In deeptutor/learning/mastery.py, compute_mastery is documented as "recency-weighted accuracy" (newer attempts count more, so recovery after early mistakes is rewarded). But the scoring line binds the loop variables in the wrong order:

score = sum(w * (1.0 if c else 0.0) for w, c in zip(recent, weights, strict=True)) / sum(weights)

recent is the correctness list and weights is the recency weights, so zip(recent, weights) yields (correctness, weight) pairs — but they're unpacked as w, c, binding w to the outcome and c to the weight. Since a weight is always truthy, (1.0 if c else 0.0) is always 1.0, so each term reduces to the raw correctness bit. The result:

• the numerator becomes a plain correct-count (recency weights only affect the denominator), so
• mastery is order-independent — a recovering history scores the same as a declining one.

Minimal repro:

>>> compute_mastery([False, True, True])   # miss then two hits (recovering)
0.714...
>>> compute_mastery([True, True, False])   # two hits then a recent miss
0.714...   # identical — recency has no effect

Expected: [F,T,T] should score higher than [T,T,F] (recent attempts weigh more).

Fix — swap the unpack to match zip's order:

score = sum(w * (1.0 if c else 0.0) for c, w in zip(recent, weights, strict=True)) / sum(weights)

After the fix: [F,T,T] = 0.696 vs [T,T,F] = 0.643; the confidence caps are unaffected.