# Conditional matched depth-training comparison

Specified while the initial depth pilot is running, before training these candidates. Run only after a shallow depth passes the pilot's recorded admission gate.

Both arms use the validated expanded-leg actor: the learned 0.25 residual is preserved, and a zero-initialized leg branch adds up to 0.75 on the twelve leg actions. The two arms start from the same retained 400-rollout checkpoint. Architecture, observation inputs, all baseline rewards, torque limits, failure criteria, action noise 0.1, 84 eligible motions, four fixed terrain families and original reference timing match. Each arm has 256 rollouts of 24 steps across four environments (24,576 transitions), with four PPO epochs (1,024 optimizer steps). Actor and critic learning rates are 1e-4, with the existing KL and replay guards active.

Use four native processes of 64 rollouts each, warm-starting policy and critic weights from the prior phase with fresh optimizers. Both arms restart equally and use the same phase seeds [23, 24, 25, 26]. This is one training-seed sequence, not four independent training replicas. The depth-scheduled arm uses [0.04, 0.08, 0.14, 0.14] m if 4 cm is admitted; otherwise [0.02, 0.06, 0.14, 0.14] m if only 2 cm is admitted. The matched control uses 0.14 m in every phase. Ground is present in all phases. Depth is static during each process, so geometry never changes underneath a contacting robot. This is a fixed schedule admitted by a practice pilot, not competence-based promotion. There is no deep-soil rehearsal during the first two scheduled phases; the final half of the training budget uses the original target depth.

The first phase verifies that the added branch contributes exactly zero initially. Subsequent phases strictly load the expanded checkpoint. Log sampled motion coverage, per-terrain failures and clip-tail completions, actual material contact, replay error, post-update KL, finite parameters, teacher preservation and checkpoint hashes. A failed guard ends and preserves the attempt. A learning-rate retry, if required, restarts both complete arms from the common original checkpoint and uses 5e-5 for actor and critic; never silently continue a rejected update.

Both final candidates receive the same three original-speed development clips, seed 23, 500 control steps, all four evaluation families, original 14 cm depth and 2 cm grid. The fresh 14 cm pilot supplies the unchanged-policy control. Acceptance requires at least 20% fewer pooled soil task failures than both that starting policy and the equal-budget target-only arm, zero ground task failures, no per-soil-family root/planar-velocity RMSE regression versus start, and at least 95% retained soil contact. Report full-reference completions and per-episode contact separately; do not select intermediate checkpoints.

Only a passing final candidate triggers paired validation with reset seed 41 against the starting policy on the same clips, frozen checkpoint and unchanged settings. Require the failure reduction to remain at least 20% pooled across seeds, no failure-count regression on seed 41 alone, and the same pooled tracking/contact checks. This would still not establish independent-training-seed or held-out generalization. Reserved motions remain unused during this comparison. Negative results do not justify claiming a stable policy.