trueskill-tt

Three production sites use std transcendentals, and hypot diverges more than the exp case the rule cites numerics testing

#63

by logaritmisk was closed 2026-09-09 16:10:52 +00:00

posterior_of is nondeterministic across processes: float sums over a HashMap bug numerics

#62

by logaritmisk was closed 2026-09-09 16:10:52 +00:00

Gaussian::from_ms overflows for tiny sigma, and three constructors still absorb a negative sign silently api bug numerics

#61

by logaritmisk was closed 2026-09-09 16:10:51 +00:00

Truncation 1 - w cancellation gives NaN skills from plausible parameters bug numerics

#60

by logaritmisk was closed 2026-09-09 16:10:51 +00:00

quality() returns 0 then NaN past ~150 groups: a determinant ratio in linear space bug numerics

#59

by logaritmisk was closed 2026-09-09 16:10:51 +00:00

tuple_max erases NaN, so "NaN is never convergence" is a coin flip across processes bug numerics

#58

by logaritmisk was closed 2026-09-09 16:10:51 +00:00

A tiny non-zero drift returns a posterior variance 12 000x too small, as Ok bug numerics

#57

by logaritmisk was closed 2026-09-09 16:10:51 +00:00

expected_information_gain returns 4.67x its own ln k ceiling bug numerics

#56

by logaritmisk was closed 2026-09-09 16:10:51 +00:00

predict_outcome returns probabilities greater than 1: the grid clamps with no detection bug numerics

#55

by logaritmisk was closed 2026-09-09 16:10:51 +00:00

The joint precision matrix is 0.19% dense and we factorise it densely enhancement numerics performance

#52

by logaritmisk was closed 2026-09-10 05:08:22 +00:00

2

A node's reported sigma is ~10x wider than the estimate's measured reproducibility decision numerics

#47

by logaritmisk was closed 2026-09-08 17:46:28 +00:00

5

No joint posterior: anything derived from two or more competitors has to add marginals and overstate its width api enhancement numerics

#46

by logaritmisk was closed 2026-09-08 04:49:07 +00:00

2

erfc is a 1.2e-7 approximation and sets the accuracy floor for the whole crate numerics performance

#41

by logaritmisk was closed 2026-09-07 20:25:11 +00:00

2

Spec's tolerance-not-bit-identity caveat for filtered estimates is conservative docs numerics

#30

by logaritmisk was closed 2026-08-27 15:22:05 +00:00

1

log_evidence_internal's forward flag is only a filtering quantity before convergence docs numerics

#28

by logaritmisk was closed 2026-08-27 15:22:14 +00:00

1

Hot-path Gaussian arithmetic round-trips through sqrt; convergence deltas computed in moment space numerics performance

#15

by logaritmisk was closed 2026-08-04 20:13:42 +00:00

1

Evidence accumulates as a linear product, so log-evidence underflows to -inf on large games bug numerics

#12

by logaritmisk was closed 2026-08-04 20:13:34 +00:00

1

Draws with the default p_draw = 0.0 silently produce NaN skills and report converged: true bug numerics

#8

by logaritmisk was closed 2026-08-04 20:13:31 +00:00

1

Opt-in synchronous-EP Schedule with snapshot + barrier merge enhancement numerics

#6

by logaritmisk was closed 2026-09-08 00:18:14 +00:00

2

converge() returns all-NaN once a time slice exceeds ~75 competitors (regression vs 0.1.0) bug numerics

#3

by logaritmisk was closed 2026-06-12 20:26:27 +00:00