Weight puzzle selection and stop repeats inside a session (#183)

Selection took a hard rating band (medium = user rating ± 100), removed
everything attempted in the last 30 days, and drew uniformly from what was
left. Because the bulk importer gives every endgame in a batch the same
rating, the band held either the whole library or — once a user's rating
drifted past it — only the endgames that user had already played. With
`fresh` empty, the fallback re-drew from the same band ignoring recency, so
recently played puzzles came straight back. An empty band used a
deterministic nearest-match, alternating between the same two puzzles.

Difficulty is now a bias rather than a filter: candidates are weighted by
`exp(-|rating - target| / spread)` and drawn with an exponential race, so
the whole library stays reachable and a thin region around the user's rating
can never empty the pool. `easy` and `hard` keep one-sided guards so the
labels stay honest. When every candidate is excluded, selection serves the
least recently attempted endgame instead of a random recent one.

Sessions now carry the ids they have served, which fixes the observed
in-session repeat (only the current puzzle used to be excluded), and guest
training gets the same treatment. The 30-day window is configurable as the
repeat window and is now named in GLOSSARY.md, distinct from the rating
window and from the tablebase cooldown.

Difficulty copy is qualitative, since the band numbers no longer describe
how selection works.
This commit is contained in:
2026-11-17 14:28:24 +01:00
committed by GitHub
parent fa1fef7b62
commit e0b446bbfb
8 changed files with 212 additions and 115 deletions
+15 -5
View File
@@ -105,6 +105,14 @@ carries a start time and a rating baseline for the timer and rating-gain
monitor. Ephemeral — it lives only in the running LiveView, never persisted.
_Avoid_: Run, game, attempt
**Repeat window**:
The minimum interval before the same endgame may be served to the same user
again. Selection skips endgames attempted inside the window; only when nothing
else is available does the least recently attempted endgame come back. It is
independent of the rating window, which decides whether a replay may move a
rating.
_Avoid_: Cooldown (reserved for tablebase rate-limiting), due date
## Ratings
**Rating**:
@@ -128,11 +136,13 @@ systems; the trainer deliberately does not track it.
_Avoid_: Variance, inconsistency
**Difficulty**:
A choice — `easy`, `medium` or `hard` — that selects which puzzles to serve
relative to the current user's rating: `medium` is near the user's rating,
`easy` below it, `hard` above it. The band width is a configured constant. It is
a property of a session, not of an endgame — the same endgame is easy for one
user and hard for another.
A choice — `easy`, `medium` or `hard` — that biases which puzzles are served
relative to the current user's rating: `medium` centres on the user's rating,
`easy` below it, `hard` above it. Candidates are weighted by how close their
rating is to that centre, so a session draws mostly from near the target but the
whole library stays reachable. The spread is a configured constant. It is a
property of a session, not of an endgame — the same endgame is easy for one user
and hard for another.
_Avoid_: Level, rank, tier
## Infrastructure