projects / easemysemester
EaseMySemester numbers, v2
v2-specific defensible numbers for EaseMySemester, with their caveats.
numbers · current · updated 2026-09-16 · EaseMySemester case study
See overview for the project overview, and numbers for v1 and cross-cutting figures.
| Number | Measures | Caveat |
|---|---|---|
| 809 questions, 34 chapters, 218 topics | v2's question bank, verified 2026-09-06: Anatomy 253, Biochemistry 253, Physiology 303 | The parser emits 816 questions from six corpus files; seven are asked in both Paper 1 and Paper 2 of a subject and stored once, kept as two occurrences (five in Anatomy, two in Biochemistry, none in Physiology). |
| 909 occurrences, 0 zero-occurrence questions | v2's question bank appearances, verified 2026-09-06 | Counted per appearance (a question can appear at more than one institution) rather than per question, so the reconciliation gate catches an import that silently drops occurrence tags (details). Confirmed against a fixed tag detector: parsing all six source files reports zero skipped lines and 910 appearance tags, one collapsed by a uniqueness constraint, not a missed truncation (details) |
| 7,731 chunks, 3,955,261 tokens, $0.079 | v2 corpus ingest across nine book rows, verified 2026-09-07 | BD Chaurasia split into four separate rows for its four volumes; Lippincott dropped; Netter kept as a plate-cited figure reference only |
| 435 chunks, 222,604 tokens | Netter ingested over its plate range (pdf 42-825), every chunk carrying a plate label | After two matcher fixes: a plural match reading plate numbers off the table of contents, and Electronic Bonus Plates colliding with the main plate series |
| ~89 percent vs ~14 percent | v2 topic-level recurrence ceiling vs question-level recurrence ceiling | The topic score is the strong claim; question ordering within a topic is a weaker tiebreaker kept structurally separate (details). Crediting near-duplicate questions as one moves the question-level figure to at most 17.3% on the 2025 exam year, still well below the topic ceiling (details) |
| 7.67 vs 6.00 vs 7.33 | v2 ranking backtest, topic recall@10 on a held-out 2025 exam year: frequency-only vs hand-tuned weights vs the weight sweep's tuning-best | Frequency-only won; the honest verdict per the harness's own rule is to ship it, not the tuned weights (details) |
| 11 of 23 sittings | Hand & Wrist topic's actual recurrence, v2 Anatomy, verified against the live corpus | The view shows true sittings a student can verify against their own papers, not a computed probability |
| 809 questions regenerated, zero errors | v2 full-bank answer regeneration under the composed per-subject prompt, 2026-09-08 | One prompt version per subject; single-part decline rate fell 10.9 percent to 7.9 percent, multi-part decline rate rose 9.2 percent to 28.7 percent (details) |
| 809 of 809, 721 complete, 88 declined, ~$0.65 | v2's first full-bank answer generation, one prompt version, 2026-09-07 | Predates the type-aware prompt overlay above; decline by subject was Anatomy 9.1%, Biochemistry 6.3%, Physiology 16.2%, inverting the expectation that Biochemistry (single-source) would decline most (details) |
| 82 of 809 (10.1%) declined | v2's current answer set, verified against the mobile design's honest-gap surface, 2026-09-09 | 7.9% of single-part questions decline against 28.7% of multi-part ones, evidence that one missing sub-part is sinking most of the remaining declines (details) |
| 720 of 809 (89.0%) appear once; 9 questions reach 3 or 4 appearances | Recurrence distribution across the whole corpus, counted per paper, verified against production 2026-09-09 | Counted per paper tag, not per distinct sitting; a question tagged on both papers of one exam counts twice. Confirmed stable under fuzzy near-duplicate merging up to a 0.90 similarity threshold, at which the ceiling is still 4 appearances and zero questions reach 7 (details) |
| 34 of ~50 | Students moved from v1 to v2, as of 2026-09-16 | Mid-migration, and counted as accounts moved rather than daily active. Expected to rise when the semester starts. The remaining students are still on v1, which stays hosted. |
| 0 | Paying users on v2 | v2 ships with no paid tier at all. The six who paid on v1 (numbers) are among the 34 and pay nothing, because there is nothing to pay for. The paywall was removed deliberately to build daily habit and word of mouth first, and because the features planned next need usage data that a paywalled minority would not produce. Paid scope is deferred, not abandoned: further subjects, supplementaries, and an expanded corpus. |
| 23 / 24 / 18 papers | Per-subject paper counts backing the mobile design's recurrence bands: Anatomy, Physiology, Biochemistry | Denominator for a topic's band is per subject, never a flat 10; produces 22 topics in "Every year," 78 in "Most years," 118 in "Some years" (details) |