The model steps out of the photo
Last week AR mode found its bearings. This week it started to feel like the thing we’ve been picturing all along: you reach a stop, the notice board appears in front of you, the story plays from it, and with a tap the old cottage or the tram shed steps out of the photo and stands there in front of you.
For walkers: the notice board comes alive
When a waypoint has a 3D model, its AR notice board now carries a Pop out button on the cover photo. Tap it and the model grows out of the photo towards you over about a second, then holds still at roughly waist height, so you can look down over the top of it the way you would a model in a museum case. The button then reads Pop in, and the model shrinks and flattens back into the photo. The first version had a separate button under the model. One toggle on the photo turned out simpler after actually trying it.
The board itself got a voice. Each stop’s arrival story now plays through spatial audio anchored to the board, so it sounds like it’s coming from where the board is standing, not from the phone in your hand. If you’ve turned away and the board is off screen, a small Board this way arrow at the edge points you back the shorter way round. The narration controls also moved: they used to float under the header, and now they sit as a strip on top of the dock card, naming what’s playing. All three changes came straight out of the 26 September test walk.
Outside AR, the regular follow-map now turns with you. It uses heading-up, following your GPS direction while you’re moving and the compass while you’re standing still, smoothed so it doesn’t jitter. A compass button flips it back to north-up if you’d rather. This is the one we expect families to notice first, because it’s how most people already expect a walking map to behave.
Two fixes also came from real walks. Tapping End Trail while an AR board was showing could crash the app, which was found in a crash report from the 27 September walk, and that’s fixed. And a resumed walk that already had later stops ticked off would jump to the end-of-trail screen too early. The app now works out “have I reached the end?” from where you actually are on the route, not from how many stops are marked done.
For walkers: why AR kept pointing the wrong way
Last week’s heading fix helped, but some trails still pointed noticeably off. Replaying the diagnostics logs that the app started recording last week showed the real cause. ARKit’s own sense of north was rotated by about 50, 60, and 120 degrees across three separate sessions, so every trail placed from GPS came out rotated by the same amount. The app now measures that rotation, first from the compass and then by comparing your GPS track with the camera’s own record of where you’ve walked, and places the trail on a corrected anchor.
The same logs showed that Apple’s geo-tracking could report itself as “localized” while drifting more than 200 metres from your actual GPS position. The app now checks geo-tracking against GPS, and if they disagree by more than a sensible margin, GPS wins.
For historians: wake the GPU yourself, and models with their colours
3D generation runs on a GPU box that costs real money every hour it’s on, so until now only an admin could switch it on. Now anyone can wake it: the API-status pills in Suburb Historian show whether it’s asleep, starting, switched off, or about to go to sleep, and you can wake it from there. Once it’s idle for five minutes it switches itself back off, and it never stays on for more than three hours. Admins get a three-way switch instead of an on/off toggle: Off (nobody can wake it), On demand, or Always on.
3D models also look the way they’re meant to in AR on iPhone. Generated models were showing up flat grey. The colour was in the model, but Apple’s AR renderer ignores the way it was stored. Generated models now have their colours baked into a proper texture during conversion, and every existing waypoint model in production has been regenerated. Getting there also fixed the reason Generate USDZ was rejecting every real model outright.
One frustrating bug is gone too. Replacing a waypoint’s 3D model could fail with a permission error, because a large upload took longer than the sign-in token it started with lasted. Models now upload straight to storage and are checked once they’ve landed, so a slow connection no longer costs you the upload.
For anyone asking an AI about a local walk
Story Trail’s published trails are now readable by AI assistants. There’s a public /llms.txt summary of every published trail, and a read-only MCP server that lets assistants search trails by keyword, suburb, or nearby location. When someone asks their assistant for a history walk near them, it can now find a real one, written by a real local historian, with that historian credited by name. Trails also have a clean, permanent /trail/… link, the one QR codes and app links will use from now on.
Historians get a counterpart: a signed-in MCP server for Suburb Historian. With it, you can read, edit, structure, and publish your own studies from an AI assistant, using the same rules as the web app, and see the walk score for every trail you’ve written.
Under the hood
Each stop’s facts on the three New York trails were checked against sources, which are now recorded per stop. A backfill found that 28 of 36 published trails had never had their suburb sent through to Story Trail, and four had the wrong one. All of them are corrected, and every trail update now sends its suburb along. For testing, the app also grew an AR Board Lab under Profile: tap the floor and any waypoint’s notice board appears right there, Pop out, narration, and arrow included, so the board can be checked indoors without walking a trail.