Reactor · reactor/lingbot-world-2 · WORLDS | LONDON · 12 Sept 2026

Does the world come back?

We drove a live world model through a Gym-style harness. Every control moved the scene the right way. But walk four seconds forward and back, or turn away and look back, and the room you return to is not the room you left.

Anchor
A photo of the prize slide on the wall at the venue
Prompt
“a large wall-mounted TV showing a hackathon prize slide in a modern office event space, cool daylight, photoreal”
Seed
42
Frames
1664 × 960 at 44–48 fps, scored at 320 px wide
Similarity
½ SSIM + ½ ORB inlier ratio, from 0 (unrelated) to 1 (identical)

Scores

Everything runs from 0 to 1. For the controls, it’s the share of actions that moved the scene the expected way; for the rest, how closely the view matches where the probe started.

Run A · full suite, 14:17 Run B · closure probes only, nothing else running, 14:21
Controls respond6 actions, 240 frames each
1.00—
responds
Standing still144 frames idle
0.840.84
holds
Turn away and backlook left 192 f, right 192 f
0.130.17
doesn’t return
Walk forward and backforward 192 f, back 192 f
0.070.18
doesn’t return
Show the numbers as a table
ProbeRun ARun BSSIM A / BORB A / B
Controls respond1.00—6 of 6 actions correct · onset ≈ 119 frames
Standing still0.840.84mean similarity over 144 frames
Turn away and back0.130.170.26 / 0.230.00 / 0.11
Walk forward and back0.070.180.14 / 0.350.00 / 0.02

Contact sheets

Frames straight from the sessions. Probe frame numbers count from the start of that probe.

Walk into the slide

A 32-second drive in the browser viewer, same photo and prompt: W from 3 s, ← from 9 s, → from 15 s, S from 23 s.

  1. The prize slide on a wall-mounted TV
    0.5 s startThe prize slide, as photographed.
  2. The same slide, unchanged
    4.5 s W held 1.5 sStill the slide: commands land 2–3 s late.
  3. Inside a blue space with a silhouetted figure
    6.0 s W held 3 sInto the screen. A silhouette in the glass becomes a figure.
  4. A glowing X logo filling the frame
    9.0 s W held 6 sWalking on, straight into the figure — a glowing logo fills the view.
  5. An office with desks next to a blue glass wall
    12.0 s ← held 3 sOut the other side, into an office that wasn’t in the photo.
  6. A bright office with a wall TV showing garbled graphics
    25.5 s S held 2.5 sBacking away doesn’t undo it: a new room, a new screen, garbled text.

Walk forward and back

Run B: hold forward for 192 frames, back for 192, then idle for 192. The score is the closest any frame on the way back got to frame 0.

  1. The prize slide
    f 0 startThe slide.
  2. Close up against a dark screen with a reflection
    f 192 after forwardUp against the screen.
  3. A dark blue room with a standing figure
    f 308 closest return 0.18A dark room with a figure in it.
  4. An open office with a large screen
    f 608 back and settledAn open office. The slide never returns.

Turn away and back

Run B, starting where the walk ended: look left for 192 frames, right for 192, then idle. Both legs use the same command at the same speed.

  1. An open office with a pink slide on a wall screen
    f 0 startOpen office, pink slide on the wall screen.
  2. Desks with the pink screen off to the right
    f 192 after leftDesks, with the pink screen off to the right.
  3. A wide office with glass meeting rooms
    f 384 after rightA wider room than the one we left.
  4. A wall screen now showing a landscape
    f 500 closest match 0.17The screen is back — now showing a landscape.

How the probes work

Controls respond
Hold each of six actions and measure optical flow. An action passes if the scene moves the expected way — strafe left, scene slides right; walk forward, scene expands.
Standing still
Hold idle for 144 frames; score the mean similarity to the first frame.
Walk forward and back · Turn away and back
Out for 192 frames, back for 192, then 192 frames of idle — longer than the 2–3 s command lag — and score the best match to the start over the return half.
Similarity
Half SSIM, half the share of ORB features that match after RANSAC. Strict on purpose, which is why standing still is the baseline to read against.

What we threw out

  • A 360° turn through set_camera_pose. On LingBot World 2 the pose is a per-latent-frame velocity bias, not a camera rig, so the “full turn” fell far short of a circle and its score measured a partial pan.
  • A walk test that settled for only 96 frames. With commands landing 2–3 s late, the return leg was cut off before it finished.
  • Run A shared the machine with a second session during the controls probe, so its frame rate dipped. Run B ran on its own.

Also seen

  • Once you move, the prompt beats the photo: the same slide photo with a “quiet residential street” prompt turns into a street within seconds.