Confirmation update · 6 September 2026
All three relative-progress confirmation seeds pass the allocation gate. The fixed four-arm campaign will test whether that exposure improves held-out tracking.
All U/A/R/D runs for seeds 21 and 22, plus R23 and D23, have completed 4,000 iterations and passed their declared training gates. Every saved sampler state replays exactly, with zero recorded invalid-start, invalid-reference or censored-reset events. A is retained under its frozen comparator rule even when allocation contrast is weak.
Static snapshot · 6 September 2026, 22:17 EDT. U23 is running, with checkpoint 1100 saved; A23 is queued. The scheduler and follow-on workers remain active. No held-out evaluation cell has started. Both D calibration seeds and all four seed-51 confirmation-entrypoint smokes passed before the immutable freeze. The original overnight attempt timed out waiting for GPU capacity before any scientific job; its record is preserved. The active operational re-entry uses the same scientific contract and schedule.
Each completed run below has 512 environments, 4,000 PPO iterations and 41 verified saved states. Mean TV averages the 37 saved states at iterations 400–3999. Incomplete U23/A23 runs are excluded from the measured table and figure; no value is imputed.
| Arm and seed | Mean TV | Completed trials | Decision |
|---|---|---|---|
| U · seed 21 | 0.000000000 | 1,086,702 | Training gate pass |
| U · seed 22 | 0.000000000 | 1,086,071 | Training gate pass |
| A · seed 21 | 0.029759170 | 1,084,430 | Training gate pass |
| A · seed 22 | 0.029338594 | 1,082,658 | Training gate pass |
| R · seed 21 | 0.083484949 | 1,083,411 | Training gate pass |
| R · seed 22 | 0.082491997 | 1,087,418 | Training gate pass |
| R · seed 23 | 0.082332844 | 1,089,129 | Training gate pass |
| D · seed 21 | 0.083910689 | 1,116,479 | Training gate pass |
| D · seed 22 | 0.085865093 | 1,116,999 | Training gate pass |
| D · seed 23 | 0.083983853 | 1,117,687 | Training gate pass |
R passes in all three confirmation seeds (mean TV 0.083485, 0.082492, 0.082333); D also passes in all three. A's completed runs remain near 0.029–0.030. D begins near the prior and increases its allocation contrast later; R changes exposure earlier and follows a different trajectory. This reproduces the exposure distinction that motivated relative normalization. It does not show that R selects learnable motions better than D or that either improves tracking.
Development evidence · before confirmationBoth seed-1 training arms finished, but absolute-progress sampling reached mean post-warm-up TV of only 0.029659, below the frozen 0.05 minimum. Policy endpoints stayed closed and seeds 2–3 stopped. This is an inadequate intervention, not a measured policy null.
The fixed-policy comparison finished on 26 clips and 656 paired conditions per reference arm. Repaired minus raw TrackingScore was −0.001003, with clip-bootstrap 95% CI [−0.008588, +0.008020]. It demonstrated no aggregate tracking gain. The separate 22/26 repair qualification rate describes the physics screen.
One fixed candidate completed two fresh runs of 512 environments and 4,000 PPO iterations. Both passed full sampler replay and the allocation gate. All recorded invalid-start, invalid-reference, and censored-reset counters stayed at zero.
Replay of E4’s saved sampler states showed that the fixed additive floor increasingly dominated the focus distribution as absolute progress shrank. We therefore normalized the existing progress signal by its deployment-weighted mean, leaving the estimator and admitted starts fixed.
| Run | Mean TV | Minimum effective units | Completed trials | Decision |
|---|---|---|---|---|
| Relative R · seed 11 | 0.083211 | 616.25 | 1,087,814 | Manipulation pass |
| Relative R · seed 12 | 0.082728 | 618.61 | 1,086,192 | Manipulation pass |
| Failure D · seed 31 | 0.084955 | 625.06 | 1,113,773 | Calibration pass |
| Failure D · seed 32 | 0.085939 | 624.29 | 1,118,388 | Calibration pass |
TV is total variation: half the summed absolute probability differences from the deployment prior. Effective units summarize distribution breadth using entropy. Each run has 41 saved snapshots, including 37 from iteration 400 onward. Those correlated snapshots are not independent training seeds. E4 used a different saved cadence, so its mean is a diagnostic reference rather than a matched estimate.
Let b be the deployment prior over admitted motion units and g the existing nonnegative absolute change in smoothed conditional success. Set its deployment-weighted mean to μ = Σu bugu. With complete history and positive μ, use:
wᵤ = gᵤ / μ + 2 qᵤ = bᵤ wᵤ / 3 pᵤ = 0.40 bᵤ + 0.60 qᵤ Apply the existing joint unit and clip caps.
Before active caps, this is exactly 80% deployment prior + 20% normalized progress distribution. A common positive rescaling of all progress values leaves it unchanged. With incomplete history or zero mean progress, the existing prior/cap fallback applies. This is a different normalization of the same information source.
The controlled setting stays fixed: Unitree G1 in pinned mjlab 1.6.0, a flat scene, 800 training motions, 1,184 admitted units, 368,951 exact legal starts, and 50-step trials that cannot wrap or leave an admitted interval. PPO, observations, actions, rewards, the progress estimator, and the 0.05 unit / 0.25 clip probability caps remain matched.
Main unresolved mechanism: absolute progress includes both improvement and decline. About 46.6% and 46.8% of the positive excess allocation mass in the two R histories went to declining success estimates, averaged over post-warm-up snapshots. That might reflect useful recovery from forgetting or estimation noise. These percentages describe excess probability mass, not the fraction of all training samples. More TV alone cannot resolve this.
Learning-progress curricula and failure-based motion sampling already have precedents. The proposed contribution is an auditable composition with exact support and a controlled policy test; the normalization alone is a small engineering change. No novelty or best-performance claim is established here.
| Arm | Allocation rule | Question |
|---|---|---|
| U | Deployment-uniform prior | Does adapting exposure help at all? |
| A | Existing absolute-progress floor | Does the scale correction matter? |
| R | Fixed relative-progress rule above | Does sustained progress allocation improve tracking? |
| D | Conditional failure, fixed ρ = 0.8, power 1, floor 0 | Is progress more useful than practicing failures? |
Run all four arms on fresh paired training seeds 21, 22 and 23, with 512 environments and 4,000 iterations per run. Evaluate checkpoints 1000, 2000, 3000 and 3999 on 100 held-out clips, including 25 feasible-hard clips, using 2,800 paired conditions per evaluation cell. This gives 12 training jobs and 48 evaluation cells. All 12 training manipulation checks precede evaluation.
The primary endpoint is the final-checkpoint R−U difference in feasible-hard clip-mean, liveness-weighted TrackingScore. A benefit decision requires a mean gain of at least +0.02, a positive lower two-sided seed-level 95% t-confidence bound (df = 2), and an all-panel lower bound above −0.01. With valid preceding gates, a hard-panel upper bound below +0.02 rules out the target-sized benefit; other unresolved cases remain inconclusive under the fixed decision rules. Gate failures retain their declared not_tested or invalid status.
R−A, R−D, learning curves, survivor-conditioned pose error with coverage, all-condition mechanical work with exposure, and elapsed cost are secondary descriptions. R−D is not a pure ranking-only ablation: startup behavior and floor/cap composition differ. A paired hierarchical bootstrap is supplementary to the primary seed-level interval.
D31/D32 and all four seed-51 confirmation-entrypoint smokes passed. Source bindings, replay logs and all 12 configuration hashes were verified before sealing the current contract. Scientific settings remain fixed.
Ten training gates pass at this snapshot. U23 is running, followed by A23. The scheduler opens policy evaluation only after all twelve gates pass; a failure retains its declared disposition.
Run the 48 evaluation cells and validate all training, pairing and provenance records before aggregation. The queued efficiency postprocessor requires exact reproduction of the original complete analysis.
Precision is a real limitation. With three paired seeds, the primary 95% t-interval half-width is 2.484 times the observed seed-difference standard deviation. At an observed mean of +0.02, a positive lower bound needs that SD below about 0.00805. A prospective precision audit explores hypothetical normal-distribution scenarios; it is not measured power or a forecast. An inconclusive result is plausible and will be retained.
The implementation includes checkpoint-linked sampler telemetry, complete history replay, paired-evaluation provenance checks and a strict campaign analyzer. All four actual seed-51 smokes passed before the contract freeze. Sample-efficiency analysis, the frozen-policy pilot and CUDA lifecycle checks are queued with separate prerequisites; these queued stages are not completed evidence.
The downloads below export already measured results and identify their original artifacts by repository-relative path and SHA-256. They contain no trained weights or licensed motion payloads. Incomplete training runs are excluded from the completed-run comparison. No held-out policy outcomes were read by the public exporter.
Labels: sealed means the protocol or decision is preserved; measured means observed in the stated experiment; exploratory limits generalization; pending means the evidence does not yet exist. Simulator qualification does not establish hardware feasibility. Repair uncertainty here resamples clips for one fixed policy, not training seeds; numerical repeatability and repair-specific effects remain separate questions.