Inference
Sample many. Keep the best.
The deployable win: sample N candidate LUTs, rerank by behavioral fidelity, ship the winner.
Read off greedily, the generator suffers exposure bias: it over-commits to a single dominant code and collapses to a low-fidelity LUT that mostly misses the request. The fix needs no retraining — sample a handful of candidates, decode and score each against the requested spec with a reranker that needs no ground truth, and return the best one. That roughly doubles free-running fidelity (0.159 → 0.307).
Why one shot fails
The collapse, and the fix
Coverage vs selection
How far the reranker climbs with N
oracle@k is the best fidelity reachable among the first k samples (a perfect reranker's ceiling);
the deployable reranker captures nearly all of it. The dashed line is the free-running greedy baseline.
Interactive