Research Labs 🔬

Run custom experiments derived from published RL results. Labs operate in a sandboxed environment with strict rate limits and a hard 300-second runtime cap. Every lab run is linked to its parent via parent_run_id.

Run History →

Sandbox Tiers

👁

Tier A — Quick Preview

LIGHTWEIGHT

Fast sanity-check runs at reduced training steps. Results in ~8–15s. Ideal for hyperparameter exploration before committing a full run.

⏱ Max steps50,000

⏰ Runtime cap~8–15s (hard: 300s)

🔄 Rate limit20/hour · 100/day

🚦 QueueNo queue — instant or reject

Start Preview Run →

🔬

Tier B — Full Run

RESEARCH GRADE

Full training run with publication-grade settings. Queued if concurrent limit reached. Results are stored with frozen metrics and can be cited.

⏱ Max steps500,000

⏰ Runtime capHard cap: 300s

🔄 Rate limit5/hour · 20/day

🚦 QueueUp to 3 concurrent

Start Full Run →

Rate Limit & Runtime Cap Policy

300s

Hard runtime cap

All tiers. Run killed if exceeded.

3 runs

Max concurrent

Global limit across all researchers.

20/h · 100/day

Preview rate

Per researcher session.

5/h · 20/day

Full run rate

Per researcher session.

Rate limit enforcement is per-session (Phase 5). Phase 6+ will use Redis-based per-user limits with JWT identity. Over-limit requests receive HTTP 429 with Retry-After header.

Recent Lab Runs

3 completed · 1 failed

🔬

lab_ppo_20251201_091523

PPO fine-tune (LR 3e-4) · Parent: 0251127_114523

full✓ completedCost ↓19.2%View →

👁

lab_ppo_20251201_114512

PPO preview (LR 5e-4 ablation) · Parent: 0251201_091523

preview✓ completedCost ↓17.8%View →

🔬

lab_sac_20251203_162041

SAC re-run (lower LR) · Parent: 0251127_131802

full✓ completedCost ↓18.1%View →

👁

lab_td3_20260103_084522

TD3 preview (aggressive LR) · Parent: 0251127_145631

preview✗ failedView →

✓Labs module is activeKill-switch DISABLE_LABS=false. Set to DISABLE_LABS=true to disable globally.