Why Our Resume Rewrite Can Promise Not to Fabricate

In July we ran a test: four real Chinese-written resume bullets, three models, same task. The short version: one model produced clean output in 3 to 5 seconds without inventing facts; another inflated "participated in" into "led"; a third had a coin-flip chance of returning nothing at all.

The most common question after that test: "How does your own product avoid the same mistake?" This post is the answer. We'll walk the pipeline step by step and show you how to check it yourself. You don't have to believe us.

The deadliest error isn't grammar. It's quiet inflation

Grammar is the surface layer; one pass cleans it. The dangerous error is different: the output reads like a textbook US resume bullet, but the details aren't what you wrote. It takes three usual forms:

• Verb upgrade: your "participated in developing" comes back as "Led the development of...";
• Number appearance: your original has no metrics, the output has "improving performance by 40%";
• Scope expansion: you were one member of a team, the output reads like you owned the whole module.

Resumes like that can survive a first screen. The first follow-up question in the interview exposes them, and the background check ends it. For a job seeker this isn't "over-polishing." It's the kind of hole that blows up an offer.

Why models inflate on their own

This isn't one model's bug. It's all of them. In the July test, all three models did it, just to different degrees. The reason is simple: the training objective is "write like a resume bullet," not "write like you."

Feed it "participated in part of a system's development" and the high-probability continuation is "Led the development of...": because in the internet's English resume corpus, that's what tends to follow. The model has no "stay faithful to the source" objective unless you give it one explicitly.

So the right engineering move isn't "find a smarter model." It's changing the process: make faithfulness an independent, checkable step.

Our approach: two passes, the second one only checks

US Tech Resume Rewrite runs twice, not once.

Pass one: rewrite. Chinese or Chinglish experience becomes American English bullets. This pass has one goal: sound native. It may restructure sentences, swap synonyms, reorder clauses. It may not add facts your original doesn't contain. That constraint doesn't rely on the model being well-behaved. Pass two guarantees it.

Pass two: check, line by line. After the rewrite, a second AI compares every output bullet against your original on a fixed four-item list:

• Numbers: can every number, percentage, and amount in the output be found in your original? If not, it gets deleted, not swapped for another number;
• Verb strength: your original says "participated / assisted / supported"; has the output been upgraded to "Led / Owned / Drove"? It gets dialed back to match your original;
• Scope: has the responsibility boundary you described been widened? It gets pulled back;
• Titles and tools: are titles, companies, and technology names the same as your original? They get restored to your wording.

Any bullet that fails a check gets revised or rewritten, not passed through. The version that reaches you has already been through one round of nitpicking. That's why our product page can say "Never fabricates numbers or upgrades participated into Led." It isn't a line in a prompt. It's a step with explicit checks.

One limit, stated plainly: the check catches "doesn't match the source," but if your original is vague ("responsible for some optimizations"), all we can do is faithfully preserve the vague. Sharpening a fuzzy original is a different job, and we won't decide for you what you actually did back then.

How to verify it yourself: one free bullet

You don't have to believe the step exists. The product has a verification door built in: try one bullet free. Here's how we'd test it:

• Paste the most concrete bullet in your resume, the one with real numbers and clear actions;
• When the output comes back, don't read it for flow. Check item by item: is every number in your original? Did verb strength go up? Did scope get bigger?
• Then deliberately paste a "participated-style" original and see whether the output upgrades it.

The third step is the real test. Inflation happens exactly on "participated-style" experience. If the output contains a number your original never had, or an upgraded verb, that's a bug. Tell us at Contact and we'll fix it.

What we don't promise

To finish the thought: we don't promise interviews, we don't promise offers. No resume service can, and the ones that promise are lying. We promise the tool's part: faithful to your original, only the expression gets Americanized, and every line is checkable. AI output is a starting point. Read it again yourself before you submit. That's our position on every product.