The higher guard: retraining every skill onto a new stance

I asked for a new guard: both wrists at the brow, palms facing forward and turned about 25° in, the forearms framing the face with the eyes clear. Before, the fists sat at eye and nose level. Every reference table was rebuilt from the new pose (rhythm, recovery, the five punches and their aim grids, blocks, walking). I wanted the current models fine-tuned onto it rather than retrained from scratch, and balance, walking, punches and blocks all had to still work.

The parent was opp15-14.5M, the blocking keeper on the old guard.

The runs

run	what changed	outcome
guard1	opp15 fine-tuned in the opponent env on the new guard	stopped at 34.2 M: punches, walk, blocks short
guard2	more punch, block and walk episodes	stopped at 159 M: everything held but the punches
guard3	level-head and palm-to-ear references	lead hook about 1 m/s, 0/12 (it got ~3 % of episodes)
guard4	three times the lead-hook presses	56 M at the 3 h cap: no lead hook, punches 52 → 44
guard5	no walk teacher while punching (teacher-free block-punch)	kept 26.2 M: every row at or above opp15, opponent 125/160
guard6	the lead hook from my capture	trained on a body with the ankle widened for the retarget: walk 10/40, not kept
guard7	guard6 with the ankle back	lead hook 11/12 but punches 47 → 38, walk 31–33/40; guard5 itself lands the captured hook 10/12, so it stays
guard8–10	the lead hook’s form: impact and swing rewards, a follow-through	the swing still ~25 cm low
guard11	the reference’s feed-forward torque on the hook arm	swing −14 to −18 cm, blocks and opponent lost
guard12	guard5 as the teacher outside the lead hook, an arm-pose term	other skills held (blocks 31, opponent 118), elbow still 10 cm low
guard13	the arm term narrower, with the trunk joints	swing −10, elbow −9 cm
guard14	the swing term ×3	kept 45.9 M: elbow level with the fist, lead hook 12/12, every row at guard5’s
guard15	the jab’s aim	stopped at 0.7 M when I asked for the original elbows
guard16–18	the original elbows, the jab at the reference’s pace, its pose and line terms	guard18-124.5M: the base for new training (jab 2–4 cm off; blocks 27/32)

guard14 is the parent of the first body-straight and knee teachers; guard18 is the base of the distillation and every later clone.

Where the money went

	machines	machine-hours	cost
guard1–14 on vast.ai, plus gate, clip and render jobs	29	about 21.6	about $17.62
guard15–18 on RunPod, plus an angle render and the gate	7	about 2.9	about $3.42
total	36	about 24.5	about $21.04

Several vast.ai boxes never came up (created stopped, no SSH after 4 minutes, or still loading after 10). They are billed for the minutes they existed and are in the totals. RunPod first refused every request filtered to CUDA 13 (“no longer any instances available”) while 5090s were in stock; without the filter it deployed at once.

Every run

From the run records in kakuto-ml: each run's machines, their billed hours and cost (EC2 runs are estimates from the run logs), plus the tests and renders this log covers.

• lead-hook reference renders for the higher guard, 22 October: 2 machines, 0.1 h, $0.09

← All the work logs for “Shadow clones: four Muay Thai techniques learnt in parallel”