Claim 1 · Distribution
Applied verification
Three seeds; every PCS model improves held-out censored NLL over a goal-conditioned repeated-hazard baseline.
ICML 2026 reproduction · applied distribution audit
SVL learns a real navigation time-to-goal law
Released survival tuples, critic, and censored likelihood on the full official PointMaze training pool. Disjoint validation and a standalone NumPy audit test the distribution directly—not through the value identity.
Claim 2 · Identity
Verified
Bellman, survival-matrix, and value iteration agree within 1.2e-12; two scope controls break it.
Claim 3 · Benchmarks
Inconclusive
No actors, HIQL rerun, or long-horizon success-rate claim. C1 evidence is not relabeled as C3.
Held-out censored negative log-likelihood · lower is better
0.760mean NLL improvement
paired gains 0.715–0.807
paired gains 0.715–0.807
Independent reconstruction
Mass error ≤ 1.34e-15; zero monotonicity violations; raw NLLs and hashes match exactly; target shuffling costs 1.755–2.082 NLL.
Scope & cost
1M train pool
100k validation pool
3 seeds · 87.13 s
648 MiB peak RSS
USD 0 · 7/7 tests