About NumberSmith

A wrong answer key is the single worst failure mode for a practice platform. It teaches the wrong method, and it does it with total confidence. So NumberSmith does not rely on an author double-checking their own work.

The 109 hand-written problems are checked by a script that independently re-solves each one from scratch — by brute-force search wherever the problem allows one — and compares its own result against the stored answer key. The script never reads the stored answer to decide what the answer should be. For multiple-choice problems it first resolves the stored letter through the choices list, so a correct value paired with the wrong letter still fails.

The generated problems work the other way round: no answer is ever authored, so there is nothing for the key to disagree with. Each template draws its numbers, computes the answer from them, and then recomputes it a second time by a deliberately different route — a closed-form formula checked against a brute-force search, or vice versa. Any instance where the two disagree is discarded rather than published. Both mechanisms run in the same command:

$ npm run verify:answers

verified:  109/109
mismatches: 0

--- generated bank ---
generators:        91
generated:         7,668
answer mismatches: 0
structural faults: 0

The check fails loudly if a hand-written problem lacks an independent verification, so one cannot be added without a check being written for it. The structural pass separately catches what a per-problem check cannot: duplicate options, an answer letter with no corresponding choice, missing hints, or a difficulty outside the scale.

This catches mathematical errors in answer keys. It does not certify that every hint and worked solution is perfectly worded — those are reviewed by hand. If you find an error of any kind, please report it to supportnumbersmith@gmail.com and it will be corrected.

Competition problems are copyrighted by the organizations that write them. NumberSmith does not scrape or republish AMC, MATHCOUNTS, Math Kangaroo, HMMT, or any other organization's archives.

Every problem is written specifically for NumberSmith, and each one carries its source, license, and author as columns in the database rather than as an assumption. All 7777 currently published problems are recorded as NumberSmith Original.

The training tracks are modeled on each contest's published format — its topic mix, difficulty curve, timing, and answer format. NumberSmith is not affiliated with, endorsed by, or sponsored by any competition organization, and practicing here confers no standing in any actual contest.

Nothing that decides your level, your rating, or your training plan involves a language model. Every one of those numbers comes from a fixed algorithm you could work out on paper, which means the same performance always produces the same result.

• Placement starts near the middle of the difficulty range and steps up after correct answers and down after incorrect ones, accelerating on streaks. It stops once your recent difficulty range has settled, which is usually well before the maximum question count.
• Your rating is computed from the average difficulty you settled at, adjusted by your accuracy — an Elo-inspired scale that then moves with every problem you solve afterward.
• Topic mastery uses an exponential moving average, so recent work counts for more than work from months ago without ever fully erasing your history.
• Your training plan ranks topics by how weak you are in them weighted by how much they matter to the competitions you chose, with your top-priority contest weighted double. A student training for AMC 8 and a student training for AIME get genuinely different plans from identical mastery scores.

One deliberate consequence: your skill breakdown only shows domains the placement test actually tested you on. Untested topics are left out rather than filled in with a plausible-looking percentage.

Most of what makes a study tool trustworthy is what it refuses to say. For the record:

• No claim that it raises contest scores. No such study has been done, so no such number is quoted anywhere on this site.
• No user counts, testimonials, or star ratings. There are not enough users yet for any of those to be honest.
• No claim of affiliation with, or endorsement by, any competition organization.
• No guarantee of qualification for any contest, and no prediction of your contest result.
• No representation that the problem library covers every topic on any syllabus. It is growing.
• No sale of your data, and no advertising network. The Privacy Policy states this in binding terms.

NumberSmith is built and maintained by Aarit Chakraborty, operating as a sole proprietorship in the State of California, United States. It is not a company with a support department, and this page is not going to pretend otherwise.

Practically, that means support email goes to a real person and is answered within about 2 business days, and that bug reports get read. It also means NumberSmith moves at the pace of one person, which is the honest trade-off for everything above.

Questions, corrections, and mathematical disputes are all welcome at supportnumbersmith@gmail.com.