What we tried

1. An ending image can erase the move that gets you there.

Three pairs repeated the same LASERPAW/VOLTJAW finisher with seeds 111, 222 and 333. Each pair used the same starting art, prompt and settings; one added a target ending image.

All three text-only controls retained a yarn sphere and rainbow trail, although literally riding the sphere remained ambiguous. All three end-image trials skipped that setup. Both approaches showed the shark falling and staying down. The target improved a particular grooming pose, not whether the intended winner won.

What we learned: the end-image runs started 7.525 seconds slower on average. This is three repeats of one matchup, with a fixed run order and provider variability—not a general latency benchmark. Seed 222’s target run reached the 40-second time limit. Keep end images off by default; test them selectively when the staging benefit is worth the delay.

2. A longer connection still needs recovery.

One session lasted 120.978 seconds, showing 112.030 seconds of video. It still had a 0.416-second gap between decoded frames. A separate short control had a 5.200-second frame gap. Reported session allowances changed during testing, so a fixed maximum timer would be a poor substitute for reading the live allowance.

The longer session also ran a scripted practice match. Moves appeared roughly 6–11 seconds after submission, even when the service accepted the prompt earlier. Character proportions and staging drifted, and unwanted text-like markings appeared.

What we learned: retain session recovery. These timings exclude the game’s judge and choreographer, spectators, balances, commentary mixing and finish verification. A complete Studio match and maximum-session replacement test remain unverified.

3. Supplying sound worked. Running out produced silence.

The recording contained the supplied 110 Hz sound, then the queued 880 Hz sound, then the 330 Hz replacement. Replacing audio did not immediately interrupt material already dispatched. Exhaustion produced very quiet or silent samples; an invalid audio source was rejected without ending the video session.

Default audio and explicit 96, 128 and 192 kbps settings all produced recordings. The test used synthetic tones to make source identity measurable. These were not production effects, and no supplied impact was shown to line up with a game-authorized hit.

What we learned: keep commentary on its separate track. Signal measurements and model-assisted audio reviews do not establish which setting sounds best; direct human listening is still needed. The saved files were re-encoded, so their bitrate does not measure the incoming audio bitrate.

4. More pixels worked, but did not fix the fight.

Each short live sample reported zero dropped frames, but higher resolutions took longer to generate. A separate ten-second offline re-recording produced 2.66 / 2.78 / 5.07 MB at 480p / 768p / 1080p. The 1080p encoder used a higher bitrate, so this is not an isolated resolution comparison. A late 1080p close-up also hid the defeated fighter.

What we learned: retain 480p as the default. Trial 768p selectively; higher detail alone did not improve move fidelity or finish framing. These tests did not validate the live spectator path.

The next hypothesis: finish with both faces visible.

A local prompt change made at 16:11:21 PDT on September 8 asks for both full faces to remain recognizable through the final second and actual last frame. It preserves masks, nonhuman anatomy and the resolved result: a defeated fighter stays down, and the action cannot be shortened or reset just to reveal a face.

Not yet established: whether this reduces distortion in the next scene. The eleven recordings below used the earlier prompts. Thirty prompt checks passed after this change, but no new video trial or production deployment was performed for it.