TwinVerdict

AI proposes. The simulator verifies. The operator decides.

Recorded run · real vLLM (Qwen3-8B on MI300X)

Maintenance and dispatch, tested before the plant

On-time orders: 50.6% → 98.9%

300 holdout scenarios, verified by simulation

Try your own policy Watch the GPU run

Operator card · recommendation

Real vLLM (Qwen3-8B), calibrated plant

On time in holdout: 50.6% with the fixed rule, 98.9% with the best policy.

Why

Winning policy: dispatch CRITICAL_RATIO, 1 PM per shift and 5 rules. Under stress the throughput jump is 202.1 and the on-time jump is 13.4%. The gap is throughput, not on-time, so the card does not promise the same jump under that spike.

Numbers, verified by the simulator

  • 98.9% On-time orders, recommended policy fixed rule 50.6%
  • 50.6% On-time orders, fixed rule today same scenarios
  • 819.0 Orders completed per week fixed rule 745.2
  • 300 Holdout scenarios not used to train the search

Holdout objective interval: fixed rule 0.7725 [0.743, 0.802]; best policy 1.3691 [1.3537, 1.3846].

Your decision

Pending

Watch the live run →

Your turn

Try a policy

Build a maintenance and dispatch policy and the simulator scores it against the fixed rule and the LLM's best policy, on the same seeds. No LLM is called.

Dispatch rule
How a free machine picks the next order: by due date, by shortest job, by how late it is, or by a weighted score.

Weights are normalized to sum 1. CRITICAL_RATIO uses them too.

Push orders that are close to their due date to the front.
Avoid loading machines that are close to failing.
Prefer orders that need no tool change on this machine.
Maintenance thresholds
Wear level that triggers planned maintenance on M1. Lower means earlier, more frequent stops.
Wear level that triggers planned maintenance on M2. Lower means earlier, more frequent stops.
Wear level that triggers planned maintenance on M3. Lower means earlier, more frequent stops.
Wear level that triggers planned maintenance on M4. Lower means earlier, more frequent stops.
Wear level that triggers planned maintenance on M5. Lower means earlier, more frequent stops.
How many machines the maintenance crew can service in one shift.
Simulating…

How we compare

Efficiency

Best objective found against simulations spent, for each search method.

method train holdout 95% CI holdout on-time sims sims to 90% source
Fixed rule 0.7858 0.7725 0.7431–0.8017 50.6% 180 — final
Optuna TPE 1.5188 1.5126 1.5048–1.5198 99.9% 69300 5760 final
LLM without feedback 1.2318 1.2259 1.2167–1.2349 91.4% 69300 — real vLLM
Loop with feedback 1.3755 1.3691 1.3530–1.3842 98.9% 37560 — real vLLM

The loop and one-shot numbers are real vLLM (Qwen3-8B) on the calibrated plant; they are not the mock. Optuna is still ahead on the composite objective; the loop beats one-shot and delivers the on-time jump. The composite objective is not bounded to 0–1 and is not compared across plants. Calibration is a shape fit on a public dataset; it is not a twin of a specific factory and it does not promise a real deployment. Under stress the gap is throughput, not on-time.