In a couple of days, HoleRecord goes up on the App Store, and I'm going to ask you to do something that requires a little faith: walk off a green, tap a gold button, say what happened, and trust that the scorecard filling itself in got it right.
That's a big ask. A scorecard you can't trust is worse than no scorecard — it's a lie you'll carry into your next practice session. So before I asked you to trust it, I wanted to earn it. Here's the work we did to make sure the thing actually keeps your round the way you played it.
We built a driving range for the AI — and then we played it
In an earlier post I described the idea of a "driving range for the AI": instead of shipping and hoping, you give the system a bucket of known shots and grade what comes back. This is the part where we actually hit the bucket.
We played five full rounds. Ninety holes. Two real, recognizable courses — Pebble Beach and TPC Sawgrass's Stadium course, island green and all. And here's the important part: every single hole had a known-correct scorecard attached to it before we ever spoke a word. The exact score. The putts. The fairway, the green, the club, the penalty. An answer key.
Then we said each hole out loud, let the app do its thing, and graded it against the truth — hole by hole, stat by stat.
Two rules made it honest. First, we recorded in a mess, on purpose — next to a running fan, outside in the wind, with a cart rattling past, with people talking in the background. Not in a silent recording booth where everything is easy. Second — and this is the part I'm proudest of — we did it iteratively. We didn't record all five rounds and then write a nice report. After each round, we studied every miss, figured out the cause, fixed it, and shipped the fix before the next round. So round four was tested against a better system than round three. The thing got better on purpose, in the open, one round at a time.
The surprise: it was almost never the microphone
Going in, I assumed the failures would be about hearing. Golf is a linguistic minefield, and I figured wind and chatter would be where it fell apart.
It mostly wasn't. Even outside, near noise, the speech-to-text held up. The words made it through. The misses were somewhere more interesting: the AI hadn't fully learned to think like a golfer yet. And those, honestly, are the fun ones to fix, because every one of them is a little rule that's obvious to us and invisible to a machine. A few real examples from the five rounds:
- It was handing out up-and-downs it hadn't earned. If you missed the green, chipped on, and two-putted, the app was calling that an up-and-down. Any golfer knows that's not a save — that's a bogey with extra steps. An up-and-down means you got down in one putt. We taught it the actual rule.
- It didn't know that "green in regulation" depends on the hole. "I hit the green in three" is a green-in-regulation on a par 5 and a miss on a par 4 — same words, opposite meaning. It wasn't checking the par. Now it does.
- It didn't trust you. When you flat-out say "double bogey seven," you should get a seven on the card — even if the app's own shot-by-shot count wandered off somewhere. Golfers know what they made. Now, when you state your score, the app believes you.
- It had blind spots on the little stuff. It had learned to fill in the club off the tee but was leaving the approach club blank entirely. It wanted to call every wedge a pitching wedge, and it choked on "nine iron" when it expected "9 iron." All cleaned up.
None of these are dramatic. That's the point. Keeping the whole record right is a hundred small, unglamorous rules, and the only way to find the ones you got wrong is to grade yourself honestly against an answer key.
The numbers moved — and that's the part that matters for trust
I'll share a couple, not to brag but because they're the actual reason I feel okay asking you to trust this.
That up-and-down mistake? It started the eval getting it right about two-thirds of the time and finished at roughly nineteen out of twenty. The approach club went from the app filling in nothing to filling in almost everything. And the final round, after all the fixes, projected to essentially a clean sheet — score, putts, penalties, and the club off the tee all correct across eighteen holes, with greens nearly perfect.
The shape of that story — start rough on a specific thing, find it, fix it, watch it climb — repeated across the whole eval. Which is exactly what you want to see. Not "it was perfect out of the box" (nobody should believe that), but "they measured the thing that matters, found the ugly numbers themselves, and drove them down."
What we're not claiming
Here's the honest part, because a results post without one isn't worth reading.
The misses that survived to the end weren't the AI misunderstanding golf. They were genuine transcription slips on the noisiest, most chaotic holes — the app mishearing a word when a fan and a conversation were both fighting the microphone. On one hole "par five" came out "this putt five." Real-world audio, in real-world noise, is the next thing we're chasing, and I'm not going to pretend we've nailed it. We've got a hedge in place — when you state your score, we trust it even if the rest of the sentence got garbled — but the deeper work is still ahead.
I'd rather tell you that than post a hero number from an easy test. A great score on the wrong test is just a comfortable lie. We'd rather find the ugly numbers ourselves, on the range, than have you find them on the 14th.
Why we did this before launch
The name is the whole promise: HoleRecord is about keeping the whole record of your round — the story, hole by hole, not just a number on a card headed for the laundry. But that promise is only worth something if the record is right. A round you can't trust isn't a record. It's noise with a nice font.
So we graded ourselves before we asked you to, because you're about to hand us your rounds, and that trust has to be earned rather than assumed. Ninety holes, an answer key, messy conditions, and a fix after every round — that's the homework we did before we put it in your pocket.
It launches this week. When you speak your first hole, I want you to be able to glance at the scorecard, see it's right, and just keep walking. That's the entire job. We did the work so it can disappear.