The 350,000-Question Mirage: Why Massive Nigerian Past-Question Databases Are Mathematically Impossible Without Duplication
Any API claiming 350,000+ historical Nigerian examination questions is counting multi-source OCR duplicates and unverified answer keys. Here is the combinatorial proof of West Africa's real historical question ceiling—and how our SHA-256, pgvector, and Maker-Checker pipeline audits every item.
Every few months, an engineering lead at an African EdTech startup sends us the same procurement question. While comparing assessment providers for a mock CBT platform or an AI tutor, they notice a vendor advertising 350,000+ or 370,000+ Nigerian past questions on a landing page, right next to ALOC Station's live quality telemetry reporting 12,645 verified, curriculum-aligned items.
On a surface-level spreadsheet, three hundred and fifty thousand looks twenty-eight times larger than twelve thousand. To a non-technical buyer, raw row count feels like safety.
To an assessment systems engineer, however, claiming 350,000+ authentic historical Nigerian examination questions is an immediate due-diligence red flag. You do not even need to inspect the vendor's database to know why—you only need basic combinatorics and knowledge of how West African examination boards have operated since 1978.
Let us walk through the exact arithmetic of every major Nigerian examination body, trace how multi-site web scraping inflates database row counts by 2.5x to 3.5x, and examine the four-stage verification architecture we built at ALOC Station to keep poisoned data out of student-facing products.
The Combinatorial Ceiling: Counting Every Paper Ever Administered
Suppose a hypothetical archive possessed a pristine copy of every single multiple-choice examination paper ever printed or administered in Nigeria across the last forty-seven years, with zero missing booklets and zero lost sittings.
Start with the Joint Admissions and Matriculation Board (JAMB UTME), first conducted in 1978. Across roughly 47 annual sittings and approximately 25 offered subjects, standard papers historically carried 50 questions (later 40 questions in the modern CBT era, alongside 60 in Use of English). Even if every single subject paper from 1978 to 2025 survived intact—ignoring the reality that post-2013 CBT item banks are no longer released as public paper booklets—the theoretical upper bound of unique JAMB Paper 1 items is roughly 50,000 to 55,000 questions.
Now add the West African Examinations Council (WAEC WASSCE) across both the May/June school diet and the November/December private candidate diet from 1988 through 2025. Across 38 years, two annual sittings, and roughly 22 core and elective subjects averaging 50 objective items per paper, complete historical capture yields an upper bound between 75,000 and 85,000 questions.
Include the National Examinations Council (NECO SSCE), established in 2000. Twenty-five examination years across roughly 20 subjects at 60 objective items per sitting adds at most 30,000 questions. Finally, layer in Post-UTME university screening papers (2005–2024) across Nigeria's twenty most active federal and state universities—typically four screening subjects of 25 to 50 questions—adding roughly 15,000 to 20,000 accessible items.
Sum every single theoretical ceiling together: 55,000 + 85,000 + 30,000 + 20,000. Even under impossible 100% archival perfection across nearly half a century, the entire universe of unique historical Nigerian multiple-choice questions tops out around 170,000 to 190,000 items—and the genuinely recoverable, legible subset in existence is substantially smaller.
How Multi-Site Scraping Manufactures 3x Database Bloat
Where do an extra two hundred thousand database rows come from? They come from running automated Python scrapers across four or five Nigerian cram blogs, forum threads, and pirated PDF OCR dumps, then executing INSERT INTO questions without canonical normalization.
Consider a classic JAMB Physics projectile question asking for the maximum height of a body projected at 20 m/s at an angle of 30° to the horizontal. On Blog A, the text reads 20m/s with g = 10m/s2. On Blog B, an HTML wrapper stores <p>20 m/s</p> with g = 10 ms-2. On Forum C, an OCR scan misreads the degree symbol as 300 and swaps options B and C. On PDF Dump D, the exact same 2004 question was re-uploaded under exam year 2014.
A naive SQL DISTINCT query or raw string comparison treats all four strings as completely separate questions. Multiply that pattern across twenty years of scraped revision sites, and a 90,000-row raw scrape balloons into a 360,000-row marketing headline overnight.
x²), Greek angles (θ), and chemical subscripts (H₂SO₄) into unreadable plain-text gibberish.The Silent Killer: 14%–18% Wrong Answer Keys in Unaudited Dumps
Duplication is annoying, but poisoned answer keys destroy EdTech retention. During our longitudinal ingestion audits at ALOC Station, we measured raw error rates across legacy community booklets and scraped web repositories: between 14% and 18% of computational STEM items carried disputed, transposed, or outright wrong answer keys.
Think about what happens when a high-performing student solves a quadratic inequality correctly, selects Option C, and your app marks them wrong because a scraped database stored Option A from a 2011 typo. Worse, if your product feeds that corrupted answer key into an LLM prompt to explain the solution, the language model will confidently hallucinate bad algebra to justify the wrong letter.
Parents, schools, and paying subscribers do not blame the upstream API vendor when an app teaches wrong physics—they churn from your product and warn their classmates.
Inside ALOC Station's 4-Stage Deduplication & Maker-Checker Pipeline
At ALOC Station, we made a foundational architectural decision: we would rather publish 12,645 verified, deduplicated, syllabus-mapped questions with live public audit metrics than inflate our homepage with 350,000 unverified rows. Every question entering our Unified Ingestion Pipeline must survive four sequential quality gates before public API endpoints serve it.
Gate One runs inside DuplicateDetectorService during staging (questionImportRows), enforcing structural (examType, subject, year, questionNumber) collision checks so legacy migrations and spreadsheet batches never double-insert an existing paper slot.
Gate Two executes Stage A Exact Hash Deduplication in our FastAPI intelligence layer: we strip HTML noise, normalize whitespace, canonicalize mathematical symbols, and compute a zero-cost SHA-256 `questionHash` to reject exact duplicates across examination sources.
Gate Three runs Stage B Semantic Near-Duplicate Detection using pgvector cosine similarity over dense embeddings (QuestionEmbedding). Even when OCR variants alter unit formatting or punctuation, high-similarity vector matches flag near-clones for human inspection before enrichment credits are spent.
Gate Four is our Independent Solver & Maker-Checker Governance (`QuestionReviewService`). When our L3 pedagogical engine generates a worked step-by-step derivation, it independently computes derivedAnswer from first principles and compares it against the historical correctAnswer. Any discrepancy automatically quarantines the edit into a non-destructive QuestionReviewProposal (awaiting_checker), requiring dual-control Reviewer and Super Admin Checker sign-off before mutating live data.
questionHash) with pgvector ANN cosine similarity to catch both exact clones and OCR mutations.derivedAnswer against the stored key, exposing historical booklet typos automatically.QuestionReviewProposal diffs that cannot alter production responses until a Checker approves.Radical Honesty on Coverage—and How CTOs Should Audit Any API
Infrastructure trust comes from radical transparency, including being honest about what we are still building. Today, ALOC Station serves 8,907 JAMB, 2,256 WAEC, 1,217 Post-UTME, and an initial baseline of 267 NECO verified questions—all 100% mapped to curriculum topic nodes with calibrated difficulty scores, exposed live without authentication at aloc.com.ng/data-quality and aloc.com.ng/coverage.
Where an exam body or subject is still thin, our API never masks gaps with scraped junk or silent synthetic padding: public retrieval endpoints strictly default to contentSource=historical and reviewStatus=approved, returning structured coverage hints when a slice is still expanding.
Before signing an annual contract or wiring any assessment API into your application, run a fifteen-minute technical audit. Pull 50 random Mathematics and Chemistry items, check how many times identical stems reappear under different IDs, solve ten computational items by hand against the returned answer key, and ask whether the vendor exposes a public, unauthenticated quality telemetry endpoint.
/api/v1/metrics/quality.contentSource=historical) and reviewStatus=approved are enforced by default across REST, SDKs, and MCP tools.Combinatorial Reality Check: Theoretical Historical Ceiling vs. Scraped Claims
| Examination Body (Window) | Theoretical Max Unique MCQs | ALOC Verified Baseline (Live) |
|---|---|---|
| JAMB UTME (1978–2025, ~25 subjects) | ~50,000 – 55,000 items | 8,907 deduplicated, syllabus-mapped |
| WAEC WASSCE (1988–2025, May/June & Nov/Dec) | ~75,000 – 85,000 items | 2,256 deduplicated, syllabus-mapped |
| NECO SSCE (2000–2025, ~20 subjects) | ~30,000 items | 267 verified baseline (expanding) |
| Post-UTME (2005–2024, top 20 universities) | ~15,000 – 20,000 items | 1,217 deduplicated, syllabus-mapped |
| Combined Nigerian Historical Ceiling | ~170,000 – 190,000 max possible | 12,645 audited (0% silent bloat) |
Engineering Due-Diligence Matrix: Scraped Aggregate Feeds vs. ALOC Station
| Architecture Dimension | 350k+ Scraped Question Feeds | ALOC Verified Infrastructure |
|---|---|---|
| Deduplication Pipeline | None (2.5x–3x multi-blog OCR clones) | Composite slot + SHA-256 + pgvector ANN |
| Answer Key Governance | Raw booklet keys (14%–18% STEM error rate) | AI derivedAnswer check + Maker-Checker |
| Curriculum Metadata | Flat subject & year strings only | 100% syllabus tree + Bloom L1–L4 difficulty |
| Public Quality Telemetry | Opaque marketing banner claims | Live unauthenticated /api/v1/metrics/quality |
| Developer & AI Integration | Legacy query endpoints | TS & Python SDKs + Official MCP Server |
The Strategic Takeaway
In financial infrastructure, no CTO chooses a payment processor based on who has the largest unverified ledger table; you choose the rail with cryptographic idempotency, double-entry reconciliation, and auditable uptime. Educational assessment infrastructure deserves the exact same engineering rigor.
Vanity question counts are easy to scrape over a weekend. Canonical deduplication, independent mathematical answer verification, Bloom cognitive calibration, and dual-control Maker-Checker moderation are hard—and that is precisely why we built them into the infrastructure layer so your team never has to.
Inspect our live telemetry at aloc.com.ng/data-quality, explore 20-year syllabus weights at aloc.com.ng/benchmarks, or test verified L1–L4 payloads right now in the API Playground.
Build your assessment products on ALOC Station
Start building with 1,000 free monthly credits. Integrate our official TypeScript/Python SDKs or connect your AI agent using the Model Context Protocol.