Methodology: mapping the math_oge bank to official OGE exam task numbers

Question this answers: the open FIPI task bank tags every math_oge question with a КЭС (content codifier, e.g. 7.3 Многоугольники) and an answer_type_code (see scraper/answer_types.py) — but not the official exam task number (№1 through №25). This document derives that mapping from FIPI's own official documents, sourced and read directly (not guessed), and is honest about where the derivation is exact vs. a best-effort group. It mirrors OGE_METHODOLOGY.md's approach for biology and METHODOLOGY.md's for EGE math — same philosophy, separate document since math_oge's own numbering, task shapes, and codifier sections are entirely different from either.

Primary sources (official, read directly — not summaries of summaries)

Fetched from FIPI's own document server (doc.fipi.ru), the authoritative source — not a third-party test-prep site. PDFs are re-hosted here so you can open them directly: Методология on the dashboard links every one of these inline.

This document uses the 2026 spec as primary, same reasoning as biology's: it's the current, ratified version, not a draft. Unlike EGE math's METHODOLOGY.md (which targets 2027 because the profile-level EGE is being restructured 19→20 tasks), math_oge's 2027 draft is structurally identical to 2026 — cross-checked directly: both specs' Таблица 2 (part/type/count/score) and Таблица 3 (КЭС-section distribution) are byte-identical (25 tasks, same 2/17/6 split, same 7/1/2/1/1/1/5/1 per-section counts). No restructuring, so no forward-looking reason to prefer the draft. The 2027 documents are still hosted above for reference/comparison.

Exam structure (from the 2026 specification, §6–7, Table 9)

25 tasks, two parts, 31 max points, 235 minutes (3h 55m): - Part 1, tasks 1–19 (19 tasks, 19 points): short-answer. Of these, only tasks 7 and 13 are single_choice ("Ответы к заданиям 7 и 13 запишите ... в виде одной цифры, которая соответствует номеру правильного ответа" — FIPI's own instruction text to students, unambiguous, quoted directly from the demo variant). Every other Part 1 task (1–6, 8–12, 14–19) is short_answer or matching — "число или последовательность цифр" per the same instruction. - Part 2, tasks 20–25 (6 tasks, 12 points): answer_type_code = 'extended_answer' — this is Развёрнутый ответ in our schema, and, exactly like EGE math and OGE biology's open-answer parts, has no FIPI answer-checker at all. Deliberately untouched by this project's verification pipeline, same reasoning as every other subject's extended-answer part (see OPEN_ANSWER_METHODOLOGY.md).

Task-by-task content code — from FIPI's own "Обобщённый план" appendix, not inferred

Unlike biology's demo-variant-only derivation, math_oge's 2026 specification has an explicit per-task content-code table (Приложение, "Обобщённый план варианта КИМ ОГЭ 2026 года по МАТЕМАТИКЕ") — the same kind of authoritative table METHODOLOGY.md uses for EGE math. Every row below is that table's own content code (КЭС section 1–8), cross-checked against real scraped codifiers values and real question text (see "What's exact vs. best-effort" below) — not guessed from the demo variant alone, though the demo variant's actual task-by-task content (read directly, page images) confirmed every row.

Answer-submission format (separately confirmed against the real FIPI checker)

Same closed-form encodings already reverse-engineered for bio_oge (BIO_OGE_ANALYSIS.md) apply unchanged — confirmed empirically for math_oge too, not assumed: single_choice = bare selected digit; multi_choice = positional binary string over every checkbox in order (not the selected numbers concatenated); matching = concatenated digits in position order; short_answer = plain text/number (decimal point, not comma — see scraper/solve_free_models.py's normalize_answer()). Verified live before the solving rollout: one real question per answer-type/solve_method combo, all four solved and confirmed correct on the first model tried (see WORKLOG.md 2026-09-20).

What's exact vs. best-effort — read this before trusting a number

Same principle as METHODOLOGY.md/OGE_METHODOLOGY.md: where the bank's own metadata (codifier + answer type) uniquely determines a task number, it's exact; where several task numbers share the same signature and can't be told apart, they're grouped honestly rather than guessed.

scraper/assign_math_oge_task_numbers.py encodes these rules. math_oge turned out to classify unusually cleanly:

• Exact via answer type + codifier section, zero ambiguity:
• matching → always task 11 (the only official matching task).
• multi_choice, or any row (regardless of answer type) tagged with a bare section-7 codifier ("7 Геометрия", no subsection number) → task 19. The bare-vs-subsection distinction matters here specifically: every other geometry task (15–18, 23, 25) is tagged with a specific subsection (7.1–7.5), never bare — confirmed directly against real scraped rows, not assumed.
• single_choice + codifier section 3 → task 13; single_choice + section 1, 2, or 6 → task 7.
• short_answer + section 4 → task 14; + section 8 → task 10; + codifier 3.1 specifically → task 9; + codifier 2.1 specifically → task 12 (confirmed by reading real sample question text for both: 2.1-tagged items are consistently "compute via formula" word problems like task 12's demo example, while 2.2–2.5 are arithmetic-expression evaluations that belong with tasks 1–8 instead).
• Grouped (the bank's metadata genuinely can't separate them further — and for the geometry clusters, this isn't even a limitation of OUR data: FIPI's own appendix table assigns tasks 15–18 the identical content+requirement code, and the same for 23–25):
• short_answer + section 1, or 2.2–2.5, or 3.3 → grouped "1,2,3,4,5,6,8" (the real-life scenario module plus the two arithmetic-expression tasks — genuinely indistinguishable from each other by codifier alone, and task 1–4's own official content code in the appendix is literally the blanket range "1–8").
• short_answer + section 7 → grouped "15,16,17,18".
• extended_answer + section 7 → grouped "23,24,25"; + section 2 or 3 → grouped "20,21".

Result on the real 3884-question bank (2026-09-20 run): 1250 exact, 2633 grouped, 1 unclassified (a single question with an empty codifiers array on FIPI's own side — a real scraping gap, not a rule-table gap; left NULL rather than guessed). This is a notably cleaner split than biology's (978 exact / 510 grouped / 491 unclassified out of 1979) — mainly because FIPI's own math spec ships an explicit per-task content code (the Обобщённый план appendix), where biology's spec only gives aggregate section counts.

Implementation

scraper/assign_math_oge_task_numbers.py applies the rules above over codifiers + answer_type_code, writing questions.oge_task_number / oge_task_group / oge_task_confidence — the same three columns bio_oge uses, not a separate pair, since both subjects share exam = 'oge' (see schema.sql's reasoning for why EGE and OGE get separate column pairs but two OGE subjects can share one: each row belongs to exactly one subject, and this script — like assign_oge_task_numbers.py — is scoped to WHERE subject = 'math_oge' only). Because these columns are now genuinely shared between two subjects with unrelated numbering, web/app.py's dropdown/label/filter code was generalized to always resolve a task label and build a filter clause per subject, not by task number alone — see WORKLOG.md 2026-09-20 for the bug this would otherwise have caused (task 11 silently showing biology's label on a math question, or vice versa).

Re-run any time this rule table is revised or the bank is re-scraped — it's idempotent, just re-derives and overwrites these three columns for every math_oge row.