# Format specification Everything below was read out of the reference `.docx` by unzipping it and inspecting the XML. Nothing here is inferred from how the document looks in a viewer. Line references are to the reference document's `word/document.xml` (141 paragraphs, 259 runs, 13 distinct paragraph shapes). ```bash unzip -q reference.docx -d unpacked/ ``` The package has 11 parts and nothing else — no headers, footers, images, tables, text boxes, bookmarks, comments or content controls: ``` [Content_Types].xml _rels/.rels docProps/app.xml docProps/core.xml word/document.xml word/_rels/document.xml.rels word/fontTable.xml word/numbering.xml word/settings.xml word/styles.xml word/theme/theme1.xml word/webSettings.xml ``` --- ## 1. Page setup ```xml ``` A4 (`w:code="9"`), 0.5″ margins all round. Text area is 7.27″ × 10.69″. US Letter is `w:w="12240" w:h="15840"` with `w:code="1"` — the `--page-size` flag swaps exactly these three attributes and nothing else. ## 2. Style layer ### docDefaults (`styles.xml`) ```xml ``` `kern`, `ligatures` and the theme-font mapping are inherited by everything and are easy to miss. They are one of the main reasons the template-clone strategy was chosen over generating a document from scratch. ### Normal ```xml ``` **`Normal` is 11 pt, not 10 pt.** It also carries `w:lang="en-IN"`. The 10 pt body size exists *only* as direct run formatting on every single run. A run emitted without an explicit `` silently renders at 11 pt in the theme font (Aptos), not 10 pt Bookman. `validate.py` checks that every run pins its own `rFonts` and `sz` for exactly this reason. `spacing after="0" line="240" lineRule="auto"` = single-spaced, no gap between paragraphs. Vertical rhythm in this resume comes entirely from deliberately placed empty paragraphs (§5). ### ListParagraph ```xml ``` `contextualSpacing` on bullets is **inherited from this style**, not written on each paragraph. The `ind left="720"` is overridden per-paragraph to `360`. ### Hyperlink ```xml ``` ### Theme `theme1.xml` sets major = **Aptos Display**, minor = **Aptos** (the Word 2024 defaults). Body runs use `w:cstheme="minorHAnsi"`, so complex-script fallback resolves to Aptos while Latin text is Bookman Old Style. ## 3. Numbering ```xml ...levels 1-8 tentative... ``` The glyph is real list numbering, not a typed character. **No `•` ever appears inside a `` element** — `validate.py` asserts this, because a typed bullet both breaks list semantics and confuses ATS parsers. ## 4. The paragraph kinds XML samples below show the real markup with **fictional placeholder text** substituted for the reference author's details. Every run in the document carries the same font block: ```xml ``` `rPr` children must appear in schema order: `rStyle, rFonts, b, bCs, sz, szCs, u`. `pPr` children likewise: `pStyle, numPr, ind, contextualSpacing, jc, rPr`. ### 4.1 Name and target title (paragraphs 0, 1) ```xml {FONT}{SIZE} {FONT}{SIZE}Jordan Rivera ``` ### 4.2 Contact line (paragraph 2) Centered and bold like the two above, but the email is wrapped in a hyperlink: ```xml {FONT}{SIZE} jordan.rivera@example.com {FONT}{SIZE} | +1 555-0100 ``` with, in `word/_rels/document.xml.rels`: ```xml ``` Only the email is linked. A LinkedIn URL or city on the same line stays plain text, matching the reference. When the input has no email, the generator drops the `rId5` relationship entirely rather than emitting an empty target. ### 4.3 Section headers (paragraphs 4, 22, 32, 138) ```xml {FONT}{SIZE} {FONT}{SIZE} PROFESSIONAL SUMMARY: ``` ALL CAPS, trailing colon, bold, single underline, left-aligned, no indent. > **Reference inconsistency.** `EDUCATION:` (paragraph 138) additionally carries > ``; the other three do not. On a single line that never > wraps, justification is indistinguishable from left alignment. The generator > emits the 3-of-4 majority form for all four. ### 4.4 Bullets (107 paragraphs) ```xml {FONT}{SIZE} {FONT}{SIZE}… ``` Regular weight — **zero of the 107 bullets contains a bold run**. The per-paragraph `ind left="360"` overrides the style's `720`, while the abstract level's `hanging="360"` still applies, so the glyph sits at 0″ and text at 0.25″. ### 4.5 Skills lines (8 paragraphs) — not bullets ```xml {FONT}{SIZE} {FONT}{SIZE}Programming & Data: {FONT}{SIZE} Python, SQL, Pandas, … ``` Bold label including the colon, then the separating space at the head of the regular run. > **Reference inconsistency.** 7 of 8 skills lines put the separating space in > the regular run; the 8th (`AI Platforms, Observability & Delivery: `) puts it > inside the bold run. A bold space is fractionally wider, so that one line's > justification differs by well under a pixel. The generator uses the 7-of-8 > majority form. ### 4.6 Job blocks Three bold lines, then bullets: | Line | `pPr` rPr | Run | |---|---|---| | `Company, City ST \| Month Year – Month Year` | *not* bold | bold | | `Job Title` | **bold** | bold | | `Responsibilities:` | *not* bold | bold | The paragraph-mark formatting differs between line 2 and lines 1/3. It affects only the ¶ glyph, never rendered text; the generator reproduces it anyway because it costs nothing. All four date ranges use **EN DASH U+2013** with a space either side. The whole document contains exactly 5 non-ASCII characters: 5 × U+2013, plus 2 × U+00A0 (see §7). ### 4.7 Blank paragraphs — the vertical rhythm Seven empty paragraphs, and their placement is the entire spacing system: - exactly one **before each of the 4 section headers** (¶3, ¶21, ¶31, ¶137) - exactly one **before each job block after the first** (¶64, ¶93, ¶120) - **none** between a header and its first content line - **none** at the end of the document > **Reference inconsistency.** Those 7 paragraphs have 5 different `pPr`, each > having inherited whatever the preceding block's paragraph mark carried — > `jc=center`+bold, `ListParagraph`+`ind 360`+`jc=both`, `contextualSpacing`, > plain (×3), and `jc=both`. All seven render as an identical empty 10 pt line > because `Normal` sets `after="0"` and none of those properties affects an > empty paragraph's height. This is Word editing residue, not design. The > generator emits one canonical form (the 3-of-7 plurality: `pPr` with only the > font/size run-mark properties). ## 5. settings.xml ``, `compatibilityMode = 15`, `themeFontLang en-US`, `characterSpacingControl doNotCompress`. No tabs are used anywhere in the body — the 259 runs contain 259 `` elements and zero `` or ``. ## 6. fontTable.xml Declares Symbol, Times New Roman, Courier New, Wingdings, Aptos, Aptos Display and: ```xml ``` The panose and signature are what let Word pick a sane substitute on a machine without the font. `validate.py` asserts every `w:ascii` used in the body is declared here, so `--font Georgia` would be flagged if Georgia were missing. ## 7. Content defects in the reference (not formatting) - ¶5 ends `…across the USA and India.. ` — a doubled period and a trailing space, from two runs each contributing a `.` - ¶50 and ¶54 end with a trailing **U+00A0** non-breaking space The parser strips trailing whitespace including NBSP, so these three vanish on round-trip. The doubled period survives (it is content, and the skill does not edit content). ## 8. What was stripped to make `assets/template/` Performed once by `reference/make-template.py`: | Removed | From | Why | |---|---|---| | `w:rsid*` attributes (~530) | document, settings, styles | editing-session fingerprints | | `w14:paraId` / `w14:textId` (282) | document | per-paragraph ids | | `` (116) | document | spell/grammar markers | | `` (4) | document | stale pagination cache | | ``, `w14:docId`, `w15:docId` GUID, `` | settings | document identity | | `dc:creator`, `cp:lastModifiedBy`, `cp:revision`, timestamps | core.xml | authorship | | `Pages/Words/Characters/Lines/Paragraphs/TotalTime/Company` | app.xml | stale statistics | | the entire `` content | document | replaced by a `{{BODY}}` token | | `Target="mailto:…"` | document.xml.rels | replaced by a `{{MAILTO}}` token | `styles.xml` (all 376 latent style entries), `numbering.xml`, `fontTable.xml`, `theme1.xml`, `webSettings.xml`, `[Content_Types].xml` and `_rels/.rels` are otherwise byte-identical to the original. ## 9. Fidelity evidence Extracting the reference back to plain text and rebuilding it produces a body XML diff of **7 hunks in 3,215 lines**, every one of them in the documented invisible set: 4 blank-paragraph `pPr` variants, the stray `jc` on `EDUCATION:`, and 3 trailing-whitespace normalizations. Paragraph count matches (141 = 141) and paragraph text matches in order. Rendered through Microsoft Word to PDF on a machine with Bookman Old Style installed, `pdftotext -layout` output of the original and the rebuild is **identical — 289 lines, zero differences**: every line wraps at the same word and both documents are 5 pages. A raster diff at 100 dpi shows 2,339 of 4,837,950 pixels differing (0.048%), max delta 65/255, all of it anti-aliasing around the one bold-space skills line described in §4.5.