trane-project / trane / 22074080966

100%

push

github

More improvements to PowerLawScorer (#370)

- Interval-aware spacing effects: Successful recalls now increase
stability more when they happen
after lower pre-review retrievability (longer spacing), instead of
treating all successful
  reviews the same.
- Split success vs lapse stability updates: Lapses are no longer treated
as simple negative
growth; successful and failed recalls now use separate transition logic,
with lapses applying a
  bounded drop based on difficulty and surprise.
- Calibrated forgetting curve from decay (FSRS-style invariant): The
forgetting-curve factor is
now derived from decay so R(t=S) is consistently calibrated to 90% for
both exercise types,
  improving interpretability and cross-type consistency.
- Explicit high-stability damping: Successful growth now includes a
damping term so relative
gains shrink as stability increases, preventing runaway growth and
producing smoother long-run
  saturation.
- Dynamic, mean-reverting difficulty updates: Difficulty now evolves per
review during stability
chaining, allowing it to respond to performance trends while remaining
bounded through mean
  reversion.

101 of 101 new or added lines in 1 file covered. (100.0%)

4592 of 4592 relevant lines covered (100.0%)

2393404.12 hits per line