v2_lam03_s123_v2, cwm3d_v2_lam03_s42_v2, cwm3d_v2_nohaz_s42_v2, cwm3d_v2_nospatial_batonpre_trunkp123_s1011_v2, cwm3d_v2_nospatial_batonpre_trunkp123_s123_v2, cwm3d_v2_nospatial_batonpre_trunkp123_s42_v2, cwm3d_v2_nospatial_batonpre_trunkp123_s456_v2, cwm3d_v2_nospatial_batonpre_trunkp123_s789_v2, cwm3d_v2_nospatial_s1011_v2, cwm3d_v2_nospatial_s123_v2, cwm3d_v2_nospatial_s42_v2, cwm3d_v2_nospatial_s456_v2, cwm3d_v2_nospatial_s789_v2, cwm3d_v2_plain_s42_v2, cwm3d_v2_q060_s42_v2, cwm3d_v2_q090_s42_v2); 22306515fd4a42a0 (2 runs: cwm3d_kp_batonpre_trunkp123_s1011_v2, cwm3d_kp_batonpre_trunkp123_s789_v2); 652c4ba49ba937f2 (1 runs: cwm3d_v2_s42_v2); d44d169e8c2a9bcb (1 runs: cwm3d_v2_nospatial_batonpre_pretrain_s123).
K-v2 programme — CockpitWM-K improvement, pre-registered 2026-09-14 (protocol C23) — val n = 405 (selection), test n = 609 (one evaluation per run), paired per-clip bootstrap (10000)
Generated 20260915T013147Z by aide/eval/report_kv2_program.py from pred_{val,test}.npz / metrics_{val,test}.json / run json / aide/results/kv2/*.json only (locked protocol: driver first, grid floor, val HM top 40 (threshold 83.33 px, NEVER a ranking statistic), test HM top 60 (threshold 72.05 px), PCK @ 96 px; every run selected on val All-MPJPE of the EMA weights only, test evaluated once by the harness). Δ = paired per-clip difference (negative = better for errors, positive = better for PCK), CI = clip-bootstrap percentile interval. Verdict rule (as arch_attribution.md): BEATS / WORSE = Δ CI excl. 0 on ≥ ceil(2n/3) seeds (5 seeds: 4) with the same sign on every seed; NOT_WORSE = all CIs cover 0 and |Δ| < 0.25; else INCONCLUSIVE. Confirmatory cells need all 5 seeds; otherwise "pending".
0. Pre-registration (verbatim decisions row) and timestamp check
- S0 UTC: 2026-09-14T22:51:07Z; earliest programme run
t_start_utc: 2026-09-14T23:08:22Z (42 programme runs finished of 189 possible names); runs started at or before S0: none.
| 2026-09-14 | K-v2 programme — S0 pre-registration (CockpitWM-K improvement; new model file aide/models/cwm3d_v2.py, --model cwm3d_v2; approved plan <local-notes>/plans/elegant-dazzling-stallman.md Part A; row written at S0 UTC 2026-09-14T22:51:07Z BEFORE any K-v2 / MB-grid run; the same UTC is in aide/results/kv2/S0_UTC.txt and protocol C23). Honesty clause: the K-v2 design (hierarchical part trunk kept; joint tokens + 2 spatial blocks + attention pooling; lean shared read-out; onset-hazard head gating the anchor mixture) was motivated by the TEST diagnostics of arch_attribution.md §3.2 (K vs MotionBERT-kp: right hand +0.71, body +0.28, h=2–4 +0.2–0.3, HM h=3/4 +2.4/+2.0) → therefore ALL K-v2 configuration choices are fixed on VAL only, and MotionBERT-kp (MB) receives the same val selection budget (S3, 10 val points each). Locked protocol unchanged: val-only selection on All-MPJPE of the EMA weights; the official test split is evaluated exactly once per run by the existing harness (append-only results.csv); no flip TTA; ensembles only as separately labelled rows; Stage-A recipe verbatim (--lr 3e-4 --wd 0.05 --warmup 10 --epochs 200 --bs 32 --ema 0.999 --clip 1.0 --patience 40 --aug-window 1 --noise-px 2.0 --joint-drop 0.1 --translate 20 --scale 0.05 --flip-p 0; K loss defaults euc + 0.1 NLL + 0.05 bone + 0.01 jerk; tensor-output models add --loss euc); paired per-clip clip-bootstrap 10,000 resamples seed 0 (aide/eval/metrics.paired_diff); verdict rule of arch_attribution.md: BEATS / WORSE = Δ CI excl. 0 on ≥ ceil(2n/3) seeds (5 seeds: 4/5) with the same sign on every seed; NOT_WORSE = all CIs cover 0 and |Δ| < 0.25; else INCONCLUSIVE. Every decision below is computed mechanically by aide/eval/kv2_gate.py from pred_val.npz + harness.load_split("val") (+ metrics.hm_mask on val) and written as UTC-stamped json in aide/results/kv2/ (s1_decision.json, s1b_<cfg>_decision.json, s3_decision.json, pretrain_check_<cfg>.json); downstream runs may start only after the corresponding json exists. *S1 val screen (seed 42, n = 405)**: 9 configs vs the existing cwm3d_kp_s42_v2 (val All 55.367): cwm3d_v2_s42_v2 (full), cwm3d_v2_nospatial_s42_v2 (--no-v2-spatial), _khead (--no-lean-head, K's per-part read-out), _nohaz (--no-hazard), _hazaux (--haz-mod 0: hazard BCE on, gating off), _plain (--no-physics --w-nll 0), _q060 / _q090 (--haz-quantile 0.60 / 0.90), _lam03 (--haz-weight 0.3); statistic = paired per-clip Δ val All (config − K); SURVIVE iff mean Δ ≤ −0.30 px (≈ 1.4 val seed-sd); at most 2 survivors (ranked by mean Δ; candidates within 0.10 px of each other are ties → the one with FEWER components wins, components = enabled optional modules of {spatial blocks, lean head, hazard BCE, hazard gating, physics+NLL} — full / q060 / q090 / lam03 = 5, nospatial / khead / hazaux / plain = 4, nohaz = 3; remaining ties by name); HM guard (never a selection statistic): KILL iff Δ val HM (40 clips) mean > +3.0 px AND its CI excludes 0. The 40-clip val HM subset is never used for ranking (level CI width ≈ 37 px vs 8.3 for val All; seed sd 0.93 vs 0.21; val HM threshold 83.3 px ≠ test 72.05 px). Test numbers of S1 runs are produced once by the trainer as usual and are RECORDED, never used for any decision. S1b: each survivor at seed 123 vs cwm3d_kp_s123_v2; CONTINUE to seeds 456 / 789 / 1011 iff Δ val(123) < 0 AND the two-seed mean ≤ −0.25 px; otherwise stop and report "val-killed at seed 123". **S2 (5 seeds 42 / 123 / 456 / 789 / 1011), without and with BATON trunk-only init**: K-v2's tokenizer / trunk tensors differ from K's → own pretraining cwm3d_v2_batonpre_pretrain_s123 (pretrain_cwm3d_baton.py --model cwm3d_v2 --seed 123: same corpus sha256 465dac1cf0a8265a, bs 256, 6 epochs, held-out-route selection, --zeroshot 0 = no zero-shot test evaluation) UNLESS kv2_gate.py pretrain-check <cfg> (dry train_kp.load_init_ckpt(parts="trunk") of the survivor model on CPU against aide/results/runs/cwm3d_kp_batonpre_pretrain_s123/ckpt_best.pt) reports n_skipped_shape = 0 AND zero missing TRUNK_PREFIXES tensors, in which case K's checkpoint is reused; fine-tunes cwm3d_v2_<cfg>_batonpre_trunkp123_s<k>_v2 (--init-parts trunk); K's own +pre arm is completed to 5 seeds with cwm3d_kp_batonpre_trunkp123_s{789,1011}_v2 (recipe and init of the existing s42/123/456 runs verbatim). **S3 MB fairness arm**: (i) same-budget val grid at seed 42: motionbert_kp_{d192, L4, lr1e-4, lr1e-3, wd0.01, wd0.1, do0.1, d192_L4, d192_L4_do0.1}_s42_v2 (--model motionbert_kp[_d192][_L4] name-encoded width / depth, --lr, --wd, --dropout flags; --loss euc) + the existing default motionbert_kp_s42_v2 = 10 val points = S1's 10; **MB-best = argmin val All** (ties within 0.10 px → fewer parameters); if MB-best ≠ default it is run at the 5 seeds; (ii) MB-best + BATON trunk-only init at 5 seeds: motionbert_kp_batonpre_trunkp123_s{789,1011}_v2 are run UNCONDITIONALLY (completing the attribution table); if MB-best changes lr / wd / dropout only, the existing motionbert_kp_batonpre_pretrain_s123 checkpoint is reused (motionbert_kp_<cfg>_batonpre_trunkp123_s<k>_v2); if MB-best changes d or L, a new pretraining motionbert_kp_<tag>_batonpre_pretrain_s123 (same corpus / schedule) precedes 5 fine-tunes. **K-v2(best)** = the S2 survivor with the lowest mean VAL All over its finished seeds (ties 0.10 → fewer components); with one survivor it is that one. **PRIMARY (two co-primary endpoints, test All-MPJPE, 5 seeds, both sides in the same pretraining state, beats rule above)**: P1 = K-v2(best) − MB-best (no pretraining); P2 = K-v2(best)+pre − MB-best+pre. **ADOPTION**: A1 = K-v2(best) − cwm3d_kp_s<k>_v2 BEATS on All over 5 seeds → K-v2 becomes the main-table CockpitWM-K row (K's row itself unchanged), otherwise K-v2 is an ablation row; A2 = K-v2(best)+pre − K+pre (cwm3d_kp_batonpre_trunkp123_s<k>_v2) BEATS → the "+ BATON trunk" row. **SECONDARY (descriptive with CIs, never an adoption basis)**: d-n, HM, PCK@0.05, HM h=5, per-part, per-horizon, AUC of the hazard probability haz_P against HM membership. **Paper-wording consequence table** (mirrors the 2026-09-14 attribution row): P1 ∧ P2 BEATS → an architecture advantage under the keypoint formulation may be written FOR K-v2 (K's own attribution sentence unchanged); P1 BEATS ∧ ¬P2 → "advantage without pretraining that the BATON trunk equalises", attribution sentence kept; ¬P1 ∧ P2 → "advantage conditional on the pretrained trunk", attribution sentence kept; neither → wording unchanged verbatim, K-v2 = ablation family + gen_tables.py NEGATIVES entry; any WORSE is reported as is. **Multiplicity**: ≤ 9 (S1) + 18 (S2) + 2 (K+pre seeds) + 9 (S3 i) + 4 (MB-best seeds) + 7 (S3 ii) ≈ 49 new single test evaluations, each required by the protocol, none used for selection; confirmatory claims are exactly P1, P2, A1, A2. **Already published numbers untouched**; all results only in the new table aide/results/tables/kv2_program.md (+ .json) written by aide/eval/report_kv2_program.py (reads only pred_{val,test}.npz, metrics_{val,test}.json, run json; append-only; machine-readable line **ADOPTION A1: ADOPTED\|NOT ADOPTED**). **Execution**: aide/train/run_kv2_program.sh lanes A (S1 → gate s1), B (S3 i → gate s3 → MB-best seeds / +pre), C (motionbert_kp_batonpre_trunkp123_s{789,1011}_v2, cwm3d_kp_batonpre_trunkp123_s{789,1011}_v2), D (pretrain-check → v2 pretraining if needed → S1b → gate s1b → seeds ± pre); every launch passes the project gate (free VRAM ≥ need + 4 GB, MemAvailable ≥ 28 GB, `pgrep -fc "python.(train_cwm3|train_kp|pretrain)"< 4, OMP/MKL 4;setsid nohup bash; a run whose run json exists is never re-run; pids in aide/results/kv2/pids.txt); --no-lock --max-minutes 45, --notes "kv2_program;;cfg="; --refs zvshrink=zv_shrink_val,simlpe=simlpe_kp_s<k or 42>_v2,kp=cwm3d_kp_s_v2,mb=motionbert_kp_s_v2(+seedinit=_s_v2for +pre runs; refs are evaluation-only). Lanes A / D start only afteraide/models/cwm3d_v2.pyexists andpytest -q aide/tests/test_cwm3d_v2.pypasses. Honest expectation (recorded): HM likely improves, All likely still ties MB; every result enters the table regardless. | plan Part A §A.2–A.4;aide/eval/kv2_gate.py, aide/eval/report_kv2_program.py, aide/train/run_kv2_program.sh, aide/tests/test_kv2.py, aide/models/motionbert_kp.py::build(name variants),aide/docs/protocol.mdC23,aide/results/kv2/S0_UTC.txt; motivation aide/results/tables/arch_attribution.md§3.2 | Every K-v2 / MB-grid run json must havet_start_utclater than 2026-09-14T22:51:07Z (checked byreport_kv2_program.py` §0); gate json must predate the downstream launches; "S1 outcome" and final "K-v2 outcome" rows + C23 outcome follow; paper text changes only after the outcome, per the consequence table |
1. S1 val screen (seed 42; statistic = paired per-clip Δ val All vs cwm3d_kp_s42_v2; SURVIVE iff mean Δ ≤ −0.30 px; ≤ 2 survivors, ties 0.10 → fewer components; HM guard KILL iff Δ val HM > +3.0 with CI excl. 0 — the val HM column is SECONDARY)
- Decision source:
aide/results/kv2/s1_decision.json written 2026-09-14T23:45:48Z (complete = True). Rule: paired per-clip Δ val All (config − cwm3d_kp_s42_v2), n = 405, 10000 clip-bootstrap seed 0; SURVIVE iff mean Δ ≤ -0.3 px; ≤ 2 survivors ranked by mean Δ, ties within 0.1 px -> fewer components (full/q060/q090/lam03 = 5, nospatial/khead/hazaux/plain = 4, nohaz = 3); HM guard (never a ranking statistic): KILL iff Δ val HM (40 clips) mean > +3.0 AND CI excl. 0.
| cfg |
run |
flags |
params |
best ep / epochs / stop / GPU min |
val All |
Δ val All [CI] |
Δ val HM [CI] (secondary) |
guard |
S1 status |
test All (recorded, not used) |
| K (baseline) |
cwm3d_kp_s42_v2 |
— |
5771449 |
58 / 99 / early_stop(patience=40) / 0.6 |
55.367 |
0 |
0 |
— |
— |
48.767 |
| full |
cwm3d_v2_s42_v2 |
(full) |
2736282 |
75 / 116 / early_stop(patience=40) / 2.25 |
55.170 |
-0.197 [-0.798,+0.387] |
+0.60 [-2.54,+3.45] |
ok |
KILL |
48.696 |
| nospatial |
cwm3d_v2_nospatial_s42_v2 |
--no-v2-spatial |
2666906 |
87 / 128 / early_stop(patience=40) / 1.69 |
54.788 |
-0.579 [-1.275,+0.058] |
-0.64 [-4.19,+1.99] |
ok |
SURVIVE |
48.267 |
| khead |
cwm3d_v2_khead_s42_v2 |
--no-lean-head |
5091964 |
52 / 93 / early_stop(patience=40) / 4.75 |
55.482 |
+0.115 [-0.442,+0.662] |
+1.06 [-1.16,+3.09] |
ok |
KILL |
48.642 |
| nohaz |
cwm3d_v2_nohaz_s42_v2 |
--no-hazard |
2702199 |
106 / 147 / early_stop(patience=40) / 5.71 |
55.325 |
-0.042 [-0.873,+0.723] |
+2.94 [-0.38,+6.86] |
ok |
KILL |
48.648 |
| hazaux |
cwm3d_v2_hazaux_s42_v2 |
--haz-mod 0 |
2736267 |
87 / 128 / early_stop(patience=40) / 4.51 |
55.140 |
-0.227 [-0.888,+0.393] |
+2.41 [+0.04,+4.95] |
ok |
KILL |
48.592 |
| plain |
cwm3d_v2_plain_s42_v2 |
--no-physics --w-nll 0 |
2736282 |
65 / 106 / early_stop(patience=40) / 3.69 |
55.218 |
-0.149 [-0.722,+0.397] |
+1.73 [-0.72,+4.18] |
ok |
KILL |
48.728 |
| q060 |
cwm3d_v2_q060_s42_v2 |
--haz-quantile 0.60 |
2736282 |
76 / 117 / early_stop(patience=40) / 4.32 |
55.320 |
-0.047 [-0.674,+0.606] |
+3.26 [+0.21,+7.50] |
KILL |
KILL |
48.704 |
| q090 |
cwm3d_v2_q090_s42_v2 |
--haz-quantile 0.90 |
2736282 |
71 / 112 / early_stop(patience=40) / 4.19 |
55.128 |
-0.239 [-0.774,+0.291] |
+1.33 [-1.27,+4.02] |
ok |
KILL |
48.380 |
| lam03 |
cwm3d_v2_lam03_s42_v2 |
--haz-weight 0.3 |
2736282 |
87 / 128 / early_stop(patience=40) / 5.51 |
55.019 |
-0.348 [-1.018,+0.310] |
+1.92 [-0.73,+5.20] |
ok |
SURVIVE |
48.512 |
S1 survivors (decision): nospatial, lam03.
1.1 S1b (survivors at seed 123 vs cwm3d_kp_s123_v2; CONTINUE iff Δ val(123) < 0 and 2-seed mean ≤ −0.25)
| cfg |
seed-123 run |
val All |
Δ val All s42 [CI] |
Δ val All s123 [CI] |
2-seed mean |
decision |
source |
| nospatial |
cwm3d_v2_nospatial_s123_v2 |
54.801 |
-0.579 [-1.275,+0.058] |
-0.566 [-1.121,-0.038] |
-0.573 |
CONTINUE |
decision json 2026-09-15T00:00:38Z |
| lam03 |
cwm3d_v2_lam03_s123_v2 |
55.297 |
-0.348 [-1.018,+0.310] |
-0.070 [-0.678,+0.534] |
-0.209 |
KILLED |
decision json 2026-09-15T00:00:39Z |
S1b: continuing to 5 seeds: nospatial.
2. S2 — 5 seeds (42 / 123 / 456 / 789 / 1011), TEST (one evaluation per run), K-v2 survivors without / with BATON trunk-only init
- Pretraining
cwm3d_v2_batonpre_pretrain_s123 (architecture-matched trunk for lam03): not run / not finished.
- Pretraining
cwm3d_v2_nospatial_batonpre_pretrain_s123 (architecture-matched trunk for nospatial): finished (6 epochs, best epoch 5, held-out All 27.18, 10.72 GPU-min, corpus sha256 465dac1cf0a8265a, command aide/train/pretrain_cwm3d_baton.py --model cwm3d_v2 --no-v2-spatial --seed 123 --name cwm3d_v2_nospatial_batonpre_pretrain_s123 --zeroshot 0).
- pretrain-check
nospatial: vs K checkpoint → NEED_V2_PRETRAIN (loaded 47 / 54 trunk tensors, shape-skipped 0, missing 7); vs own pretraining → trunk_complete = True (loaded 54 / 54, shape-skipped 0, missing 0).
- pretrain-check
lam03: vs K checkpoint → no json (loaded — / — trunk tensors, shape-skipped —, missing —); vs own pretraining → no json.
2.1 nospatial (--no-v2-spatial) — S1b: CONTINUE
| seed |
arm |
run |
params |
best ep / epochs / stop |
GPU min |
val All |
All [CI] |
d-n [CI] |
HM [CI] |
PCK@0.05 [CI] |
HM h=5 [CI] |
| 42 |
K-v2 |
cwm3d_v2_nospatial_s42_v2 |
2666906 |
87 / 128 / early_stop(patience=40) |
1.69 |
54.788 |
48.27 [44.54, 52.40] |
2.191 [2.022, 2.379] |
128.74 [106.37, 157.12] |
87.10 [85.94, 88.24] |
159.70 [116.06, 218.87] |
| 42 |
K-v2 + pre |
cwm3d_v2_nospatial_batonpre_trunkp123_s42_v2 |
2666906 |
44 / 85 / early_stop(patience=40) |
1.21 |
54.689 |
47.98 [44.26, 52.04] |
2.178 [2.009, 2.363] |
128.58 [106.36, 156.90] |
87.20 [86.08, 88.32] |
159.20 [115.08, 219.21] |
| 123 |
K-v2 |
cwm3d_v2_nospatial_s123_v2 |
2666906 |
84 / 125 / early_stop(patience=40) |
3.71 |
54.801 |
48.53 [44.75, 52.63] |
2.203 [2.031, 2.389] |
129.11 [106.68, 157.62] |
86.95 [85.78, 88.12] |
161.40 [117.80, 220.27] |
| 123 |
K-v2 + pre |
cwm3d_v2_nospatial_batonpre_trunkp123_s123_v2 |
2666906 |
42 / 83 / early_stop(patience=40) |
1.09 |
54.321 |
47.95 [44.22, 52.05] |
2.177 [2.007, 2.363] |
129.25 [106.85, 157.68] |
87.29 [86.15, 88.40] |
160.85 [116.52, 221.12] |
| 456 |
K-v2 |
cwm3d_v2_nospatial_s456_v2 |
2666906 |
91 / 132 / early_stop(patience=40) |
3.67 |
55.150 |
48.53 [44.76, 52.59] |
2.203 [2.032, 2.388] |
129.92 [107.92, 158.13] |
87.01 [85.86, 88.16] |
160.73 [117.34, 219.34] |
| 456 |
K-v2 + pre |
cwm3d_v2_nospatial_batonpre_trunkp123_s456_v2 |
2666906 |
44 / 85 / early_stop(patience=40) |
1.12 |
54.352 |
48.04 [44.34, 52.10] |
2.181 [2.013, 2.365] |
128.45 [106.26, 156.89] |
87.09 [85.95, 88.21] |
159.24 [115.23, 219.46] |
| 789 |
K-v2 |
cwm3d_v2_nospatial_s789_v2 |
2666906 |
80 / 121 / early_stop(patience=40) |
1.65 |
55.048 |
48.44 [44.65, 52.56] |
2.199 [2.027, 2.386] |
130.96 [108.49, 159.39] |
86.96 [85.82, 88.10] |
161.72 [117.96, 220.66] |
| 789 |
K-v2 + pre |
cwm3d_v2_nospatial_batonpre_trunkp123_s789_v2 |
2666906 |
40 / 81 / early_stop(patience=40) |
1.08 |
54.710 |
47.81 [44.12, 51.85] |
2.170 [2.003, 2.354] |
128.48 [106.16, 157.11] |
87.25 [86.14, 88.36] |
159.53 [115.32, 219.61] |
| 1011 |
K-v2 |
cwm3d_v2_nospatial_s1011_v2 |
2666906 |
81 / 122 / early_stop(patience=40) |
1.58 |
54.744 |
48.30 [44.53, 52.43] |
2.192 [2.021, 2.380] |
131.03 [108.45, 159.36] |
87.09 [85.94, 88.23] |
164.10 [120.54, 222.22] |
| 1011 |
K-v2 + pre |
cwm3d_v2_nospatial_batonpre_trunkp123_s1011_v2 |
2666906 |
44 / 85 / early_stop(patience=40) |
1.12 |
54.559 |
48.10 [44.37, 52.18] |
2.183 [2.014, 2.369] |
129.87 [107.84, 158.26] |
87.28 [86.16, 88.39] |
161.58 [118.12, 220.56] |
Δ = K-v2 − K, per seed (n = 5 seed pairs finished; n_req = ceil(2n/3) = 4)
| cell |
A s42 |
A s123 |
A s456 |
A s789 |
A s1011 |
B s42 |
B s123 |
B s456 |
B s789 |
B s1011 |
Δ s42 [CI] |
Δ s123 [CI] |
Δ s456 [CI] |
Δ s789 [CI] |
Δ s1011 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
48.27 |
48.53 |
48.53 |
48.44 |
48.30 |
48.77 |
48.66 |
48.46 |
48.54 |
48.57 |
-0.50 [-1.00,-0.01] |
-0.13 [-0.69,+0.33] |
+0.07 [-0.42,+0.59] |
-0.10 [-0.44,+0.24] |
-0.28 [-0.60,+0.05] |
-0.19 ± 0.21 |
INCONCLUSIVE |
| d-n |
2.191 |
2.203 |
2.203 |
2.199 |
2.192 |
2.214 |
2.209 |
2.200 |
2.204 |
2.205 |
-0.023 [-0.045,-0.001] |
-0.006 [-0.031,+0.015] |
+0.003 [-0.019,+0.027] |
-0.005 [-0.020,+0.011] |
-0.013 [-0.027,+0.002] |
-0.009 ± 0.010 |
INCONCLUSIVE |
| HM-MPJPE |
128.74 |
129.11 |
129.92 |
130.96 |
131.03 |
132.49 |
131.05 |
131.37 |
131.73 |
131.22 |
-3.75 [-5.86,-1.76] |
-1.94 [-3.80,-0.25] |
-1.45 [-3.86,+0.83] |
-0.77 [-2.46,+1.04] |
-0.19 [-2.01,+1.73] |
-1.62 ± 1.36 |
INCONCLUSIVE |
| PCK@0.05 |
87.10 |
86.95 |
87.01 |
86.96 |
87.09 |
87.00 |
86.99 |
87.13 |
87.00 |
87.00 |
+0.10 [-0.17,+0.36] |
-0.04 [-0.34,+0.28] |
-0.11 [-0.42,+0.16] |
-0.04 [-0.26,+0.17] |
+0.09 [-0.10,+0.29] |
-0.00 ± 0.09 |
NOT_WORSE |
| HM h=5 |
159.70 |
161.40 |
160.73 |
161.72 |
164.10 |
164.19 |
163.36 |
162.55 |
163.04 |
163.70 |
-4.49 [-9.02,-0.43] |
-1.96 [-4.68,+0.59] |
-1.82 [-6.12,+2.26] |
-1.33 [-4.21,+1.76] |
+0.39 [-2.73,+3.48] |
-1.84 ± 1.75 |
INCONCLUSIVE |
| All body |
70.64 |
70.83 |
70.85 |
70.96 |
70.73 |
71.39 |
71.29 |
71.02 |
71.05 |
71.03 |
-0.75 [-1.33,-0.20] |
-0.45 [-1.15,+0.14] |
-0.18 [-0.74,+0.48] |
-0.09 [-0.57,+0.39] |
-0.30 [-0.78,+0.17] |
-0.35 ± 0.26 |
INCONCLUSIVE |
| All face |
25.58 |
25.80 |
25.58 |
25.75 |
25.76 |
25.64 |
25.86 |
25.57 |
25.62 |
25.75 |
-0.07 [-0.60,+0.46] |
-0.06 [-0.60,+0.47] |
+0.01 [-0.55,+0.59] |
+0.14 [-0.28,+0.58] |
+0.01 [-0.37,+0.39] |
+0.01 ± 0.08 |
NOT_WORSE |
| All lhand |
64.79 |
65.01 |
65.17 |
64.71 |
64.37 |
65.67 |
65.11 |
65.03 |
65.05 |
65.30 |
-0.89 [-2.10,+0.31] |
-0.10 [-1.17,+0.93] |
+0.14 [-0.88,+1.21] |
-0.34 [-1.20,+0.51] |
-0.93 [-1.84,-0.06] |
-0.42 ± 0.47 |
INCONCLUSIVE |
| All rhand |
77.51 |
78.01 |
78.57 |
77.77 |
77.40 |
78.72 |
78.03 |
78.09 |
78.41 |
77.94 |
-1.22 [-2.47,+0.02] |
-0.02 [-1.20,+1.11] |
+0.48 [-0.86,+1.86] |
-0.65 [-1.58,+0.31] |
-0.54 [-1.30,+0.23] |
-0.39 ± 0.65 |
INCONCLUSIVE |
| All h=1 |
34.88 |
34.80 |
34.89 |
34.86 |
34.77 |
35.08 |
35.03 |
35.07 |
35.00 |
35.14 |
-0.20 [-0.82,+0.48] |
-0.24 [-0.81,+0.21] |
-0.18 [-0.57,+0.20] |
-0.15 [-0.54,+0.27] |
-0.37 [-0.88,+0.04] |
-0.23 ± 0.09 |
INCONCLUSIVE |
| All h=2 |
44.18 |
44.31 |
44.52 |
44.47 |
44.25 |
44.84 |
44.66 |
44.50 |
44.55 |
44.60 |
-0.66 [-1.44,+0.01] |
-0.35 [-0.97,+0.16] |
+0.02 [-0.48,+0.53] |
-0.09 [-0.52,+0.34] |
-0.36 [-0.71,-0.00] |
-0.29 ± 0.27 |
INCONCLUSIVE |
| All h=3 |
51.58 |
51.88 |
51.93 |
51.77 |
51.55 |
52.24 |
52.04 |
51.87 |
51.93 |
51.85 |
-0.66 [-1.39,+0.01] |
-0.16 [-0.81,+0.44] |
+0.06 [-0.62,+0.79] |
-0.16 [-0.65,+0.32] |
-0.30 [-0.76,+0.17] |
-0.24 ± 0.27 |
INCONCLUSIVE |
| All h=4 |
53.26 |
53.69 |
53.54 |
53.55 |
53.40 |
53.91 |
53.76 |
53.50 |
53.68 |
53.60 |
-0.65 [-1.36,+0.02] |
-0.07 [-0.73,+0.55] |
+0.04 [-0.70,+0.83] |
-0.13 [-0.64,+0.39] |
-0.20 [-0.65,+0.26] |
-0.20 ± 0.26 |
INCONCLUSIVE |
| All h=5 |
57.43 |
57.96 |
57.79 |
57.58 |
57.50 |
57.77 |
57.81 |
57.38 |
57.55 |
57.67 |
-0.34 [-1.05,+0.33] |
+0.15 [-0.46,+0.75] |
+0.41 [-0.37,+1.28] |
+0.02 [-0.51,+0.52] |
-0.17 [-0.62,+0.29] |
+0.02 ± 0.29 |
INCONCLUSIVE |
Δ = K-v2+pre − K+pre, per seed (n = 5 seed pairs finished; n_req = ceil(2n/3) = 4)
| cell |
A s42 |
A s123 |
A s456 |
A s789 |
A s1011 |
B s42 |
B s123 |
B s456 |
B s789 |
B s1011 |
Δ s42 [CI] |
Δ s123 [CI] |
Δ s456 [CI] |
Δ s789 [CI] |
Δ s1011 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
47.98 |
47.95 |
48.04 |
47.81 |
48.10 |
47.98 |
48.04 |
48.01 |
48.12 |
48.11 |
+0.00 [-0.37,+0.36] |
-0.09 [-0.47,+0.27] |
+0.03 [-0.40,+0.45] |
-0.31 [-0.72,+0.07] |
-0.02 [-0.37,+0.32] |
-0.08 ± 0.14 |
INCONCLUSIVE |
| d-n |
2.178 |
2.177 |
2.181 |
2.170 |
2.183 |
2.178 |
2.181 |
2.179 |
2.184 |
2.184 |
+0.000 [-0.017,+0.016] |
-0.004 [-0.022,+0.012] |
+0.001 [-0.018,+0.020] |
-0.014 [-0.033,+0.003] |
-0.001 [-0.017,+0.015] |
-0.003 ± 0.006 |
NOT_WORSE |
| HM-MPJPE |
128.58 |
129.25 |
128.45 |
128.48 |
129.87 |
128.43 |
129.13 |
128.51 |
129.32 |
128.73 |
+0.15 [-2.22,+2.25] |
+0.12 [-1.63,+1.72] |
-0.06 [-2.48,+1.92] |
-0.84 [-2.64,+0.85] |
+1.14 [-0.67,+2.94] |
+0.10 ± 0.71 |
INCONCLUSIVE |
| PCK@0.05 |
87.20 |
87.29 |
87.09 |
87.25 |
87.28 |
87.21 |
87.20 |
87.16 |
87.06 |
87.03 |
-0.00 [-0.21,+0.22] |
+0.08 [-0.10,+0.28] |
-0.07 [-0.34,+0.23] |
+0.19 [-0.06,+0.46] |
+0.24 [+0.02,+0.50] |
+0.09 ± 0.13 |
INCONCLUSIVE |
| HM h=5 |
159.20 |
160.85 |
159.24 |
159.53 |
161.58 |
159.35 |
160.11 |
159.02 |
161.39 |
159.72 |
-0.15 [-4.17,+3.78] |
+0.74 [-2.12,+3.40] |
+0.22 [-4.10,+4.02] |
-1.86 [-5.12,+1.28] |
+1.86 [-2.13,+5.96] |
+0.16 ± 1.36 |
INCONCLUSIVE |
| All body |
70.48 |
70.17 |
70.36 |
70.27 |
70.26 |
70.64 |
70.68 |
70.78 |
70.48 |
70.37 |
-0.16 [-0.58,+0.27] |
-0.50 [-1.04,-0.02] |
-0.41 [-0.89,+0.05] |
-0.21 [-0.68,+0.24] |
-0.11 [-0.49,+0.28] |
-0.28 ± 0.17 |
INCONCLUSIVE |
| All face |
25.45 |
25.64 |
25.52 |
25.28 |
25.66 |
25.35 |
25.40 |
25.28 |
25.47 |
25.37 |
+0.09 [-0.33,+0.49] |
+0.24 [-0.13,+0.60] |
+0.24 [-0.29,+0.76] |
-0.19 [-0.67,+0.23] |
+0.28 [-0.11,+0.65] |
+0.13 ± 0.20 |
INCONCLUSIVE |
| All lhand |
63.92 |
63.86 |
63.41 |
63.85 |
63.42 |
64.02 |
63.91 |
64.19 |
64.16 |
64.39 |
-0.10 [-1.13,+0.95] |
-0.04 [-1.02,+0.97] |
-0.78 [-1.81,+0.23] |
-0.30 [-1.39,+0.77] |
-0.98 [-2.02,+0.04] |
-0.44 ± 0.42 |
INCONCLUSIVE |
| All rhand |
77.14 |
76.77 |
77.93 |
76.92 |
78.00 |
77.14 |
77.47 |
77.21 |
77.73 |
77.91 |
+0.00 [-0.98,+0.98] |
-0.70 [-1.73,+0.30] |
+0.72 [-0.29,+1.74] |
-0.81 [-1.85,+0.20] |
+0.09 [-0.96,+1.14] |
-0.14 ± 0.63 |
INCONCLUSIVE |
| All h=1 |
34.69 |
34.59 |
34.78 |
34.56 |
34.79 |
34.53 |
34.34 |
34.69 |
34.56 |
34.62 |
+0.15 [-0.23,+0.53] |
+0.25 [-0.14,+0.66] |
+0.10 [-0.31,+0.48] |
+0.00 [-0.36,+0.34] |
+0.17 [-0.18,+0.52] |
+0.13 ± 0.09 |
NOT_WORSE |
| All h=2 |
43.96 |
44.01 |
44.06 |
43.76 |
44.04 |
44.11 |
44.08 |
44.13 |
44.10 |
44.20 |
-0.15 [-0.62,+0.33] |
-0.07 [-0.58,+0.40] |
-0.06 [-0.59,+0.44] |
-0.35 [-0.79,+0.09] |
-0.15 [-0.60,+0.28] |
-0.16 ± 0.12 |
INCONCLUSIVE |
| All h=3 |
51.22 |
51.14 |
51.28 |
51.05 |
51.43 |
51.31 |
51.50 |
51.36 |
51.42 |
51.54 |
-0.09 [-0.55,+0.37] |
-0.36 [-0.88,+0.11] |
-0.08 [-0.64,+0.46] |
-0.36 [-0.87,+0.12] |
-0.10 [-0.59,+0.37] |
-0.20 ± 0.15 |
INCONCLUSIVE |
| All h=4 |
53.06 |
52.95 |
53.05 |
52.88 |
53.12 |
53.02 |
53.19 |
53.05 |
53.22 |
53.22 |
+0.03 [-0.50,+0.56] |
-0.24 [-0.81,+0.28] |
-0.01 [-0.62,+0.60] |
-0.34 [-0.94,+0.23] |
-0.09 [-0.66,+0.47] |
-0.13 ± 0.16 |
INCONCLUSIVE |
| All h=5 |
56.97 |
57.07 |
57.01 |
56.79 |
57.09 |
56.92 |
57.10 |
56.81 |
57.28 |
56.99 |
+0.05 [-0.51,+0.62] |
-0.03 [-0.60,+0.50] |
+0.20 [-0.44,+0.83] |
-0.49 [-1.15,+0.12] |
+0.10 [-0.50,+0.70] |
-0.04 ± 0.27 |
INCONCLUSIVE |
Δ = K-v2 − MB-best[do0.1], per seed (n = 5 seed pairs finished; n_req = ceil(2n/3) = 4)
| cell |
A s42 |
A s123 |
A s456 |
A s789 |
A s1011 |
B s42 |
B s123 |
B s456 |
B s789 |
B s1011 |
Δ s42 [CI] |
Δ s123 [CI] |
Δ s456 [CI] |
Δ s789 [CI] |
Δ s1011 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
48.27 |
48.53 |
48.53 |
48.44 |
48.30 |
48.39 |
48.41 |
48.21 |
48.37 |
48.92 |
-0.12 [-0.75,+0.49] |
+0.12 [-0.31,+0.54] |
+0.32 [-0.26,+0.92] |
+0.07 [-0.43,+0.59] |
-0.62 [-1.24,-0.03] |
-0.05 ± 0.36 |
INCONCLUSIVE |
| d-n |
2.191 |
2.203 |
2.203 |
2.199 |
2.192 |
2.197 |
2.197 |
2.188 |
2.196 |
2.221 |
-0.006 [-0.034,+0.022] |
+0.005 [-0.014,+0.025] |
+0.015 [-0.012,+0.042] |
+0.003 [-0.020,+0.027] |
-0.028 [-0.056,-0.001] |
-0.002 ± 0.016 |
INCONCLUSIVE |
| HM-MPJPE |
128.74 |
129.11 |
129.92 |
130.96 |
131.03 |
130.48 |
129.51 |
128.99 |
130.78 |
130.61 |
-1.74 [-3.99,+0.38] |
-0.40 [-2.37,+1.65] |
+0.93 [-1.89,+3.74] |
+0.18 [-2.95,+3.59] |
+0.43 [-2.06,+2.97] |
-0.12 ± 1.02 |
INCONCLUSIVE |
| PCK@0.05 |
87.10 |
86.95 |
87.01 |
86.96 |
87.09 |
87.26 |
87.11 |
87.30 |
87.20 |
86.90 |
-0.16 [-0.51,+0.16] |
-0.16 [-0.47,+0.12] |
-0.28 [-0.62,+0.04] |
-0.24 [-0.56,+0.06] |
+0.20 [-0.08,+0.48] |
-0.13 ± 0.19 |
INCONCLUSIVE |
| HM h=5 |
159.70 |
161.40 |
160.73 |
161.72 |
164.10 |
163.78 |
162.75 |
162.24 |
164.33 |
163.65 |
-4.08 [-8.62,+0.25] |
-1.34 [-5.07,+2.43] |
-1.51 [-6.34,+3.12] |
-2.62 [-7.44,+2.14] |
+0.44 [-4.18,+4.84] |
-1.82 ± 1.67 |
INCONCLUSIVE |
| All body |
70.64 |
70.83 |
70.85 |
70.96 |
70.73 |
70.88 |
70.94 |
70.85 |
70.68 |
71.10 |
-0.23 [-0.86,+0.38] |
-0.10 [-0.79,+0.53] |
+0.00 [-0.73,+0.75] |
+0.28 [-0.33,+0.89] |
-0.37 [-1.03,+0.28] |
-0.09 ± 0.25 |
INCONCLUSIVE |
| All face |
25.58 |
25.80 |
25.58 |
25.75 |
25.76 |
25.39 |
25.70 |
25.53 |
25.45 |
25.92 |
+0.19 [-0.43,+0.87] |
+0.10 [-0.33,+0.57] |
+0.06 [-0.50,+0.66] |
+0.31 [-0.32,+1.00] |
-0.16 [-0.77,+0.47] |
+0.10 ± 0.17 |
INCONCLUSIVE |
| All lhand |
64.79 |
65.01 |
65.17 |
64.71 |
64.37 |
65.53 |
65.27 |
65.04 |
65.36 |
66.48 |
-0.74 [-2.22,+0.64] |
-0.26 [-1.39,+0.82] |
+0.13 [-1.25,+1.42] |
-0.65 [-2.00,+0.58] |
-2.11 [-4.05,-0.42] |
-0.73 ± 0.85 |
INCONCLUSIVE |
| All rhand |
77.51 |
78.01 |
78.57 |
77.77 |
77.40 |
77.89 |
77.19 |
76.79 |
77.97 |
78.36 |
-0.38 [-1.90,+1.02] |
+0.81 [-0.36,+1.98] |
+1.78 [+0.27,+3.40] |
-0.20 [-1.44,+1.03] |
-0.96 [-2.40,+0.47] |
+0.21 ± 1.09 |
INCONCLUSIVE |
| All h=1 |
34.88 |
34.80 |
34.89 |
34.86 |
34.77 |
34.97 |
34.99 |
34.94 |
35.20 |
35.23 |
-0.09 [-0.55,+0.36] |
-0.19 [-0.63,+0.19] |
-0.05 [-0.55,+0.42] |
-0.35 [-0.89,+0.19] |
-0.45 [-1.04,+0.10] |
-0.23 ± 0.17 |
INCONCLUSIVE |
| All h=2 |
44.18 |
44.31 |
44.52 |
44.47 |
44.25 |
44.28 |
44.37 |
44.14 |
44.29 |
44.95 |
-0.10 [-0.77,+0.56] |
-0.07 [-0.57,+0.42] |
+0.38 [-0.18,+0.94] |
+0.17 [-0.41,+0.80] |
-0.71 [-1.40,-0.04] |
-0.06 ± 0.41 |
INCONCLUSIVE |
| All h=3 |
51.58 |
51.88 |
51.93 |
51.77 |
51.55 |
51.69 |
51.85 |
51.49 |
51.80 |
52.39 |
-0.11 [-0.89,+0.66] |
+0.03 [-0.59,+0.64] |
+0.43 [-0.33,+1.23] |
-0.03 [-0.74,+0.70] |
-0.84 [-1.65,-0.03] |
-0.10 ± 0.46 |
INCONCLUSIVE |
| All h=4 |
53.26 |
53.69 |
53.54 |
53.55 |
53.40 |
53.38 |
53.41 |
53.00 |
53.10 |
53.86 |
-0.12 [-0.98,+0.70] |
+0.28 [-0.34,+0.90] |
+0.54 [-0.36,+1.46] |
+0.44 [-0.31,+1.22] |
-0.46 [-1.26,+0.32] |
+0.14 ± 0.42 |
INCONCLUSIVE |
| All h=5 |
57.43 |
57.96 |
57.79 |
57.58 |
57.50 |
57.63 |
57.43 |
57.46 |
57.45 |
58.16 |
-0.19 [-1.05,+0.65] |
+0.53 [-0.09,+1.15] |
+0.32 [-0.55,+1.27] |
+0.12 [-0.60,+0.84] |
-0.66 [-1.50,+0.16] |
+0.03 ± 0.46 |
INCONCLUSIVE |
Δ = K-v2+pre − MB-best[do0.1]+pre, per seed (n = 5 seed pairs finished; n_req = ceil(2n/3) = 4)
| cell |
A s42 |
A s123 |
A s456 |
A s789 |
A s1011 |
B s42 |
B s123 |
B s456 |
B s789 |
B s1011 |
Δ s42 [CI] |
Δ s123 [CI] |
Δ s456 [CI] |
Δ s789 [CI] |
Δ s1011 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
47.98 |
47.95 |
48.04 |
47.81 |
48.10 |
47.78 |
47.44 |
47.74 |
47.74 |
47.77 |
+0.20 [-0.34,+0.72] |
+0.51 [-0.00,+1.05] |
+0.30 [-0.21,+0.82] |
+0.07 [-0.42,+0.55] |
+0.33 [-0.24,+0.90] |
+0.28 ± 0.16 |
INCONCLUSIVE |
| d-n |
2.178 |
2.177 |
2.181 |
2.170 |
2.183 |
2.169 |
2.153 |
2.167 |
2.167 |
2.168 |
+0.009 [-0.015,+0.033] |
+0.023 [-0.000,+0.048] |
+0.014 [-0.010,+0.037] |
+0.003 [-0.019,+0.025] |
+0.015 [-0.011,+0.041] |
+0.013 ± 0.007 |
NOT_WORSE |
| HM-MPJPE |
128.58 |
129.25 |
128.45 |
128.48 |
129.87 |
128.16 |
127.07 |
127.75 |
127.88 |
128.48 |
+0.42 [-1.80,+2.63] |
+2.18 [-0.47,+5.11] |
+0.70 [-1.92,+3.42] |
+0.60 [-1.34,+2.58] |
+1.39 [-1.55,+4.38] |
+1.06 ± 0.73 |
INCONCLUSIVE |
| PCK@0.05 |
87.20 |
87.29 |
87.09 |
87.25 |
87.28 |
87.45 |
87.45 |
87.38 |
87.36 |
87.41 |
-0.24 [-0.54,+0.05] |
-0.17 [-0.47,+0.12] |
-0.29 [-0.61,+0.02] |
-0.11 [-0.37,+0.17] |
-0.13 [-0.45,+0.17] |
-0.19 ± 0.08 |
INCONCLUSIVE |
| HM h=5 |
159.20 |
160.85 |
159.24 |
159.53 |
161.58 |
159.24 |
157.09 |
158.75 |
158.93 |
159.83 |
-0.05 [-3.75,+3.54] |
+3.76 [-0.73,+8.20] |
+0.48 [-3.58,+4.53] |
+0.60 [-3.47,+4.54] |
+1.75 [-3.19,+6.91] |
+1.31 ± 1.52 |
INCONCLUSIVE |
| All body |
70.48 |
70.17 |
70.36 |
70.27 |
70.26 |
70.16 |
69.94 |
70.19 |
70.01 |
69.98 |
+0.32 [-0.28,+0.94] |
+0.24 [-0.39,+0.85] |
+0.18 [-0.47,+0.83] |
+0.26 [-0.36,+0.85] |
+0.28 [-0.39,+0.95] |
+0.25 ± 0.05 |
INCONCLUSIVE |
| All face |
25.45 |
25.64 |
25.52 |
25.28 |
25.66 |
25.44 |
25.26 |
25.40 |
25.38 |
25.28 |
+0.01 [-0.48,+0.47] |
+0.37 [-0.16,+0.94] |
+0.12 [-0.46,+0.73] |
-0.11 [-0.52,+0.29] |
+0.37 [-0.22,+0.99] |
+0.15 ± 0.22 |
INCONCLUSIVE |
| All lhand |
63.92 |
63.86 |
63.41 |
63.85 |
63.42 |
63.71 |
62.90 |
63.24 |
63.58 |
63.62 |
+0.21 [-1.11,+1.52] |
+0.96 [-0.26,+2.19] |
+0.17 [-1.19,+1.52] |
+0.27 [-0.97,+1.56] |
-0.20 [-1.61,+1.14] |
+0.28 ± 0.42 |
INCONCLUSIVE |
| All rhand |
77.14 |
76.77 |
77.93 |
76.92 |
78.00 |
76.50 |
75.91 |
76.78 |
76.69 |
77.22 |
+0.64 [-0.96,+2.24] |
+0.86 [-0.68,+2.38] |
+1.15 [-0.24,+2.51] |
+0.23 [-1.31,+1.71] |
+0.78 [-1.02,+2.52] |
+0.73 ± 0.34 |
INCONCLUSIVE |
| All h=1 |
34.69 |
34.59 |
34.78 |
34.56 |
34.79 |
34.53 |
34.00 |
34.63 |
34.52 |
34.48 |
+0.16 [-0.41,+0.69] |
+0.59 [+0.07,+1.10] |
+0.16 [-0.34,+0.63] |
+0.04 [-0.47,+0.51] |
+0.32 [-0.27,+0.87] |
+0.25 ± 0.21 |
INCONCLUSIVE |
| All h=2 |
43.96 |
44.01 |
44.06 |
43.76 |
44.04 |
43.93 |
43.58 |
43.88 |
43.96 |
43.95 |
+0.04 [-0.58,+0.67] |
+0.43 [-0.15,+1.02] |
+0.19 [-0.45,+0.83] |
-0.21 [-0.82,+0.38] |
+0.09 [-0.61,+0.79] |
+0.11 ± 0.23 |
INCONCLUSIVE |
| All h=3 |
51.22 |
51.14 |
51.28 |
51.05 |
51.43 |
51.06 |
51.02 |
51.34 |
51.20 |
51.42 |
+0.16 [-0.45,+0.79] |
+0.12 [-0.59,+0.85] |
-0.06 [-0.76,+0.67] |
-0.15 [-0.83,+0.53] |
+0.02 [-0.77,+0.80] |
+0.02 ± 0.13 |
NOT_WORSE |
| All h=4 |
53.06 |
52.95 |
53.05 |
52.88 |
53.12 |
52.52 |
52.25 |
52.24 |
52.36 |
52.38 |
+0.53 [-0.16,+1.23] |
+0.70 [+0.02,+1.38] |
+0.81 [+0.10,+1.53] |
+0.52 [-0.11,+1.15] |
+0.74 [-0.02,+1.54] |
+0.66 ± 0.13 |
INCONCLUSIVE |
| All h=5 |
56.97 |
57.07 |
57.01 |
56.79 |
57.09 |
56.88 |
56.33 |
56.61 |
56.62 |
56.61 |
+0.10 [-0.67,+0.84] |
+0.74 [-0.04,+1.52] |
+0.40 [-0.30,+1.11] |
+0.17 [-0.52,+0.84] |
+0.48 [-0.29,+1.26] |
+0.38 ± 0.26 |
INCONCLUSIVE |
Secondary: AUC of the hazard probability haz_P (max / mean over parts × horizons) for HM membership (descriptive)
| seed |
split |
AUC(max haz_P) |
AUC(mean haz_P) |
| 42 |
val |
0.812 |
0.832 |
| 42 |
test |
0.831 |
0.845 |
| 123 |
val |
0.793 |
0.796 |
| 123 |
test |
0.827 |
0.835 |
| 456 |
val |
0.779 |
0.785 |
| 456 |
test |
0.830 |
0.845 |
| 789 |
val |
0.796 |
0.801 |
| 789 |
test |
0.827 |
0.837 |
| 1011 |
val |
0.797 |
0.801 |
| 1011 |
test |
0.812 |
0.829 |
2.2 lam03 (--haz-weight 0.3) — S1b: KILLED → S2 NOT RUN by rule (rows below marked pending were never launched; the seed-123 run is the S1b run, shown for the record)
| seed |
arm |
run |
params |
best ep / epochs / stop |
GPU min |
val All |
All [CI] |
d-n [CI] |
HM [CI] |
PCK@0.05 [CI] |
HM h=5 [CI] |
| 42 |
K-v2 |
cwm3d_v2_lam03_s42_v2 |
2736282 |
87 / 128 / early_stop(patience=40) |
5.51 |
55.019 |
48.51 [44.76, 52.63] |
2.202 [2.032, 2.389] |
130.39 [108.20, 158.69] |
87.00 [85.84, 88.14] |
161.80 [118.48, 220.73] |
| 42 |
K-v2 + pre |
cwm3d_v2_lam03_batonpre_trunkp123_s42_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 123 |
K-v2 |
cwm3d_v2_lam03_s123_v2 |
2736282 |
76 / 117 / early_stop(patience=40) |
4.47 |
55.297 |
48.53 [44.78, 52.64] |
2.203 [2.033, 2.389] |
130.48 [108.05, 158.75] |
87.07 [85.94, 88.19] |
162.10 [118.09, 221.88] |
| 123 |
K-v2 + pre |
cwm3d_v2_lam03_batonpre_trunkp123_s123_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 456 |
K-v2 |
cwm3d_v2_lam03_s456_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 456 |
K-v2 + pre |
cwm3d_v2_lam03_batonpre_trunkp123_s456_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 789 |
K-v2 |
cwm3d_v2_lam03_s789_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 789 |
K-v2 + pre |
cwm3d_v2_lam03_batonpre_trunkp123_s789_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 1011 |
K-v2 |
cwm3d_v2_lam03_s1011_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
| 1011 |
K-v2 + pre |
cwm3d_v2_lam03_batonpre_trunkp123_s1011_v2 |
— |
pending |
— |
— |
— |
— |
— |
— |
— |
Δ = K-v2 − K, per seed (n = 2 seed pairs finished; n_req = ceil(2n/3) = 2)
| cell |
A s42 |
A s123 |
B s42 |
B s123 |
Δ s42 [CI] |
Δ s123 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
48.51 |
48.53 |
48.77 |
48.66 |
-0.26 [-0.72,+0.19] |
-0.13 [-0.71,+0.35] |
-0.19 ± 0.09 |
INCONCLUSIVE |
| d-n |
2.202 |
2.203 |
2.214 |
2.209 |
-0.012 [-0.032,+0.009] |
-0.006 [-0.032,+0.016] |
-0.009 ± 0.004 |
NOT_WORSE |
| HM-MPJPE |
130.39 |
130.48 |
132.49 |
131.05 |
-2.10 [-4.70,+0.41] |
-0.57 [-2.38,+1.33] |
-1.34 ± 1.09 |
INCONCLUSIVE |
| PCK@0.05 |
87.00 |
87.07 |
87.00 |
86.99 |
+0.00 [-0.20,+0.20] |
+0.08 [-0.18,+0.42] |
+0.04 ± 0.06 |
NOT_WORSE |
| HM h=5 |
161.80 |
162.10 |
164.19 |
163.36 |
-2.39 [-6.42,+1.45] |
-1.26 [-4.39,+1.79] |
-1.83 ± 0.80 |
INCONCLUSIVE |
| All body |
70.83 |
70.75 |
71.39 |
71.29 |
-0.56 [-1.16,+0.03] |
-0.54 [-1.23,+0.03] |
-0.55 ± 0.02 |
INCONCLUSIVE |
| All face |
25.54 |
25.61 |
25.64 |
25.86 |
-0.10 [-0.58,+0.34] |
-0.25 [-0.78,+0.20] |
-0.18 ± 0.10 |
INCONCLUSIVE |
| All lhand |
65.34 |
65.44 |
65.67 |
65.11 |
-0.33 [-1.62,+0.91] |
+0.34 [-0.79,+1.52] |
+0.00 ± 0.47 |
INCONCLUSIVE |
| All rhand |
78.43 |
78.34 |
78.72 |
78.03 |
-0.29 [-1.47,+0.96] |
+0.31 [-0.83,+1.43] |
+0.01 ± 0.42 |
INCONCLUSIVE |
| All h=1 |
34.87 |
35.11 |
35.08 |
35.03 |
-0.21 [-0.79,+0.37] |
+0.08 [-0.54,+0.59] |
-0.06 ± 0.20 |
NOT_WORSE |
| All h=2 |
44.63 |
44.53 |
44.84 |
44.66 |
-0.21 [-0.91,+0.42] |
-0.12 [-0.82,+0.46] |
-0.17 ± 0.06 |
NOT_WORSE |
| All h=3 |
51.92 |
51.97 |
52.24 |
52.04 |
-0.32 [-1.02,+0.33] |
-0.06 [-0.77,+0.56] |
-0.19 ± 0.18 |
INCONCLUSIVE |
| All h=4 |
53.62 |
53.56 |
53.91 |
53.76 |
-0.29 [-0.90,+0.31] |
-0.20 [-0.90,+0.44] |
-0.24 ± 0.06 |
INCONCLUSIVE |
| All h=5 |
57.52 |
57.48 |
57.77 |
57.81 |
-0.25 [-0.86,+0.33] |
-0.33 [-0.98,+0.26] |
-0.29 ± 0.06 |
INCONCLUSIVE |
Δ = K-v2+pre − K+pre, per seed (n = 0 seed pairs finished; n_req = ceil(2n/3) = —)
(pending — no finished seed pair)
Δ = K-v2 − MB-best[do0.1], per seed (n = 2 seed pairs finished; n_req = ceil(2n/3) = 2)
| cell |
A s42 |
A s123 |
B s42 |
B s123 |
Δ s42 [CI] |
Δ s123 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
48.51 |
48.53 |
48.39 |
48.41 |
+0.12 [-0.47,+0.68] |
+0.12 [-0.25,+0.48] |
+0.12 ± 0.00 |
NOT_WORSE |
| d-n |
2.202 |
2.203 |
2.197 |
2.197 |
+0.006 [-0.022,+0.031] |
+0.006 [-0.011,+0.022] |
+0.006 ± 0.000 |
NOT_WORSE |
| HM-MPJPE |
130.39 |
130.48 |
130.48 |
129.51 |
-0.09 [-2.76,+2.71] |
+0.98 [-0.50,+2.61] |
+0.44 ± 0.76 |
INCONCLUSIVE |
| PCK@0.05 |
87.00 |
87.07 |
87.26 |
87.11 |
-0.25 [-0.55,+0.02] |
-0.04 [-0.27,+0.21] |
-0.15 ± 0.15 |
INCONCLUSIVE |
| HM h=5 |
161.80 |
162.10 |
163.78 |
162.75 |
-1.98 [-7.13,+3.15] |
-0.65 [-3.66,+2.07] |
-1.31 ± 0.94 |
INCONCLUSIVE |
| All body |
70.83 |
70.75 |
70.88 |
70.94 |
-0.04 [-0.60,+0.49] |
-0.19 [-0.80,+0.35] |
-0.12 ± 0.10 |
NOT_WORSE |
| All face |
25.54 |
25.61 |
25.39 |
25.70 |
+0.15 [-0.52,+0.86] |
-0.09 [-0.52,+0.35] |
+0.03 ± 0.17 |
NOT_WORSE |
| All lhand |
65.34 |
65.44 |
65.53 |
65.27 |
-0.19 [-1.37,+0.87] |
+0.17 [-0.70,+1.05] |
-0.01 ± 0.26 |
NOT_WORSE |
| All rhand |
78.43 |
78.34 |
77.89 |
77.19 |
+0.54 [-0.66,+1.70] |
+1.15 [+0.10,+2.24] |
+0.85 ± 0.43 |
INCONCLUSIVE |
| All h=1 |
34.87 |
35.11 |
34.97 |
34.99 |
-0.10 [-0.61,+0.38] |
+0.13 [-0.34,+0.51] |
+0.02 ± 0.16 |
NOT_WORSE |
| All h=2 |
44.63 |
44.53 |
44.28 |
44.37 |
+0.35 [-0.32,+0.97] |
+0.16 [-0.34,+0.64] |
+0.25 ± 0.13 |
INCONCLUSIVE |
| All h=3 |
51.92 |
51.97 |
51.69 |
51.85 |
+0.23 [-0.55,+0.97] |
+0.13 [-0.40,+0.65] |
+0.18 ± 0.07 |
NOT_WORSE |
| All h=4 |
53.62 |
53.56 |
53.38 |
53.41 |
+0.24 [-0.61,+1.05] |
+0.16 [-0.33,+0.65] |
+0.20 ± 0.06 |
NOT_WORSE |
| All h=5 |
57.52 |
57.48 |
57.63 |
57.43 |
-0.11 [-0.96,+0.70] |
+0.04 [-0.41,+0.50] |
-0.03 ± 0.11 |
NOT_WORSE |
Δ = K-v2+pre − MB-best[do0.1]+pre, per seed (n = 0 seed pairs finished; n_req = ceil(2n/3) = —)
(pending — no finished seed pair)
Secondary: AUC of the hazard probability haz_P (max / mean over parts × horizons) for HM membership (descriptive)
| seed |
split |
AUC(max haz_P) |
AUC(mean haz_P) |
| 42 |
val |
0.800 |
0.798 |
| 42 |
test |
0.831 |
0.835 |
| 123 |
val |
0.792 |
0.800 |
| 123 |
test |
0.824 |
0.823 |
3. S3 — MotionBERT-kp fairness arm (same val budget: 9 new grid points + existing default = 10; MB-best = argmin val All, ties 0.10 → fewer params)
- Decision source:
aide/results/kv2/s3_decision.json written 2026-09-14T23:49:07Z (complete = True). Rule: MB-best = argmin val All over the 10 grid points (9 new + existing default); ties within 0.1 px -> fewer parameters.
| cfg |
run |
model |
flags |
params |
best ep / epochs / stop / GPU min |
val All |
Δ val All vs default [CI] |
test All (recorded, not used) |
S3 |
| default |
motionbert_kp_s42_v2 |
motionbert_kp |
(default) |
813088 |
81 / 122 / early_stop(patience=40) / 4.6 |
55.550 |
+0.000 [+0.000,+0.000] |
48.279 |
|
| d192 |
motionbert_kp_d192_s42_v2 |
motionbert_kp_d192 |
(default) |
1809440 |
64 / 105 / early_stop(patience=40) / 6.14 |
55.559 |
+0.009 [-0.573,+0.575] |
48.595 |
|
| L4 |
motionbert_kp_L4_s42_v2 |
motionbert_kp_L4 |
(default) |
1077664 |
82 / 123 / early_stop(patience=40) / 4.71 |
55.433 |
-0.117 [-0.572,+0.313] |
48.488 |
|
| lr1e-4 |
motionbert_kp_lr1e-4_s42_v2 |
motionbert_kp |
--lr 1e-4 |
813088 |
79 / 120 / early_stop(patience=40) / 3.18 |
55.310 |
-0.240 [-0.773,+0.247] |
48.493 |
|
| lr1e-3 |
motionbert_kp_lr1e-3_s42_v2 |
motionbert_kp |
--lr 1e-3 |
813088 |
60 / 101 / early_stop(patience=40) / 2.75 |
55.281 |
-0.269 [-0.732,+0.131] |
48.394 |
|
| wd0.01 |
motionbert_kp_wd0.01_s42_v2 |
motionbert_kp |
--wd 0.01 |
813088 |
73 / 114 / early_stop(patience=40) / 3.15 |
55.366 |
-0.184 [-0.659,+0.199] |
48.342 |
|
| wd0.1 |
motionbert_kp_wd0.1_s42_v2 |
motionbert_kp |
--wd 0.1 |
813088 |
64 / 105 / early_stop(patience=40) / 2.83 |
55.603 |
+0.053 [-0.324,+0.440] |
48.424 |
|
| do0.1 |
motionbert_kp_do0.1_s42_v2 |
motionbert_kp |
--dropout 0.1 |
813088 |
95 / 136 / early_stop(patience=40) / 3.89 |
55.215 |
-0.335 [-0.888,+0.129] |
48.389 |
MB-best |
| d192_L4 |
motionbert_kp_d192_L4_s42_v2 |
motionbert_kp_d192_L4 |
(default) |
2402912 |
67 / 108 / early_stop(patience=40) / 5.18 |
55.399 |
-0.151 [-0.632,+0.325] |
48.418 |
|
| d192_L4_do0.1 |
motionbert_kp_d192_L4_do0.1_s42_v2 |
motionbert_kp_d192_L4 |
--dropout 0.1 |
2402912 |
55 / 96 / early_stop(patience=40) / 3.82 |
55.425 |
-0.125 [-0.587,+0.327] |
48.477 |
|
MB-best = do0.1 (final; model motionbert_kp, flags ['--dropout', '0.1']; changes d/L: False → existing motionbert_kp_batonpre_pretrain_s123 checkpoint reused).
3.1 MB-best and MB-best + BATON trunk-only init at 5 seeds; K + pre completion (test levels)
| seed |
arm |
run |
params |
best ep / epochs / stop |
GPU min |
val All |
All [CI] |
d-n [CI] |
HM [CI] |
PCK@0.05 [CI] |
HM h=5 [CI] |
| 42 |
MB default |
motionbert_kp_s42_v2 |
813088 |
81 / 122 / early_stop(patience=40) |
4.6 |
55.550 |
48.28 [44.53, 52.37] |
2.192 [2.021, 2.377] |
130.34 [108.20, 158.53] |
87.17 [86.04, 88.28] |
163.95 [119.85, 223.86] |
| 42 |
MB default + pre |
motionbert_kp_batonpre_trunkp123_s42_v2 |
813088 |
37 / 78 / early_stop(patience=40) |
1.94 |
54.528 |
47.74 [44.07, 51.78] |
2.167 [2.000, 2.350] |
128.03 [105.81, 156.47] |
87.46 [86.37, 88.54] |
159.20 [115.09, 219.85] |
| 42 |
MB-best |
motionbert_kp_do0.1_s42_v2 |
813088 |
95 / 136 / early_stop(patience=40) |
3.89 |
55.215 |
48.39 [44.59, 52.49] |
2.197 [2.024, 2.383] |
130.48 [107.84, 158.95] |
87.26 [86.15, 88.36] |
163.78 [118.78, 224.72] |
| 42 |
MB-best + pre |
motionbert_kp_do0.1_batonpre_trunkp123_s42_v2 |
813088 |
41 / 82 / early_stop(patience=40) |
2.59 |
54.575 |
47.78 [44.10, 51.83] |
2.169 [2.002, 2.353] |
128.16 [105.87, 156.59] |
87.45 [86.35, 88.53] |
159.24 [114.93, 219.57] |
| 42 |
K + pre |
cwm3d_kp_batonpre_trunkp123_s42_v2 |
5771449 |
31 / 72 / early_stop(patience=40) |
0.48 |
55.003 |
47.98 [44.28, 52.04] |
2.178 [2.010, 2.363] |
128.43 [106.25, 157.11] |
87.21 [86.10, 88.30] |
159.35 [115.56, 219.46] |
| 123 |
MB default |
motionbert_kp_s123_v2 |
813088 |
58 / 99 / early_stop(patience=40) |
4.36 |
55.627 |
48.38 [44.65, 52.46] |
2.196 [2.027, 2.381] |
129.16 [106.70, 157.58] |
87.14 [86.02, 88.27] |
162.59 [118.87, 222.08] |
| 123 |
MB default + pre |
motionbert_kp_batonpre_trunkp123_s123_v2 |
813088 |
38 / 79 / early_stop(patience=40) |
2.16 |
54.720 |
47.53 [43.87, 51.55] |
2.158 [1.992, 2.340] |
126.11 [103.96, 154.69] |
87.40 [86.32, 88.46] |
156.47 [113.00, 215.99] |
| 123 |
MB-best |
motionbert_kp_do0.1_s123_v2 |
813088 |
54 / 95 / early_stop(patience=40) |
2.1 |
55.656 |
48.41 [44.69, 52.49] |
2.197 [2.029, 2.383] |
129.51 [107.22, 158.01] |
87.11 [85.97, 88.23] |
162.75 [118.82, 222.30] |
| 123 |
MB-best + pre |
motionbert_kp_do0.1_batonpre_trunkp123_s123_v2 |
813088 |
42 / 83 / early_stop(patience=40) |
2.68 |
54.666 |
47.44 [43.78, 51.45] |
2.153 [1.987, 2.336] |
127.07 [105.00, 155.58] |
87.45 [86.36, 88.52] |
157.09 [113.60, 216.60] |
| 123 |
K + pre |
cwm3d_kp_batonpre_trunkp123_s123_v2 |
5771449 |
33 / 74 / early_stop(patience=40) |
0.88 |
54.828 |
48.04 [44.33, 52.11] |
2.181 [2.012, 2.365] |
129.13 [106.80, 157.67] |
87.20 [86.07, 88.32] |
160.11 [116.11, 220.04] |
| 456 |
MB default |
motionbert_kp_s456_v2 |
813088 |
90 / 131 / early_stop(patience=40) |
5.17 |
55.166 |
48.48 [44.74, 52.58] |
2.201 [2.031, 2.387] |
130.23 [107.78, 158.89] |
87.14 [86.00, 88.25] |
162.24 [116.93, 223.50] |
| 456 |
MB default + pre |
motionbert_kp_batonpre_trunkp123_s456_v2 |
813088 |
44 / 85 / early_stop(patience=40) |
1.68 |
54.797 |
47.91 [44.22, 51.93] |
2.175 [2.008, 2.357] |
127.59 [105.28, 156.14] |
87.39 [86.28, 88.47] |
158.92 [114.76, 219.74] |
| 456 |
MB-best |
motionbert_kp_do0.1_s456_v2 |
813088 |
65 / 106 / early_stop(patience=40) |
2.5 |
55.356 |
48.21 [44.50, 52.28] |
2.188 [2.020, 2.373] |
128.99 [106.63, 157.76] |
87.30 [86.18, 88.39] |
162.24 [117.69, 222.93] |
| 456 |
MB-best + pre |
motionbert_kp_do0.1_batonpre_trunkp123_s456_v2 |
813088 |
43 / 84 / early_stop(patience=40) |
2.74 |
54.828 |
47.74 [44.07, 51.79] |
2.167 [2.001, 2.351] |
127.75 [105.54, 156.19] |
87.38 [86.27, 88.46] |
158.75 [114.66, 219.20] |
| 456 |
K + pre |
cwm3d_kp_batonpre_trunkp123_s456_v2 |
5771449 |
36 / 77 / early_stop(patience=40) |
0.91 |
54.670 |
48.01 [44.31, 52.12] |
2.179 [2.011, 2.366] |
128.51 [106.16, 157.20] |
87.16 [86.02, 88.27] |
159.02 [115.09, 219.40] |
| 789 |
MB default |
motionbert_kp_s789_v2 |
813088 |
77 / 118 / early_stop(patience=40) |
6.14 |
55.134 |
48.46 [44.71, 52.57] |
2.200 [2.030, 2.386] |
130.66 [108.19, 159.30] |
87.27 [86.16, 88.36] |
163.04 [118.30, 223.77] |
| 789 |
MB default + pre |
motionbert_kp_batonpre_trunkp123_s789_v2 |
813088 |
37 / 78 / early_stop(patience=40) |
3.45 |
54.882 |
47.55 [43.88, 51.56] |
2.159 [1.992, 2.341] |
127.71 [105.39, 156.61] |
87.40 [86.30, 88.48] |
159.11 [114.88, 219.30] |
| 789 |
MB-best |
motionbert_kp_do0.1_s789_v2 |
813088 |
89 / 130 / early_stop(patience=40) |
3.43 |
55.295 |
48.37 [44.60, 52.47] |
2.196 [2.025, 2.382] |
130.78 [108.51, 159.34] |
87.20 [86.08, 88.30] |
164.33 [119.77, 224.96] |
| 789 |
MB-best + pre |
motionbert_kp_do0.1_batonpre_trunkp123_s789_v2 |
813088 |
40 / 81 / early_stop(patience=40) |
2.5 |
54.786 |
47.74 [44.06, 51.78] |
2.167 [2.000, 2.350] |
127.88 [105.66, 156.44] |
87.36 [86.25, 88.45] |
158.93 [114.68, 219.29] |
| 789 |
K + pre |
cwm3d_kp_batonpre_trunkp123_s789_v2 |
5771449 |
35 / 76 / early_stop(patience=40) |
1.22 |
54.590 |
48.12 [44.44, 52.21] |
2.184 [2.017, 2.370] |
129.32 [107.17, 157.78] |
87.06 [85.94, 88.18] |
161.39 [117.32, 221.77] |
| 1011 |
MB default |
motionbert_kp_s1011_v2 |
813088 |
81 / 122 / early_stop(patience=40) |
5.69 |
55.506 |
48.68 [44.92, 52.82] |
2.210 [2.039, 2.398] |
131.49 [108.92, 160.13] |
86.97 [85.81, 88.12] |
166.22 [121.36, 227.01] |
| 1011 |
MB default + pre |
motionbert_kp_batonpre_trunkp123_s1011_v2 |
813088 |
47 / 88 / early_stop(patience=40) |
3.98 |
54.868 |
47.85 [44.16, 51.87] |
2.172 [2.005, 2.355] |
129.47 [107.17, 157.92] |
87.40 [86.30, 88.49] |
160.75 [116.16, 221.59] |
| 1011 |
MB-best |
motionbert_kp_do0.1_s1011_v2 |
813088 |
138 / 179 / early_stop(patience=40) |
4.66 |
55.253 |
48.92 [45.11, 53.00] |
2.221 [2.048, 2.406] |
130.61 [108.46, 158.67] |
86.90 [85.74, 88.04] |
163.65 [119.68, 223.28] |
| 1011 |
MB-best + pre |
motionbert_kp_do0.1_batonpre_trunkp123_s1011_v2 |
813088 |
59 / 100 / early_stop(patience=40) |
2.71 |
54.566 |
47.77 [44.07, 51.75] |
2.168 [2.001, 2.349] |
128.48 [106.50, 156.68] |
87.41 [86.31, 88.48] |
159.83 [115.53, 220.28] |
| 1011 |
K + pre |
cwm3d_kp_batonpre_trunkp123_s1011_v2 |
5771449 |
34 / 75 / early_stop(patience=40) |
1.14 |
54.609 |
48.11 [44.40, 52.21] |
2.184 [2.016, 2.370] |
128.73 [106.48, 157.39] |
87.03 [85.91, 88.16] |
159.72 [115.63, 220.35] |
Δ = MB-best[do0.1]+pre − MB-best[do0.1], per seed (descriptive; n = 5)
| cell |
A s42 |
A s123 |
A s456 |
A s789 |
A s1011 |
B s42 |
B s123 |
B s456 |
B s789 |
B s1011 |
Δ s42 [CI] |
Δ s123 [CI] |
Δ s456 [CI] |
Δ s789 [CI] |
Δ s1011 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
47.78 |
47.44 |
47.74 |
47.74 |
47.77 |
48.39 |
48.41 |
48.21 |
48.37 |
48.92 |
-0.61 [-1.17,-0.07] |
-0.97 [-1.50,-0.46] |
-0.47 [-0.96,+0.03] |
-0.63 [-1.10,-0.16] |
-1.15 [-1.82,-0.53] |
-0.77 ± 0.28 |
BEATS |
| d-n |
2.169 |
2.153 |
2.167 |
2.167 |
2.168 |
2.197 |
2.197 |
2.188 |
2.196 |
2.221 |
-0.028 [-0.053,-0.003] |
-0.044 [-0.068,-0.021] |
-0.021 [-0.044,+0.001] |
-0.029 [-0.050,-0.007] |
-0.052 [-0.082,-0.024] |
-0.035 ± 0.013 |
BEATS |
| HM-MPJPE |
128.16 |
127.07 |
127.75 |
127.88 |
128.48 |
130.48 |
129.51 |
128.99 |
130.78 |
130.61 |
-2.32 [-4.83,-0.06] |
-2.44 [-4.62,-0.37] |
-1.25 [-3.56,+0.87] |
-2.90 [-5.13,-0.88] |
-2.13 [-4.94,+0.62] |
-2.21 ± 0.61 |
INCONCLUSIVE |
| PCK@0.05 |
87.45 |
87.45 |
87.38 |
87.36 |
87.41 |
87.26 |
87.11 |
87.30 |
87.20 |
86.90 |
+0.19 [-0.09,+0.46] |
+0.35 [+0.08,+0.63] |
+0.08 [-0.16,+0.33] |
+0.15 [-0.06,+0.36] |
+0.51 [+0.19,+0.87] |
+0.26 ± 0.17 |
INCONCLUSIVE |
| HM h=5 |
159.24 |
157.09 |
158.75 |
158.93 |
159.83 |
163.78 |
162.75 |
162.24 |
164.33 |
163.65 |
-4.54 [-8.85,-0.30] |
-5.66 [-10.34,-1.22] |
-3.49 [-8.12,+0.52] |
-5.41 [-9.75,-1.54] |
-3.82 [-8.18,+0.25] |
-4.58 ± 0.95 |
INCONCLUSIVE |
Δ = K+pre − K, per seed (descriptive; n = 5)
| cell |
A s42 |
A s123 |
A s456 |
A s789 |
A s1011 |
B s42 |
B s123 |
B s456 |
B s789 |
B s1011 |
Δ s42 [CI] |
Δ s123 [CI] |
Δ s456 [CI] |
Δ s789 [CI] |
Δ s1011 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
47.98 |
48.04 |
48.01 |
48.12 |
48.11 |
48.77 |
48.66 |
48.46 |
48.54 |
48.57 |
-0.79 [-1.29,-0.30] |
-0.62 [-1.17,-0.11] |
-0.45 [-0.89,-0.02] |
-0.43 [-0.88,+0.03] |
-0.46 [-0.89,-0.02] |
-0.55 ± 0.15 |
BEATS |
| d-n |
2.178 |
2.181 |
2.179 |
2.184 |
2.184 |
2.214 |
2.209 |
2.200 |
2.204 |
2.205 |
-0.036 [-0.059,-0.014] |
-0.028 [-0.053,-0.005] |
-0.021 [-0.041,-0.001] |
-0.019 [-0.040,+0.001] |
-0.021 [-0.041,-0.001] |
-0.025 ± 0.007 |
BEATS |
| HM-MPJPE |
128.43 |
129.13 |
128.51 |
129.32 |
128.73 |
132.49 |
131.05 |
131.37 |
131.73 |
131.22 |
-4.06 [-6.57,-1.74] |
-1.92 [-4.14,+0.22] |
-2.86 [-5.09,-0.73] |
-2.41 [-4.43,-0.48] |
-2.49 [-4.79,-0.27] |
-2.75 ± 0.81 |
BEATS |
| PCK@0.05 |
87.21 |
87.20 |
87.16 |
87.06 |
87.03 |
87.00 |
86.99 |
87.13 |
87.00 |
87.00 |
+0.20 [-0.05,+0.45] |
+0.22 [-0.07,+0.52] |
+0.03 [-0.20,+0.26] |
+0.06 [-0.16,+0.28] |
+0.03 [-0.20,+0.26] |
+0.11 ± 0.09 |
NOT_WORSE |
| HM h=5 |
159.35 |
160.11 |
159.02 |
161.39 |
159.72 |
164.19 |
163.36 |
162.55 |
163.04 |
163.70 |
-4.85 [-8.95,-1.00] |
-3.25 [-6.97,+0.55] |
-3.53 [-7.22,+0.08] |
-1.65 [-5.11,+1.98] |
-3.98 [-8.22,+0.07] |
-3.45 ± 1.17 |
INCONCLUSIVE |
Δ = K − MB-best[do0.1], per seed (descriptive; n = 5)
| cell |
A s42 |
A s123 |
A s456 |
A s789 |
A s1011 |
B s42 |
B s123 |
B s456 |
B s789 |
B s1011 |
Δ s42 [CI] |
Δ s123 [CI] |
Δ s456 [CI] |
Δ s789 [CI] |
Δ s1011 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
48.77 |
48.66 |
48.46 |
48.54 |
48.57 |
48.39 |
48.41 |
48.21 |
48.37 |
48.92 |
+0.38 [-0.21,+0.96] |
+0.25 [-0.15,+0.69] |
+0.25 [-0.19,+0.69] |
+0.17 [-0.29,+0.64] |
-0.34 [-0.98,+0.25] |
+0.14 ± 0.28 |
INCONCLUSIVE |
| d-n |
2.214 |
2.209 |
2.200 |
2.204 |
2.205 |
2.197 |
2.197 |
2.188 |
2.196 |
2.221 |
+0.017 [-0.010,+0.044] |
+0.011 [-0.007,+0.031] |
+0.012 [-0.009,+0.032] |
+0.008 [-0.013,+0.029] |
-0.016 [-0.045,+0.011] |
+0.006 ± 0.013 |
NOT_WORSE |
| HM-MPJPE |
132.49 |
131.05 |
131.37 |
131.73 |
131.22 |
130.48 |
129.51 |
128.99 |
130.78 |
130.61 |
+2.01 [-0.71,+4.80] |
+1.54 [-0.56,+3.72] |
+2.38 [+0.08,+4.80] |
+0.95 [-1.79,+3.81] |
+0.62 [-1.60,+2.73] |
+1.50 ± 0.73 |
INCONCLUSIVE |
| PCK@0.05 |
87.00 |
86.99 |
87.13 |
87.00 |
87.00 |
87.26 |
87.11 |
87.30 |
87.20 |
86.90 |
-0.25 [-0.54,+0.01] |
-0.12 [-0.41,+0.14] |
-0.17 [-0.47,+0.12] |
-0.20 [-0.48,+0.06] |
+0.11 [-0.19,+0.41] |
-0.13 ± 0.14 |
INCONCLUSIVE |
| HM h=5 |
164.19 |
163.36 |
162.55 |
163.04 |
163.70 |
163.78 |
162.75 |
162.24 |
164.33 |
163.65 |
+0.41 [-4.94,+5.58] |
+0.61 [-3.26,+4.35] |
+0.30 [-3.97,+4.60] |
-1.29 [-6.68,+3.41] |
+0.05 [-4.50,+4.01] |
+0.02 ± 0.76 |
INCONCLUSIVE |
Δ = K+pre − MB-best[do0.1]+pre, per seed (descriptive; n = 5)
| cell |
A s42 |
A s123 |
A s456 |
A s789 |
A s1011 |
B s42 |
B s123 |
B s456 |
B s789 |
B s1011 |
Δ s42 [CI] |
Δ s123 [CI] |
Δ s456 [CI] |
Δ s789 [CI] |
Δ s1011 [CI] |
mean Δ ± sd |
verdict |
| All-MPJPE |
47.98 |
48.04 |
48.01 |
48.12 |
48.11 |
47.78 |
47.44 |
47.74 |
47.74 |
47.77 |
+0.20 [-0.29,+0.70] |
+0.60 [+0.11,+1.12] |
+0.27 [-0.24,+0.79] |
+0.38 [-0.12,+0.90] |
+0.35 [-0.24,+0.97] |
+0.36 ± 0.15 |
INCONCLUSIVE |
| d-n |
2.178 |
2.181 |
2.179 |
2.184 |
2.184 |
2.169 |
2.153 |
2.167 |
2.167 |
2.168 |
+0.009 [-0.013,+0.032] |
+0.027 [+0.005,+0.051] |
+0.012 [-0.011,+0.036] |
+0.017 [-0.005,+0.041] |
+0.016 [-0.011,+0.044] |
+0.016 ± 0.007 |
INCONCLUSIVE |
| HM-MPJPE |
128.43 |
129.13 |
128.51 |
129.32 |
128.73 |
128.16 |
127.07 |
127.75 |
127.88 |
128.48 |
+0.27 [-2.42,+3.03] |
+2.06 [-0.67,+4.96] |
+0.76 [-1.81,+3.62] |
+1.44 [-0.64,+3.70] |
+0.25 [-3.14,+3.84] |
+0.96 ± 0.78 |
INCONCLUSIVE |
| PCK@0.05 |
87.21 |
87.20 |
87.16 |
87.06 |
87.03 |
87.45 |
87.45 |
87.38 |
87.36 |
87.41 |
-0.24 [-0.50,+0.02] |
-0.25 [-0.53,+0.01] |
-0.22 [-0.54,+0.07] |
-0.29 [-0.58,-0.02] |
-0.38 [-0.71,-0.07] |
-0.28 ± 0.06 |
INCONCLUSIVE |
| HM h=5 |
159.35 |
160.11 |
159.02 |
161.39 |
159.72 |
159.24 |
157.09 |
158.75 |
158.93 |
159.83 |
+0.10 [-4.83,+4.85] |
+3.02 [-2.27,+8.37] |
+0.26 [-4.85,+5.29] |
+2.46 [-1.86,+6.95] |
-0.11 [-5.85,+5.56] |
+1.15 ± 1.47 |
INCONCLUSIVE |
4. Primary endpoints and adoption (test All-MPJPE, 5 seeds, both sides in the same pretraining state; BEATS rule, 4/5)
- K-v2(best) = nospatial (lowest mean VAL All over finished seeds among S1b-continued survivors; ties 0.10 → fewer components): nospatial val All 54.906 over 5 seeds, 4 components.
| endpoint |
n seeds |
Δ All per seed [CI] |
mean Δ ± sd |
verdict |
secondary (HM / PCK@0.05 / HM h=5 / d-n verdicts) |
| P1 = K-v2(nospatial) − MB-best[do0.1] |
5 |
s42 -0.12 [-0.75,+0.49]; s123 +0.12 [-0.31,+0.54]; s456 +0.32 [-0.26,+0.92]; s789 +0.07 [-0.43,+0.59]; s1011 -0.62 [-1.24,-0.03] |
-0.05 ± 0.36 |
INCONCLUSIVE |
d-n INCONCLUSIVE (-0.00); HM-MPJPE INCONCLUSIVE (-0.12); PCK@0.05 INCONCLUSIVE (-0.13); HM h=5 INCONCLUSIVE (-1.82) |
| P2 = K-v2(nospatial)+pre − MB-best[do0.1]+pre |
5 |
s42 +0.20 [-0.34,+0.72]; s123 +0.51 [-0.00,+1.05]; s456 +0.30 [-0.21,+0.82]; s789 +0.07 [-0.42,+0.55]; s1011 +0.33 [-0.24,+0.90] |
+0.28 ± 0.16 |
INCONCLUSIVE |
d-n NOT_WORSE (+0.01); HM-MPJPE INCONCLUSIVE (+1.06); PCK@0.05 INCONCLUSIVE (-0.19); HM h=5 INCONCLUSIVE (+1.31) |
| A1 = K-v2(nospatial) − K |
5 |
s42 -0.50 [-1.00,-0.01]; s123 -0.13 [-0.69,+0.33]; s456 +0.07 [-0.42,+0.59]; s789 -0.10 [-0.44,+0.24]; s1011 -0.28 [-0.60,+0.05] |
-0.19 ± 0.21 |
INCONCLUSIVE |
d-n INCONCLUSIVE (-0.01); HM-MPJPE INCONCLUSIVE (-1.62); PCK@0.05 NOT_WORSE (-0.00); HM h=5 INCONCLUSIVE (-1.84) |
| A2 = K-v2(nospatial)+pre − K+pre |
5 |
s42 +0.00 [-0.37,+0.36]; s123 -0.09 [-0.47,+0.27]; s456 +0.03 [-0.40,+0.45]; s789 -0.31 [-0.72,+0.07]; s1011 -0.02 [-0.37,+0.32] |
-0.08 ± 0.14 |
INCONCLUSIVE |
d-n NOT_WORSE (-0.00); HM-MPJPE INCONCLUSIVE (+0.10); PCK@0.05 INCONCLUSIVE (+0.09); HM h=5 INCONCLUSIVE (+0.16) |
4.1 Paper-wording consequence table (pre-registered; the realised row is marked)
| case |
wording consequence |
realised |
| P1 ∧ P2 |
an architecture advantage under the keypoint formulation may be written FOR K-v2 (K's own attribution sentence unchanged) |
|
| P1 ∧ ¬P2 |
"advantage without pretraining that the BATON trunk equalises"; attribution sentence kept |
|
| ¬P1 ∧ P2 |
"advantage conditional on the pretrained trunk"; attribution sentence kept |
|
| neither |
wording unchanged verbatim; K-v2 = ablation family + gen_tables.py NEGATIVES entry; any WORSE reported as is |
← REALISED |
ADOPTION A1: NOT ADOPTED
ADOPTION A2: NOT ADOPTED
- Consequence-table row: neither. Any WORSE verdict above is reported as is.
5. GPU minutes of the programme's new runs (from run json; the GPU was shared with a DMS feature-extraction job and other lanes, so minutes are wall-clock upper bounds)
| run |
GPU min |
| cwm3d_kp_batonpre_trunkp123_s1011_v2 |
1.14 |
| cwm3d_kp_batonpre_trunkp123_s789_v2 |
1.22 |
| cwm3d_v2_hazaux_s42_v2 |
4.51 |
| cwm3d_v2_khead_s42_v2 |
4.75 |
| cwm3d_v2_lam03_s123_v2 |
4.47 |
| cwm3d_v2_lam03_s42_v2 |
5.51 |
| cwm3d_v2_nohaz_s42_v2 |
5.71 |
| cwm3d_v2_nospatial_batonpre_pretrain_s123 |
10.72 |
| cwm3d_v2_nospatial_batonpre_trunkp123_s1011_v2 |
1.12 |
| cwm3d_v2_nospatial_batonpre_trunkp123_s123_v2 |
1.09 |
| cwm3d_v2_nospatial_batonpre_trunkp123_s42_v2 |
1.21 |
| cwm3d_v2_nospatial_batonpre_trunkp123_s456_v2 |
1.12 |
| cwm3d_v2_nospatial_batonpre_trunkp123_s789_v2 |
1.08 |
| cwm3d_v2_nospatial_s1011_v2 |
1.58 |
| cwm3d_v2_nospatial_s123_v2 |
3.71 |
| cwm3d_v2_nospatial_s42_v2 |
1.69 |
| cwm3d_v2_nospatial_s456_v2 |
3.67 |
| cwm3d_v2_nospatial_s789_v2 |
1.65 |
| cwm3d_v2_plain_s42_v2 |
3.69 |
| cwm3d_v2_q060_s42_v2 |
4.32 |
| cwm3d_v2_q090_s42_v2 |
4.19 |
| cwm3d_v2_s42_v2 |
2.25 |
| motionbert_kp_L4_s42_v2 |
4.71 |
| motionbert_kp_batonpre_trunkp123_s1011_v2 |
3.98 |
| motionbert_kp_batonpre_trunkp123_s789_v2 |
3.45 |
| motionbert_kp_d192_L4_do0.1_s42_v2 |
3.82 |
| motionbert_kp_d192_L4_s42_v2 |
5.18 |
| motionbert_kp_d192_s42_v2 |
6.14 |
| motionbert_kp_do0.1_batonpre_trunkp123_s1011_v2 |
2.71 |
| motionbert_kp_do0.1_batonpre_trunkp123_s123_v2 |
2.68 |
| motionbert_kp_do0.1_batonpre_trunkp123_s42_v2 |
2.59 |
| motionbert_kp_do0.1_batonpre_trunkp123_s456_v2 |
2.74 |
| motionbert_kp_do0.1_batonpre_trunkp123_s789_v2 |
2.50 |
| motionbert_kp_do0.1_s1011_v2 |
4.66 |
| motionbert_kp_do0.1_s123_v2 |
2.10 |
| motionbert_kp_do0.1_s42_v2 |
3.89 |
| motionbert_kp_do0.1_s456_v2 |
2.50 |
| motionbert_kp_do0.1_s789_v2 |
3.43 |
| motionbert_kp_lr1e-3_s42_v2 |
2.75 |
| motionbert_kp_lr1e-4_s42_v2 |
3.18 |
| motionbert_kp_wd0.01_s42_v2 |
3.15 |
| motionbert_kp_wd0.1_s42_v2 |
2.83 |
| total (42 runs) |
139.4 |
6. Provenance
- Inputs:
pred_val.npz / pred_test.npz / metrics_val.json / metrics_test.json / run_<name>_<utc>.json under /home/USER/cwm3/aide/results/runs; gate decisions /home/USER/cwm3/aide/results/kv2: pretrain_check_nospatial.json, pretrain_check_nospatial_20260915T000040Z.json, pretrain_check_nospatial_own.json, pretrain_check_nospatial_own_20260915T001217Z.json, s1_decision.json, s1_decision_20260914T234548Z.json, s1b_lam03_decision.json, s1b_lam03_decision_20260915T000039Z.json, s1b_nospatial_decision.json, s1b_nospatial_decision_20260915T000038Z.json, s3_decision.json, s3_decision_20260914T230840Z.json, s3_decision_20260914T234907Z.json; pre-registration row from /home/USER/cwm3/docs/decisions.md (S0 UTC 2026-09-14T22:51:07Z).
- Comparison runs
cwm3d_kp_s<k>_v2, cwm3d_kp_batonpre_trunkp123_s{42,123,456}_v2, motionbert_kp_s<k>_v2, motionbert_kp_batonpre_trunkp123_s{42,123,456}_v2 are the EXISTING runs (never retrained); every programme run is a new run dir.
- Selection: val All-MPJPE of the EMA weights inside the trainer; S1 / S1b / S3 decisions from pred_val.npz only (
aide/eval/kv2_gate.py); one test evaluation per run by the harness; no test number was used for any decision.
- Launcher
aide/train/run_kv2_program.sh (gate: free VRAM ≥ need + 4 GB, MemAvailable ≥ 28 GB, < 4 trainers, OMP/MKL 4; finished runs skipped); pids aide/results/kv2/pids.txt.
- Pending runs referenced above: 0. This block is append-only; earlier blocks are never edited.
- Trainer code hashes (
code_sha256 of each finished programme run's run json; distinct values = code changed between launches, e.g. the train_kp.py:630 guard fix of 2026-09-14T23:11:53Z, behaviour-neutral for dict-output models — see aide/results/kv2/NOTES.md): e313c0ff89746293 (20 runs: motionbert_kp_L4_s42_v2, motionbert_kp_batonpre_trunkp123_s1011_v2, motionbert_kp_batonpre_trunkp123_s789_v2, motionbert_kp_d192_L4_do0.1_s42_v2, motionbert_kp_d192_L4_s42_v2, motionbert_kp_d192_s42_v2, motionbert_kp_do0.1_batonpre_trunkp123_s1011_v2, motionbert_kp_do0.1_batonpre_trunkp123_s123_v2, motionbert_kp_do0.1_batonpre_trunkp123_s42_v2, motionbert_kp_do0.1_batonpre_trunkp123_s456_v2, motionbert_kp_do0.1_batonpre_trunkp123_s789_v2, motionbert_kp_do0.1_s1011_v2, motionbert_kp_do0.1_s123_v2, motionbert_kp_do0.1_s42_v2, motionbert_kp_do0.1_s456_v2, motionbert_kp_do0.1_s789_v2, motionbert_kp_lr1e-3_s42_v2, motionbert_kp_lr1e-4_s42_v2, motionbert_kp_wd0.01_s42_v2, motionbert_kp_wd0.1_s42_v2); 8caba4a8efccd311 (18 runs: cwm3d_v2_hazaux_s42_v2, cwm3d_v2_khead_s42_v2, cwm3d_v2_lam03_s123_v2, cwm3d_v2_lam03_s42_v2, cwm3d_v2_nohaz_s42_v2, cwm3d_v2_nospatial_batonpre_trunkp123_s1011_v2, cwm3d_v2_nospatial_batonpre_trunkp123_s123_v2, cwm3d_v2_nospatial_batonpre_trunkp123_s42_v2, cwm3d_v2_nospatial_batonpre_trunkp123_s456_v2, cwm3d_v2_nospatial_batonpre_trunkp123_s789_v2, cwm3d_v2_nospatial_s1011_v2, cwm3d_v2_nospatial_s123_v2, cwm3d_v2_nospatial_s42_v2, cwm3d_v2_nospatial_s456_v2, cwm3d_v2_nospatial_s789_v2, cwm3d_v2_plain_s42_v2, cwm3d_v2_q060_s42_v2, cwm3d_v2_q090_s42_v2); 22306515fd4a42a0 (2 runs: cwm3d_kp_batonpre_trunkp123_s1011_v2, cwm3d_kp_batonpre_trunkp123_s789_v2); 652c4ba49ba937f2 (1 runs: cwm3d_v2_s42_v2); d44d169e8c2a9bcb (1 runs: cwm3d_v2_nospatial_batonpre_pretrain_s123).
