Incident Log · RLPulse — Radical Candor, in Public

◷ Incident Log · radical candor, in public

7 Events

No Western lab publishes this mid-run. Restarts, OOMs, broken data pipelines — and the shutdown fingerprint at step 30. Read the full detective story in chapter 08, and see the restart artifacts in the KL chart.

✓ Capstone · read any RL dashboard like an engineer

Start with avg@n + benchmarks. Is the model actually getting better and does it transfer off-distribution? Everything else is commentary.
Cross-check reward vs pass rate. Rising together: healthy. Reward up + pass flat/down: hacking alarm, go look at rollouts.
Read entropy as exploration budget. Gradual fall: converging. Fast crash: premature lock-in. Sustained rise: ask about anti-collapse tuning or harder curriculum.
Check KL + grad_norm + notices as one unit. Spikes with a restart notice are bookkeeping; without one, suspicion.
Interpret flat zero-rates via the curriculum. Flat all-fail with rising avg@n is a treadmill working.
Price the progress. Context × turns × step time × dollars: longer trajectories must buy capability, not length.

Normal vs alarm, pocket edition

Normal: avg@n + reward + benchmarks climbing; KL gently rising between restart resets; entropy drifting; step time growing with context; infra_error near zero. Alarm: reward/pass divergence; grad_norm spike with no notice; infra_error spike masquerading as a capability dip; passrate/one swallowing the batch with no curriculum refresh.

Figure	Pro (step 30)	Flash (step 30)
avg@n (Δ vs step 1)	0.633 (▲0.068)	0.644 (▲0.130)
Cost so far	$2,620,670	$854,044
Tokens (step / total)	3.43B / 75B	3.7B / 81.4B
Context / turns	~137k / 43.6	~148k / 68.5
Step time	6h26m	3h27m
DeepSWE / in-house / automation	72.57 / 65.43 / 53.10	65.68 / 62.87 / 52.70

Sources: live dashboard snapshot captured 2026-09-21 (mimo.xiaomi.com/rl); final-step values from the run's step-30 state. Inferred definitions are labeled as inference in the text.