Materials · corvid/reviews/round2-visual.md

# Round 2: visual review (Corvid Labs, Research Scientist, Agent Evaluation)

Sources: the three rendered PNGs, the posting, and the "Application mechanism" section of the packet. All parsing predictions below come from the images, not the PDF text layer.

## Ranked problems (most damaging first)

1. **The job header lines are hard to parse and hard to scan.** Each job is one bold, comma-separated line: "Title, Employer (descriptor), [month] YYYY to [month] YYYY". The packet says dates and titles must be machine-readable.
   - "Research Lead, Measurement, Kestrel Media (streaming service)" has a comma inside the title. A parser will likely split it as title "Research Lead" and employer "Measurement".
   - "Independent Researcher" has no employer, so the date string may end up in the employer field.
   - "Postdoctoral Fellow, Ridgeview University Decision Lab" may be pulled into Education because of the word "University".
   - The parentheticals "(streaming service)" and "(consumer delivery app)" will probably end up in the employer names.
   - Dates use "to" rather than a dash. They sit at the end of a line that wraps, so the end date can drop onto a second line (Kestrel, Parcel and the postdoc all wrap now). Once the months are filled, most of these lines should fit on one line, but the dates are still not right-aligned, so a reader cannot scan the timeline.
   - **Fix:** put title, employer and dates in separate fields, for example "Research Lead, Measurement" on one line, then "Kestrel Media | Mar 2022 – Jun 2024". Drop the parentheticals or move them into a bullet. Use a consistent "Mon YYYY – Mon YYYY" format.

2. **Placeholders sit in fields the parser needs.** The email, phone and location (on both documents) and every start/end month are all still [ASK JORDAN]. If any survive into the upload, the portal record has no email or phone, and its dates are year-only or garbled. This is a hard blocker before submission, not a layout issue, but it is the one that would sink the application.

3. **The posting's must-haves are buried on page 2, and page 1 opens on the weakest entry.**
   - The PhD (requirement 1) and Python (requirement 2) appear only at the bottom of page 2, in Education and Skills.
   - The first entry under Experience is "Independent Researcher", with one thin bullet ("Self-funded independent research...").
   - The strongest evidence for this role is Relay: validating LLM judges against 300 hand labels, with agreement differing by property. It sits mid-page under Projects.
   - **Fix:** move Education (one line) up under the contact line, or add a short line covering the PhD and Python near the top. Fold the Relay bullets into the Independent Researcher entry, or put Projects first.

4. **Projects entries look like jobs to a parser.** Relay and the "Pricing-experiment rehearsal" have bold, dated header lines in the same format as jobs. A parser will likely create two extra positions with "Relay..." and "Pricing-experiment rehearsal (simulated)" as job titles. Relay also overlaps in dates with Independent Researcher, which can count as double or overlapping employment. Non-standard headings also make sections harder to detect: use "Publications" rather than "Papers", and merge the single-item "Other work" into Talks or Projects.

5. **The heading hierarchy is weak.** The name stands out. Section headings are only slightly larger than the bold entry lines, so Experience, Projects and Papers do not separate clearly from the job lines under them. Title, employer and dates are all the same weight inside one bold run, so nothing distinguishes them. A rule under section headings, or more space above them, plus a regular-weight employer/date line, would fix this.

6. **The publications don't follow the usual format, and links are missing.**
   - Jordan's name is not bolded in author lists.
   - The status words "Published. First author." / "Third of four authors." add clutter. Keep the role note only where it says something, like the measurement-model contribution.
   - There are no DOIs or URLs. "Posted on jordanhale.example" is not a link.
   - For a role that requires "strong Python", the absence of a GitHub or code link is noticeable. If Relay code or the walkthrough is public, link it in the contact line.

7. **Page 2 is about two-thirds empty, with no running header.** Only about 35% of page 2 is used, and there is no name or page number on it. Two pages is acceptable for a research scientist with publications, so the issue isn't length. The issue is that the page-1 overflow happens to hold the must-haves (see 3). After reordering, either balance the two pages or tighten to about 1.5 pages. Add "Jordan Hale – page 2".

8. **Legibility is acceptable but dense.**
   - Serif type at roughly 11 pt, with margins of about 0.9–1 in.
   - Lines run long, about 95–105 characters, so both documents read as dense.
   - Page 1 of the résumé is full to the bottom margin with little space between entries.
   - There are short widows: "decision.", "session.", "Published. First author.", the wrapped date fragments, and "models." in the letter.
   - Slightly wider margins or a shorter measure would help if the space freed in item 7 is used.

9. **The cover letter is the right shape, with one heavy paragraph.** It reads as a letter: name and contact block, date, recipient, "Re:" line, salutation, six paragraphs, sign-off, and it fills about 80% of one page. Paragraph 2 (Relay) is about 11 lines now and will reach 12–13 once its placeholder is filled. It is the visual centre of the page and the hardest block to read. Splitting it (the judge validation, then the configuration error and the paper) would help. The other placeholder (paragraph 4) adds a line or two. The page still fits after both are filled.

## By question

**1. First scan.**
- *Letter:* the eye goes to the name, then the dense Relay paragraph. The first paragraph names Corvid's March blog post and the grader-versus-rater problem, which is the right hook for this team, and it's visible at once.
- *Résumé p1:* the eye goes to the name, then the bold job lines, which read as product-analytics and organizational-behaviour roles. LLM-judge validation only registers if the reader gets to Projects. PhD and Python don't register at all. That is not what should register first for this role.
- *Résumé p2:* the eye goes to the white space, then Education and Skills, which should have been seen first.

**2. Hierarchy.** The name is clear. Section headings are only slightly differentiated from entry lines. Title, employer and dates are not distinguishable from each other. Body text and bullets are clean.

**3. Legibility.** See items 7 and 8 above. Type size is fine, lines are long, and density is high on page 1 and low on page 2. The placeholders currently cause extra line wraps in four header lines.

**4. Conventions.**
- Section order is wrong for a research hire, which usually leads with education and research; see item 3.
- Publications need the usual formatting and links; see item 6.
- Two pages is acceptable.
- The letter is conventional and appropriately short. The posting says the research team reads every letter, so its quality matters more than usual.

**5. Applicant-tracking parsing (predicted from the images).**
- *Name:* parsed correctly; it is the large first line.
- *Contact:* a single line separated by pipes. Once filled, email and phone usually parse. "jordanhale.example" without https:// may not be recognised as a URL. If unfilled, contact is lost.
- *Titles and employers:* see item 1. Misassignment is likely for Kestrel and Independent Researcher, and possible for the postdoc. Projects will likely become phantom jobs.
- *Dates:* "Month YYYY to Month YYYY" usually parses, but year-only or placeholder months will not. Wrapped date lines risk losing the end date.
- *Education:* "PhD, Organizational Behavior, Ridgeview University, 2016" and the BA line should parse into degree, field, school and year. The dissertation bullet may be dropped or appended.
- *Skills:* the standard heading and comma list should parse well.
- *Publications and talks:* typically kept as unstructured text or dropped. Low risk, but "Papers" may not be recognised as a section.
- *Overall:* the layout is single-column, with no tables, columns or text boxes visible, which is the main thing working in favour of parsing. I can't confirm from the images that the PDF has a real text layer or that its reading order matches what's visible.