[Rockdu]

What

• rollout/raw_reward in the two Kimi K3 4-layer e2e tests is now tracked, not gated (constraint=None). This is the same mode as perf/actor_train_mfu: the value is still recorded and reported against its own history, but it never fails CI or untrusts a run.
• No other gate changes. grad_norm, ppo_kl, train_rollout_logprob_abs_diff and train_rollout_kl keep their default bands.

Why

Both tests use --rm-type deterministic_random, because the pruned 4-layer model scores 0 on gsm8k. That reward is sha256(str(tokens) + response)[0] % 2 (miles/rollout/rm_hub/__init__.py). So raw_reward is the mean of 16 coin flips per step. It carries no model-quality signal, and no regression can lower its expected value of 0.5.

The default band floors it at 0.8 x mean. With the current baseline mean of 0.521875 the floor is 0.4175, so the gate fails at 6/16 or fewer. For 16 fair coins that happens 22.7% of the time.

Two more effects make it worse:

• Step 1 is re-drawn on every run. The step-1 tokens change from run to run even on main with the same code. For example, the nightlies of 2026-10-14, 10-15 and 10-16 have bit-identical step-0 grad_norm (2.4867029190063477), yet their step-1 raw_reward was 0.375, 0.5 and 0.5625.
• The baseline ratchets upward. A failing nightly is written with trusted=false, and the baseline reads only trusted rows. Low draws never enter the baseline, so its mean and floor keep drifting up.

Validation

Step-1 raw_reward from every scheduled PR Test nightly on main from 2026-10-09 to 10-30 (22 runs):

value	0.375	0.4375	0.5	0.5625	0.625
runs	2	3	5	9	3

• Two main runs (10-12, 10-14) scored below today's 0.4175 floor. The gate was not enforcing yet then, so they did not fail.
• On build(docker): bump SGLang to v0.5.23 #4548 (SGLang v0.5.23 bump), the gate failed with cur=0.3125 vs mean=0.521875 (n=20) outside [0.4175, inf]. Its step-0 rollout matched a main cluster (raw_reward 0.625, rollout logprob -10.22097). Its step-1 rollout logprob, logprob abs diff, train/rollout KL and grad norm were all inside the 22-run main range.

Local checks:

• The parsed specs for both files give rollout/raw_reward: steps="all", constraint=None, enforce=False.
• pytest tests/ci/test/{test_ci_register,test_history_gate,test_gate_integration,test_metric_selection,test_history_replay}.py: 374 passed. This includes test_every_training_e2e_with_a_history_gate_also_gates_mfu, which scans the e2e files.
• ruff check on both files is clean.

Not in this PR: the band in GATE_DEFAULTS and the trusted-only baseline. Other tests with a real reward still need them.

Files

• tests/e2e/megatron/model_scripts/test_kimi_k3_4layer_ci.py: rollout/raw_reward is tracked instead of gated.
• tests/e2e/megatron/model_scripts/test_kimi_k3_4layer_lora_ci.py: same change.

Checklist

• pre-commit run --all-files passes (not run; ruff check on the two changed files only)
• Added/updated tests for new behaviour (no tests added; declaration-only change, covered by the existing gate parser tests)
• pytest -x is green (partial: the tests/ci/test gate suites above, 374 passed; the full suite and the GPU e2e were not run)
• If launch flags changed, python3 train.py --help still parses (n/a: no flags changed)
• If a public flag was added, it appears in the CLI reference docs (n/a)
• If an example was added, it has a real walkthrough (n/a)

🤖 Generated with Claude Code

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}

[Rockdu]

@claude review

[claude]

Code review found no issues

No high-confidence issues detected in this change.