Issues

quality: review backlog — dead errorf_inv, untested public setters, fit_best_of_n CCN 18, clang-tidy SIMD cluster, kTrigDMax duplicate

quality

testing

#101

· OldCrow opened

on Aug 22, 2026

·

v4.4.2 — Fit accuracy & kernel hygiene

accuracy(distributions): LogNormal and Student-t weighted fits normalise by total weight while skipping out-of-support points

distributions

numerical accuracy

#100

· OldCrow opened

on Aug 22, 2026

·

v4.4.2 — Fit accuracy & kernel hygiene

accuracy(simd): exp_pd edge contract — −inf lanes return 3.3e-308 (not 0) and AVX2/AVX-512 map NaN to +inf

numerical accuracy

#99

· OldCrow opened

on Aug 22, 2026

·

v4.4.2 — Fit accuracy & kernel hygiene

quality: FileIOManager — 10 of 12 public statics have no in-library consumer; scope for the next major

quality

#98

· OldCrow opened

on Aug 22, 2026

·

v5.0.0 — API

quality: new v4.4.0 API polish — sample() skips validateInitialized(); sample_mv is snake_case; A20 allocate-then-overwrite

quality

#97

· OldCrow opened

on Aug 22, 2026

·

v4.5.0 — Algorithm Coverage

quality: configure treats every non-AArch64 target as x86 — armv7/riscv64/ppc64le fail late with 22 undefined `*_sse2` symbols

quality

#96

· OldCrow opened

on Aug 22, 2026

·

v4.5.0 — Algorithm Coverage

design: BasicTrainer::train() means "one pass" on BW/MAP-BW and "run to convergence" on Viterbi/SegmentalKMeans — the topology guidance is unsatisfiable on two trainers

enhancement

quality

#95

· OldCrow opened

on Aug 22, 2026

·

v5.0.0 — API

quality(cmake): LIBHMM_SIMD_SOURCES still applies -march=native to nine runtime-dispatched distribution TUs — measure, then trim

performance

quality

#94

· OldCrow opened

on Aug 22, 2026

·

v4.4.2 — Fit accuracy & kernel hygiene

quality(simd): NEON TU's non-AArch64 `#else` fallback defines 20 of the 22 kernels — delete the dead block

quality

#93

· OldCrow opened

on Aug 22, 2026

·

v4.4.2 — Fit accuracy & kernel hygiene

accuracy(student-t): ECME scale step uses the uncentred Σwx² − (Σwx)²/W form; silently skips or resets for |μ|/σ ≳ 1e6

distributions

numerical accuracy

#92

· OldCrow opened

on Aug 22, 2026

·

v4.4.2 — Fit accuracy & kernel hygiene

bug(io): from_json accepts NaN / inf / negative entries in pi and trans

bug

#91

· OldCrow opened

on Aug 22, 2026

·

v4.4.1 — Correctness patch

bug(student-t): location (μ) is never validated — NaN/inf μ accepted by the constructor, setLocation() and JSON load

bug

distributions

#90

· OldCrow opened

on Aug 22, 2026

·

v4.4.1 — Correctness patch

Navigation Menu

All issues

quality: review backlog — dead errorf_inv, untested public setters, fit_best_of_n CCN 18, clang-tidy SIMD cluster, kTrigDMax duplicate

accuracy(distributions): LogNormal and Student-t weighted fits normalise by total weight while skipping out-of-support points

accuracy(simd): exp_pd edge contract — −inf lanes return 3.3e-308 (not 0) and AVX2/AVX-512 map NaN to +inf

quality: FileIOManager — 10 of 12 public statics have no in-library consumer; scope for the next major

quality: new v4.4.0 API polish — sample() skips validateInitialized(); sample_mv is snake_case; A20 allocate-then-overwrite

quality: configure treats every non-AArch64 target as x86 — armv7/riscv64/ppc64le fail late with 22 undefined `*_sse2` symbols

design: BasicTrainer::train() means "one pass" on BW/MAP-BW and "run to convergence" on Viterbi/SegmentalKMeans — the topology guidance is unsatisfiable on two trainers

quality(cmake): LIBHMM_SIMD_SOURCES still applies -march=native to nine runtime-dispatched distribution TUs — measure, then trim

quality(simd): NEON TU's non-AArch64 `#else` fallback defines 20 of the 22 kernels — delete the dead block

accuracy(student-t): ECME scale step uses the uncentred Σwx² − (Σwx)²/W form; silently skips or resets for |μ|/σ ≳ 1e6

bug(io): from_json accepts NaN / inf / negative entries in pi and trans

bug(student-t): location (μ) is never validated — NaN/inf μ accepted by the constructor, setLocation() and JSON load

Search results