Methodology
How the questions are sourced, how answers are scored, and which pages we allow to be indexed. Every number below is measured against the live question bank by tools/gen_bank_integrity.py, not written by hand. Last measured 2026-08-14.
Where the questions come from
The bank holds 86,616 questions across 29 subjects and 510 topics. Most of it was generated in bulk by a language model and then imported; the Machine Learning and LLM banks were written by hand, one question at a time, because those subjects punish the plausible-but-wrong answers bulk generation produces. We say which is which on our editorial policy page rather than implying every question was expert-authored.
Every question carries an explanation of why the correct answer is correct — 100% coverage, with no exceptions. A question without an explanation teaches nothing, so it does not ship.
What we measured, and what we did about it
Every row is now a different question — 86,616 rows, 86,616 distinct. It was not always: the bank previously held 95,290 entries, of which 18,674 were the same question stored more than once. One source file alone contained 1,488 questions repeated up to fifteen times each, with the answer options shuffled so the copies did not look identical. Those were removed at the source, not just from the database, so a re-import cannot bring them back. All29 subjects and 510 topics survive.
Two measurements are easy to get wrong here, in opposite directions. Compare options exactly as stored and almost nothing looks duplicated — 86,605 of 86,616 — because the order is shuffled per row. Strip the leading scenario clause before comparing and you over-merge instead: “error handling in Bash/Shell” and “error handling in C#” collapse into one, and thousands of real questions look like duplicates. We report the measure in between, and the stem-template count separately, because it answers a different question — how varied the phrasing is, not how many questions there are.
Correction. An earlier version of this page reported the stem-template count as the number of distinct questions, which understated the bank by several thousand. The generator now measures and labels the two separately. The headline count also dropped from 95,290 to 86,616 when the duplicates above were removed — a smaller number describing exactly the same amount of actual content.
How answers are scored
Spoken interviews are scored against a versioned rubric — currently 2026.08-elite. Versioning is not cosmetic: a score is only defensible if the exact rubric used to produce it can be recovered later, so every stored score records the rubric version alongside it. Scores written under an older version are on a different scale and are not comparable.
Four rubrics cover the interview modes — technical/coding, HR and behavioural (STAR-based), mixed, and system design — each with five dimensions scored 0–10 against six named anchors, from “no real signal” at 2 to “principal-grade depth” at 10.
Why a score must point at something you said
Before any model scores an answer, a deterministic pass grades the evidence in it across seven dimensions — technical correctness, depth, specificity, problem solving, ownership, communication and trade-off awareness — into one of four grades:
- • Observed — the answer contains the evidence directly
- • Inferred — reasonable to conclude, but not actually stated
- • Insufficient evidence — they spoke, but nothing gradeable came out
- • Not tested — the competency never came up
That pass runs without any model call, which means it cannot invent a quotation, it works with no API key configured, and it gives the model a floor it has to justify departing from. The distinction between demonstrated and claimed in your report comes from here. Where a competency is graded not tested, the report says so instead of guessing — and the overall recommendation is withheld when the evidence does not support one.
The overall score is computed from the category scores rather than assigned by the model, and maps to fixed bands: 8.5 and above is a strong hire, 7.0 a hire, 5.5 borderline, below that a no-hire. Those bands are advisory practice feedback. They are not a hiring decision and no employer receives them.
Which pages we let search engines index
A bank of 86,616 questions could be turned into 86,616 thin pages. We did not do that, and will not. Topic hub pages are generated only for subject-topic-difficulty combinations backed by at least 15 distinct questions — counted after deduplication, so the gate cannot be passed by repeats. Combinations that fail the gate still work if you visit them; they are simply marked noindex rather than being added to the sitemap.
Of the combinations that pass, the strongest 100 are published in the first wave. Staging the rollout means a page that turns out to be weak can be caught before it has company.
Corrections
If you find a question with a wrong answer, a misleading explanation, or a duplicate, tell us through feedback and include the question text. Corrections are made to the bank itself, so the fix reaches every page the question appears on. When a figure on this page changes, the generator is re-run and the date at the top moves with it.
Related: what 86,616 interview questions reveal · editorial policy