Untitled-2: The Leave Index

I built The Leave Index, a database that grades companies on how much paternity leave they actually give dads.

Every company offering parental leave gives something to the birthing parent, and most give something to the dad. The gap between those two numbers says more about how a company thinks about fathers than any careers page will. On the site, each company gets a 52-week strip of the year after a birth — what the employer pays, what the state pays, what nobody pays — plus a letter grade, a confidence badge, and the date the number was last checked. There's a state selector for the thirteen paid-leave programs and a calculator running Minnesota's 2026 formula, because that's where I live.

Most dads don't take anything close to the time they're allowed. Fathers on unpaid leave are 58% more likely to report anxiety than fathers on paid leave, and the dads who fare worst are the ones who wanted leave and couldn't take it. Money is the barrier they'll say out loud. Stigma is the one they won't.

Both run on the same belief, which is that you'd be the only one. But the idea is that if you can see what your employer offers, and see that it's unremarkable among its peers, taking it stops being a statement and starts being a normal thing that normal companies do.

The site is a single HTML file on Netlify, with no backend, database, or LLM anywhere in the finished product. Every grade is computed in the browser from data hardcoded in the page.

The hardest part was that a lot of this information simply isn't public. Some companies publish their policy plainly, and those took a minute. Plenty don't. I tracked 188 companies and could only find a real number for 113. This lack of transparency is the problem the site exists to fix, and it ironically showed up as a hole in the site itself.

The data I couldn't find is scattered across offer letters, benefits PDFs, and Glassdoor threads, which is exactly the kind of thing an AI will hand you with total confidence and no source. Ask for a mid-size company's policy and you'll get a number that's correctly formatted, entirely plausible, and possibly from 2019. Trust me, I did that several times throughout this process.

So the fix was editorial. Every company carries a confidence badge (Verified, Reported, Employee-reported, Not disclosed) and a last-checked date that turns amber at 18 months and red at 30 months. It doesn't make the data right, but it at least makes it clear what I could and couldn't find.

As someone who grew up with binders of baseball cards, I thought it'd be cool for the company cards to look like baseball cards. Ultimately that proved too hard to scan quickly, so I pivoted to the strips.

Two bugs, and neither one broke anything.

First, the grades here are relative: your letter comes from where you fall against every other company in the set, which makes the median the entire grading system. That median was being calculated with zero-week companies included, dragging the middle down, so every company above it was getting a better grade than it earned. Separately, the wage calculator was returning replacement figures above 100% for some inputs, essentially telling fathers-to-be they'd earn more on leave than at work (which is not true).

The other surprise was the "Not disclosed" badge. I built it expecting an edge case and it ended up on 75 companies, which reframed the whole project for me. I set out to compare companies against each other. What I mostly documented is how many of them won't tell you anything at all, and how hard it is to find what they do say.

The thing I'd do differently is obvious in hindsight: hardcoding the data was right at company 10 and a tax by company 40, because every new company means editing code and shipping a deploy. The fix is moving the data into an actual database so adding a company is a form entry instead of a code change.

The thing worth stealing is the confidence badge. If you're publishing facts about named companies, put the uncertainty in the interface, not the footnotes. It cost an afternoon and changed what the site was claiming.

And check the numbers nobody looks at. The bugs that hurt weren't the ones that errored — they were a median and a percentage that rendered beautifully and were wrong.

One more, from the baseball cards: Claude would have generated card designs until I stopped asking. It never once mentioned that the concept was fighting the content. Knowing when to kill an idea rather than iterate on it is still on me.

Find out what your company gives dads at theleaveindex.com.

If your employer is one of the 75 that won't say, reply and tell me what you actually got. That's the number I really want.