How the checks work, and where they stop.

• PDF: opened with PDF.js and turned into positioned pieces of text.
• DOCX: read from the document's own XML.
• Plain text and Markdown: read as they are.

Import is best-effort. Unusual headings, graphics, scans, complex tables and multi-column layouts can come in incomplete or grouped wrongly. Parsing vendor Textkernel likewise separates text PDFs from scans that need OCR, and asks for the original document (its file-format guidance).

• −26: no number or scope; a weak opener such as “responsible for”.
• −12: filler phrases, vague amounts such as “various”, passive voice, over 34 words.
• −5: not opening on an action verb, first person, under 6 words, a list of tools with no result.

The overall score leans on evidence: the bullets' average, plus the share of bullets that show a scale or a result, then structure and the recruiter screen. A rule that fires more than once, such as dates missing on six old roles, costs at most a quarter more than it does once. A phrase pasted onto three or more bullets is judged as if it weren't there, and a repeated bullet or role counts for nothing. An empty bullet isn't scored at all: no file prints it.

Some things cap the score instead of costing points, so more words or keywords can't outweigh them: fewer than four bullets that show a scale or a result (70, or 55 under two), a claim that reads as impossible (70, or 45 for two), a blank left in brackets in your title, summary, skills or education (70; a blank in a bullet scores as if it weren't there until you fill it), a summary holding contact details or written at over 90 words (65), and no number anywhere in the bullets (84, below ATS-ready). A summary that an import ran into another section is pointed out but not marked down. A generic ending such as “improving efficiency by 12%” earns nothing. Below 150 distinct scored words, the score is provisional: quality is scaled to at most 64 because there is not enough content to judge. Repeated material does not clear that limit. Concise resumes of 150–249 words have no length ceiling.

With a job post in Job match, the score is the fit for that job: half for the posting's requirements shown on the page (a bullet counts more than a skills list, and a keyword list counts nothing), 30% for evidence and 20% for readability. Two missing must-haves cap it at 55, and a missing licence, registration or right to work at 40. On a benchmark of 156 resume and posting pairs in 12 fields, labelled by hand before scoring, it ranks them in close to the same order (Spearman 0.86, and 0.86 on the six fields held out while the rules were written).

A higher score isn't a promise that a recruiter will prefer the writing. Not every bullet needs a number, but one count, size or time a reader can check is the bar for ATS-ready. Keep claims you can defend; never invent one.

• That a usable text layer exists.
• Name, email, phone, headings and bullets, and their order.
• Encoding, document metadata and page count.
• The share of your words found in what came back.

The export keeps images, text boxes and tables out of resume content, and sets metadata, a language and left-to-right reading. It isn't a fully tagged accessible PDF, but it follows the rule the W3C and the PDF Association describe: keep the text, and keep it in order.

Which layouts break, from 108 generated PDFs: the research, with its data.