[jepsontaylor]

H237 prereg: AdaBelief / RAdam for the r882 adapter continuation (xbabe2 CPU, host amendment A1)

Preregisters H237 (hw_t0030_opus_h237_adabelief.md). Nothing scientific runs before this merges. Does not touch PR #2071.

Idea. Is the plateau partly an adaptive-step-size artefact? AdaBelief (Zhuang et al. 2020) scales steps by the variance of the gradient around its EMA instead of the raw second moment.

Arms. Only the optimizer differs. All arms share the CTRL loss, batches, lr 1e-4, betas (0.9, 0.999) and decoupled wd 1e-4.

• CONTROL: AdamW (exact H170 CTRL).
• ABEL (primary): inline AdaBelief with ε = 1e-16 inside s, bias correction, decoupled wd and no rectification.
• RADAM (secondary): torch.optim.RAdam with decoupled_weight_decay=True. Holm m = 2.

Machinery. Merged H234/H232 with the A1/A2 fixes. Fresh per-seed folds 237141–237143 (best-match Jaccard ≤ 0.186 against earlier waves), seeds 237101–237103, 15 cells, 1,800 updates. Read-only r882 copy, with DOPE c51a56d0 frozen.

Decision. On DL-Val91, an arm ADVANCEs iff its median seed gain against r882 is > +0.002, its family LCB95 is > 0 and its Holm p is < 0.05. Beating CONTROL alone is descriptive. Elo is reported against r882 and CONTROL, and each arm is also compared descriptively with the plateau best 1.1389572 (r3175).

Guards (memory rule from hw_t0030; host from amendment A1, hw_t0036).

• xbabe2 CPU, taskset 48–51, nice 19. These cores avoid H231, H233 and H236.
• Pause the worker's own process group on load5 > 108 or MemAvailable < 20 GiB; resume at load5 ≤ 100 and MemAvailable ≥ 22 GiB.
• Hard stops: MemAvailable < 12 GiB, RSS 4 GiB, free disk 300 GB, 4 threads, and the compute stop at 06:30Z.
• Admission needs load5 ≤ 100 and MemAvailable ≥ 22 GiB.

Amendment A1 (pre-outcome, host only). xbabe1 MemAvailable was 16–17 GiB, below the 22 GiB gate. The pre-merge smoke there was refused at admission and never ran. A1 moves the wave to xbabe2 (about 70 GiB), with inputs re-pinned to a hash-identical r882 copy. It is recorded in the PREREG Deviations section and in H237_claim.txt. No science change.

Local checks at this head.

• test_h237.py: 15 passed. They include AdaBelief against the reference update over 30 steps (to 1e-12), the belief property (larger steps than Adam on consistent gradients), RAdam with decoupled wd, and the monitor's load and memory pause hysteresis.
• ruff, git diff --check and the JSON parses are clean.
• All 7 jope-records gates return rc 0, including coverage validate.
• All input and asset pins verify on xbabe2.
• Pre-merge constructor smoke on xbabe2 at this head: PASS (47 tensors, 29,211 parameters, DOPE c51a56d0, 0 updates).

🤖 Generated with Claude Code

[jepsontaylor]

Same-agent review (rd_opus), head 26fb142e5fa943d01e7df21d2bdf9183b46a4c1a. This is not an independent review.

• Science.
  • treatment.py: inline AdaBelief equals the reference update over 30 steps (to 1e-12) and skips None gradients; RAdam is torch's with decoupled wd; every arm uses the CTRL loss.
  • analysis.py: seeds and per-seed folds 237141–237143, arms CONTROL/ABEL/RADAM, Holm m = 2 with the +0.002 margin.
  • runner.py: identical to merged H234 apart from the host and wave name.
• Monitor.
  • Memory below 20 GiB pauses the worker's own group, resuming at 22 GiB; below 12 GiB is a hard stop.
  • Load5 > 108 pauses, resuming at ≤ 100 (A1). Hysteresis is tested.
  • Only the worker's own process group is ever signalled.
• A1 (host only, pre-outcome). Recorded in the PREREG Deviations section and the claim receipt. All pins re-verified on xbabe2.
• Pre-merge smoke at this head on xbabe2: PASS.
• Checks. 15 tests passed, ruff clean, 7 gates rc 0.
• Findings. None blocking.