Find an empty classroom near you at Ohio State, free for as long as you need it.
It is a web page. No account, no search box, nothing to sign up for. It finds you, asks one question, and hands you one room you can walk to, with a photograph of it. Bin it and it hands you the next one. Take it and it shows you the way.
 |
 |
| One question. | One answer. |
On weekends, the same room search is available from 7am to 11pm even though few classes meet. Vacant still excludes buildings the Registrar publishes as closed and checks registered events when that week's snapshot is available. An empty class timetable does not mean every door is open; the app says when event coverage is missing.
Every other tool answers "is this room free right now". That is the wrong question, because the room is not where you are standing.
Vacant answers how long the room is yours once you get there. It takes the gap in the room's schedule, subtracts the walk, and leaves ten minutes at the end so you are not packing up while the next class files in. The walk runs over Ohio State's own sidewalks, from where you are standing to the nearest door the university publishes for the building, at 78 metres a minute.
Both halves of that are newer than the rest. The building table gives a polygon centroid, so until the doors landed every walk ended inside a wall, a median 23 m short. And until the pavement landed the walk itself was a straight line multiplied by 1.3, one constant applied everywhere, which cannot see a river, a stadium, a rail corridor or a fence.
Measured over 4,574 walks from 169 standing points, that constant understated one walk in five by a minute or more, and one in twenty-one by two or more. It overstated about as many, which is the point: the error is not a bias a better constant could absorb, it is spread, and a single multiplier cannot reorder anything.
The worst of them are all the same shape. Stand on the west bank and the old model quoted nine minutes to Animal Science, because it walked you across the Olentangy. It is sixteen. The bridge is the whole difference.
Read the third row of the list further down. PAES A111 says free till 4:00pm. The next class in that room starts at 4:10pm, you are four minutes away, and 4:00pm is when you have to be packed up.
Take it and the room screen says the same minute, then spells the rest out: Yours for 5h36 once you get there. The two screens agreeing is the point. They did not for a while: the room screen printed the raw class start, handing back the ten minutes the row had already taken off (#77).
A card that says no class rest of today instead of a time means no class is coming at all, and the timeline behind it ends where the building locks.
Drag the card and it tells you what letting go will do. Each badge sits on the corner the card is not leaving by, so the word is still on screen at the moment you commit to it.
 |
 |
| Left: the next room, nearest first. | Right: this one, and the way there. |
The buttons under the card do the same two things, and so do the left and right
arrow keys. A gesture nothing announces is unreachable from a keyboard and
invisible to a screen reader. They were also, until scripts/shoot.mjs tried to
press one, doing nothing at all: the card takes the pointer capture to follow a
drag, and pointer capture retargets the click onto the capturing element, so
every press of the bin went to the card and was swallowed. A hundred and twenty
presses left the first room on screen. The card steps aside for a control now,
the way the sheet already did for its handle and its search field.
Throw the card down and you are back at the question. There is no back arrow on this screen: it was a third piece of chrome on a photograph, and the duration is one tap to set again. The gesture is not printed anywhere either, because a label is one more thing to read on a screen whose whole argument is that one room is the answer -- but it is in the card's accessible name, where a screen reader reads it out, since a gesture nothing announces is not learnable at all by somebody who cannot see the card move.
What the back arrow used to be, on the two screens that no longer have one. It holds the choices a screen with one room on it has nowhere else to put.
 |
| Back first, because that is the one a screen without an arrow is missing. |
It is a disclosure and not a role="menu": that role promises the arrow keys
move between the items, and announcing a keyboard model the app does not
implement is worse than announcing none. Escape closes it, so does a press
anywhere off the panel, and so does leaving the screen it was opened on.
The map with the room lit on it, an arrow to it, one plate naming it, and the rest of the ranking peeked underneath. The room's calendar is what taking a room makes irrelevant, not the other rooms: one tap on a row moves the arrow and the plate to it, which is the cheapest change of mind in the app.
 |
| The answer, the way to it, and the runners-up. |
This screen has no back arrow either. The menu's Back leaves it, and so does pulling the sheet's grip down through its travel, which is the same gesture that leaves every other screen in the app. Both land on the card you took, deck still on the same room, so one more swipe is the next one.
Then the last card offers the ranking as a list, which is every row the deck would have shown you, one per building, nearest first.
 |
| Behind the end of the deck, for when you would rather scan. |
 |
 |
| Take a room and the map comes up pointing at it. | Tap it again for its whole day, doors included. |
The card shows the room, because a room you can see is a room you can recognise from the corridor, and "Cunz Hall 160" never told anybody that. It is the whole screen: one frosted plate near the top, and the two verdicts on the bottom corners. The plate is narrower than the phone and centred in it, so it reads as a label ON the photograph rather than as a bar across it.
The top of the frame is stretched to make it fit. A 3:2 photograph in a
1:2.2 screen is a 3.25x aspect gap, and a cover crop pays for it in width: 31%
of the room, and since every one of these is shot from the back, half of that
column is foreground carpet. So the ceiling pays instead. It is flat, the plate
sits over it, and stretching it buys back 72% of the width for the part of
the room you are looking at. drawWarp() maps the source to the screen as
y ** p over 240 bands on a canvas. The exponent is derived per photograph
rather than fixed, because 219 of the 306 are 3:2 and the rest run from 4:3 to
16:9: the knob is how much taller than natural the BOTTOM of the picture may be
drawn, and p falls out of that and the aspect. At the shipped 1.23x, the top
120px of the phone comes from the top 14 source rows of a 3:2 room and the top 6
of a 16:9 one, and both read as a high ceiling rather than as a distortion.
dev/warp.html is the bench those numbers came off: both sliders, all 306
rooms, and a line showing how much of the picture the plate hides.
The screen is the picture and two icons, the same size and the same distance from their own corner. There is no third one. The back arrow that used to sit opposite them went because the card is a photograph and every glyph on it is something between the reader and the room, and the ranking has never had a control because an unlabelled glyph is a puzzle and a label is one more thing to read.
306 of the 425 have one. The other 119 are departmental rooms nobody has photographed, and they get the same card without a picture: the words were always the answer.
They come from OTDI's Learning Spaces
directory, which serves them from
rooms.app.it.osu.edu at an address that can be built from a room id. The
Registrar publishes the same rooms behind an opaque /media/<hash>/ path with
a REGISTRAR - 2022 watermark burnt into the frame, so that copy is neither
addressable nor the one you want.
The originals are 1620x1080 JPEGs of up to 1.34 MB. scripts/fetch-room-photos.mjs
resizes them to 900 wide and re-encodes them as WebP, which is 39 KB each and
11.7 MB for all 306, and commits them. Hotlinking would have put a megabyte on
the one screen a student opens in a stairwell, would break the day OSU moves a
file, and would tell OSU's server which room each reader is looking at, which
the privacy page promises this app does not do with anything
else. None of them are precached: each arrives with the card that shows it, and
is kept from then on.
Photographs © The Ohio State University, OTDI Classroom Services.
You do not have to. But it is built to be installed, and installing it is what makes it answer in a stairwell.
On iPhone, open it in Safari and tap Share, or ... then Share if your tab bar sits at the bottom, then Add to Home Screen. iOS 26 defaults to that Compact layout, which is why the app offers you both. On Android, Chrome has Install app in its menu.
Installed, the whole app is 160 KB of shell and about 112 KB of schedule, gzipped, and none of it is fetched again to answer a question. Turn the network off, open it, and it still ranks rooms. That matters because the moment you want it most is the moment you are in a basement with one bar.
Vacant reads the class schedule. A class schedule is not a reservation calendar and it is not a door. Three things are true and the app says all three on screen:
Doors get locked. This is the failure every other tool in this category ships with, and at Ohio State it is avoidable, because the Registrar publishes an open and close time per building per weekday. Measured against the committed index, 424 of the 425 rooms have no Saturday class at all this term. A tool that stops at the schedule calls all 424 of them free on a Saturday. 367 of those sit in a building the Registrar publishes as closed that day, and 57 are in a building that is genuinely open.
A door nobody documents is not offered at all. The Registrar's table covers 47 buildings. The index touches 46 of them, and 0 of the 425 rooms sit in a building whose hours are unpublished, because a room in one does not ship.
That used to be a label. 90 rooms rode along in their own tier reading hours not published, which was honest and was still the wrong answer: "here is a room, and
we cannot tell you whether you can get into the building" is not an answer, and
it cost four blocks of explanatory text across three screens to say. Over 5,040
ranked rows measured from four origins across a week, not one of them ever
reached a top ten. They were paying for themselves in prose and returning
nothing.
The reservation picture is incomplete. The Registrar's public Room Matrix
adds registered non-class events to Vacant's class schedule for the week its
snapshot covers. Unscheduled use, ROOM BLOCK holds and locked room doors are
still unknown. Outside the snapshot week, Vacant has class data only. The
ground-truth walk is still needed to measure
how often those gaps change an answer.
Ohio State publishes, for every class, the room it meets in and the minutes it occupies. Nobody publishes which rooms are empty. But empty is the complement of busy, so the whole product is one inversion of a public dataset. A script walks the class API, drops everything that is not a real Columbus room, and writes a room-keyed list of busy intervals. The page downloads that file, subtracts today's intervals from the building's published opening hours, and ranks what is left by walk time. There is no server, no database, no build step and no dependencies.
Everything the page fetches to produce a ranked list is fifteen files, 515 KB,
or 129 KB over the wire once gzipped. Measured with node:zlib over the tree
as committed. The campus map is another 98 KB, 38 KB gzipped, and is warmed after
the first answer rather than before it.
| Where | What is in it |
|---|---|
index.html |
The whole shell. Markup and CSS, no framework. |
js/app.js |
Four screens in one sheet: the card, the list, one room, the buildings. |
js/engine.js |
The ranking, and the formula that decides how long a room is yours. |
js/map.js |
The campus map, drawn as vectors on a canvas. No tiles, no key. |
js/campus.js |
Latitude and longitude into map grid space. |
scripts/ |
The harvest, the inversion, and the screenshots. |
data/ |
The committed artefacts the page reads. |
git clone https://github.com/EnesYilmazcode/Vacant.git
cd Vacant
python3 -m http.server 8000 # or any static file serverThen open http://localhost:8000. It has to be served rather than opened as a
file, because the page is ES modules and it fetches JSON.
npm test # node --test, no network
node scripts/shoot.mjs # redraw docs/media from the real appOpen http://localhost:8000/dev/. It asks the two questions first: what
minute, and where are you standing. Tap the campus map to drop a pin anywhere,
or pick a building, or take one of seven one-tap jumps to Thanksgiving, finals
week, Saturday at 3am or winter break. Then it hands off to the real app.
Once you are in, ?dev=1 or three presses of D opens a panel that changes
the same two things without leaving the screen you are on.
?dev1 is one URL for a whole scene: the panel, already standing at Dreese
Laboratories on a Thursday at 2:15pm. It exists because the live minute is a bad
one to look at the app in. Open it at 10pm and every building on campus is shut,
the app correctly refuses to rank, and there is nothing on any screen. #dev1
and ?dev=dev1 do the same, and the scene survives the reload that strips the
query string.
The date is derived rather than written down: js/dev.js walks the teaching
range in the shipped index for the first Thursday that is neither a closed day
nor inside exams. A literal date in that file would name a Thursday in a term
that ended the moment the next one shipped. Any Thursday would do, because a
class schedule is keyed to the day of the week, so every Thursday in term
carries the same grid.
It is not a preview. It moves the same clock every screen reads and the same origin the ranking measures from, then repaints through the same code path a duration button uses, so the answer on screen is the answer a student would get. The readout at the bottom of the panel says whether the clock is live or simulated, and it says which rooms came back and why the app refused if it did.
js/dev.js is loaded on demand and is not in the service worker's shell list, so
a phone that never asks for it never downloads it.
Nothing here needs a key or an account. Each of these takes --dry-run.
node scripts/fetch-buildings.mjs # OSU's GIS layer -> data/buildings.json
node scripts/fetch-campus.mjs # campus polygons -> data/campus.json
node scripts/fetch-building-hours.mjs # Registrar table -> data/buildings-hours.json
node scripts/fetch-rooms.mjs 1268 # the class schedule -> data/harvest-1268.json.gz
node scripts/build-index.mjs 1268 # invert it -> data/rooms-1268.json
node scripts/fetch-sidewalks.mjs # the sidewalks -> data/walk-graph.jsonThe sidewalks are not on the weekly clock the schedule is on. walk-graph.json
is OSU's own centreline layer, contracted to 5,242 junctions and 7,977 edges,
and it only needs rebuilding when OSU repaves something or opens a path. Run it
after fetch-buildings, which writes the doors the graph is pruned around.
The harvest itself is not committed, because it is regenerated weekly, so a fresh
clone has to run fetch-rooms before build-index. One pass over the schedule is
136 requests against content.osu.edu, sent at about three a second. The
committed Autumn 2026 harvest took four passes and 545 requests, because paging
that API is not deterministic: the harvester repeats until two passes in a row
turn up no meeting in a room it has not already seen. Pass two found thirteen of
them. data/harvest-1268.manifest.json records what every pass saw.
data/raw/1262 and data/raw/1264 are committed and must stay. Those terms have
already left the API and cannot be refetched at any price.
- Nobody has walked to a room the app called free and checked. That is the
measurement the whole thing rests on and it has not been taken.
#26 is the walk, and
spikes/walk.htmlis the checklist to walk it with.
Roomix is the incumbent at Ohio State and deserves an honest description. It is a maintained three platform product: a Flutter app with iOS and Android builds, a backend, accounts, bookmarks, themes, a nearest facility button and a vacancy search. Anyone calling it a toy has not looked at it.
Two differences are measurable rather than matters of taste. Both were taken off its own API and its own compiled bundle on 2026-08-26.
It downloads about 3.3 MB on first launch. Its API is eleven static JSON
files and courses.json alone is 2.4 MB.
It has no building hours. Grepping the bundle for "building hours", "holiday" and
"final exam" returns nothing, and api.roomix.app/hours.json is a 404. Its
vacancy search also needs a building selected first and then stops at a 200 metre
radius from it, and it prints raw metres rather than walk time.
The full teardown, with the commands, is in docs/research/prior-art.md. It also corrects an earlier draft of this README, which said Roomix was organised building by building and ignored the clock. Both halves were wrong.
- docs/BACKLOG.md is the thirty issues, in order.
- docs/DECISIONS.md is append only, and records what each decision was made against and which measurement settled it.
- docs/research/ is the measurement itself: the live API, OSU's GIS server, the Registrar's table, Roomix's bundle.
- docs/BLUEPRINT.md is the design note this README used to be, kept whole, with the places it turned out wrong marked at the top.
- scripts/shoot.mjs regenerates every screenshot above from
the real app, on a pinned clock and a pinned location, and refuses to write a
frame of a screen that has not stopped moving. Re-running it on an unchanged
tree writes the same five files. It also writes down what each screen said, in
docs/media/frames.json, and a test holds the alt text above to it.
The code is MIT, in LICENSE.
The data is derived from Ohio State's own public services: the class schedule at
content.osu.edu, the building layer published by OSU Facilities Information and
Technology Services GIS, and the classroom pool schedule from the Office of the
University Registrar. Each file in data/ carries its source and attribution
inside it. Vacant is not affiliated with Ohio State. How that gets stated on the
site is #25.