Deterministic prediction engine

See overview.

The v1 Intelligence tier's headline feature ranked previous year questions by how likely they were to reappear, using explicit statistical rules instead of a language model. Only questions with at least two years of history were eligible. The tier and its pro gate belong to v1, which had a paid tier; v2 has no paid tier and ranks topics instead (v2 topic ranking).

How it scores

• Frequency: points per year asked, capped at four years.
• Overdue: a bonus if the question has not appeared in over two years.
• Recent repeat damping: a penalty if it was asked this year or last.
• Periodicity: a bonus if there is a roughly two year or longer cadence and the current gap fits it.
• Topic momentum: a bonus for a rising or emerging topic, a penalty for a dropping one.
• Staleness: a penalty if the question has not appeared in over seven years, likely out of syllabus.

Ties break by score, then occurrence count, then alphabetical body, so ordering is total and stable. The top 20 are tiered by position. Near duplicate questions are clustered using token similarity so the same question worded differently is not double counted.

Why no LLM

The question "how often has this been asked, and when" is arithmetic. A language model would add cost, latency and run to run variance without adding accuracy, and would make the feature unauditable exactly where trust matters. Each recommendation carries its own explanation, generated by the same rules that produced the score, so a student can see why a question is on the list and still trust the tool.

That v1 feature was measured against the 2026 exam: the top 20 caught 8, a v1 strike rate of 40 percent (numbers). The number only means something because the bank it was built from (2016 to 2025) had never seen the 2026 paper. v2 does not rank questions; its claim is the held-out 2025 backtest, 7.67 topic recall@10 for frequency-only against 6.00 for the tuned weights (v2 topic ranking).