Gaugepath — prove the policy before the plant

EXPERIMENT GP-042 · MODEL HASH 8F2A

Lock the question before seeing the answer.

QUALIFIED JUL 22

✓ State

Queues, machine state, part priority

✓ Actions

Route to station A, B, hold

✓ Constraints

Capacity, safety, max dwell

✓ Calibration

4.2% median error vs June

scenarios240 held out

random seeds1001–1240

primary metricmoves / completed batch

tail riskp95 dwell minutes

hard constraint0 blocked egress

🔒 Frozen Jul 22, 14:08 by Elena Cho. Any change creates GP-043.

POLICIES UNDER TEST

Rule R17

Current priority + nearest open station.

CP-SAT baseline

15-second rolling optimization.

Learned challenger L6

Reward mapped to moves, dwell, and throughput.

240 IDENTICAL SCENARIOS · COMPLETED 07:42

A classical method wins this one.

0 HARD VIOLATIONS

CURRENT RULE R17

79.2

median moves

p95 dwell 18.4m
runtime 0.02s

RECOMMENDED · CP-SAT

69.8

median moves

p95 dwell 15.1m
runtime 12.8s

LEARNED L6

68.9

median moves

p95 dwell 22.7m under surge
runtime 0.08s

ENGINEER RECOMMENDATION

Advance CP-SAT to shadow evaluation.

L6 improves the average but destabilizes tail dwell under 18%+ demand surge. The simpler baseline meets every frozen threshold.

LEARNED CHALLENGER L6 · REWARD MAP

When reward rises but reality gets worse.

− movementweight 1.0active

− queue dwellweight .65active

+ throughputweight .8active

− re-routeweight .15weak

Proxy failure in 17 surge scenarios
L6 improves movement reward by holding priority parts too long, increasing p95 dwell 27% while aggregate reward still rises.

Prove the policy before the plant pays for it.

Lock the question before seeing the answer.

Rule R17

CP-SAT baseline

Learned challenger L6

A classical method wins this one.

Advance CP-SAT to shadow evaluation.

When reward rises but reality gets worse.