Day 15: What happens when an AI film crew's own checks get too strict? A neon sign, a bent hook and $4.15 of GPU time

Day 14 was a long day of my notes. Day 15 was the day the crew's new rules got their first real test.

Day 15 at a glance

What does an AI film crew do when its own safety checks say no? On my film, on October 18, it mostly waited. Claude runs the crew. On Day 14 it added a breaker that stops paid GPU work after one empty pod, a hold on frames I approve with a fix note, and a check that keeps text out of the picture. On Day 15 all three fired.

The edit room still ran seven scheduled scene edits. The Midjourney runner drew frames until 12:51 p.m. RunPod made 29 clips, 15 of them in a single 39-minute run. All times here are Eastern.

What did I approve on October 18?

Two sittings, about 20 minutes in all.

• 12:27 to 12:36 a.m.: 29 frames. 23 approved, 6 rejected. One clip approved.
• 10:59 to 11:10 a.m.: 36 clips (27 approved, 9 rejected) and 17 frames (11 approved, 6 rejected).

Three of the midnight approvals came with a fix-in-post note. In a Scene 17 frame of Line Four, everything worked: the smoke, the red light, the men on the line, a bundle of pipes swinging from a chain. Everything except the crane hook in the foreground, which comes out bent back on itself, like a pretzel. My note said the hook was a strange shape. In a Scene 18 rooftop chase, a man leaps the gap between two roofs with an Enforcer robot behind him, sparks on its baton, and his legs come out split in two directions. And the pager in Scene 20 still showed the song lyric Midjourney had printed on it on Day 13.

Those notes matter now. At 12:35 a.m. the animation queue listed six frames "approved WITH ISSUES" and held them until they're fixed. A hook that can't hold anything won't be animated until it can.

I also rejected two takes of Mike on the roof in Scene 18 at 12:32 a.m. with a note to take the tears out of the prompt, and a third at 11:10 a.m. with the same note. Men in this film don't cry. Claude found tears in the prompts and scripts that same day and built a gate that blocks them.

Why did the GPUs sit idle until 6 a.m.?

My midnight approvals didn't go straight to a GPU. Every frame needs a written description of what's in it before it's animated, and 20 frames were waiting for one. A Claude helper wrote all 20 at about 2:34 a.m.

At 2:35 a.m. a batch of 15 frames went to RunPod. The pod lost its SSH connection during setup and was deleted at 2:43 a.m. with no clips, about $0.44. That tripped the waste breaker, which since Day 14 stops all paid batches after one empty pod. Between 2:47 and 6:10 a.m. it turned away 55 batch launches. The edit room didn't clear it, because the likely cause was the VPN on this Mac, and that needs me.

Once the pause was cleared, the 6:14 a.m. batch got a cheaper A100 GPU and made 15 clips in 39 minutes, about $0.89. At 11 a.m. I judged all 15 and approved 12, most of them from the Scene 20 night-mission montage.

How did a neon sign stop a batch?

At 11:08 a.m. I approved a Scene 06 frame: John at the far curb, a big man at a table in a lit diner window sipping from a cup, and the neon PARADISE sign glowing upside down in the puddle. The script asks for that sign, and Midjourney spelled it right, which after the garbled pager lyric on Day 13 felt like progress.

The next three batches, at 11:21, 11:25 and 11:29 a.m., stopped before any pod was rented. Claude's text checker had found "PARADISE" in the frame and treated it as painted-in caption text, and one text frame failed the whole batch.

Claude changed two things. Text that the shot's own script names is now allowed. Any other text frame is parked by itself, so it can't hold up the rest. The 11:33 a.m. batch launched. The clip of that frame finished at 1:44 p.m. On October 20 I approved it, with a note that a couple of the letters should blink at different times, like a broken sign.

What did the GPUs cost?

RunPod pods closed on October 18 cost about $4.15. There were six of them, and together they made 29 clips. Three made nothing, about $2.20. Six of the clips were new takes of shots where a man had been crying, made again without the tears.

Claude explained those empty pods three different ways in one day. First, a bad host: it put the host on a block list. Then, healthy pods so busy downloading 40 GB of models that SSH timed out. Then the third empty pod never answered SSH at all, and that killed the second theory too. The pattern was in the spend log the whole time. That day, every empty pod was an H100 from RunPod's pricier fallback tier, and every pod that made clips was an A100. The fallback is gone now, and the rule is to compare failures with successes before naming a cause.

What did the edit room do?

Seven scheduled runs finished, one scene each. Four critics scored picture, pacing, sound and lip sync before and after each run:

• Scene 07 (2:28 a.m.): 4.75 to 6.25. The same extreme close-up of Lucy's eye, looking up at John with the Spire in it, had played twice, a few shots apart. The repeat is now Lucy at her wall with a brush, her shadow about three times her size, and the music under the Victor bulletin is down to the bass.
• Scene 11 (4:22 a.m.): 5.38 to 5.50.
• Scene 02 (6:27 a.m.): 5.13 to 5.38. Lucy was answering John's question before he asked it. More on that below.
• Scene 02 again (9:28 a.m.): 5.38 to 5.75. Our own color mix had turned John's olive jacket brown.
• Scene 12 (11:28 a.m.): 5.75 to 6.5. Four clips I'd approved that morning went in, playing whole.
• Scene 11 again (3:28 p.m.): 5.0 to 5.75. A Scene 11 render at 2:32 p.m. still opened on a clip I'd rejected at 11:01 a.m. The picture critic gave that render a 3.0, and this run took the clip out.
• Scene 02 a third time (6:28 p.m.): 5.5 to 6.0.

A run that started at 9:28 p.m. rebuilt Scene 11 and then stopped before its second round of critics. It left no scores.

The Scene 02 fix is the one I keep coming back to. Lucy's retake was animated to her scripted answer, "Everybody who lives here and can't see it." On screen her mouth moved to it in silence, while an old draft line played over her closed mouth. Then her answer was heard again off screen, after John's question. So she answered before he asked, then answered twice. Claude only caught it by listening to what the take's own audio track said. The 6:28 p.m. run found the other half: John's retake was animated to his question, "Who's gonna buy a painting of this?", but the cut had covered his mouth with chuckles and played the line over the back of Lucy's head. Now he asks on his own mouth and she answers on hers. It took three Scene 02 runs in one day to get a couple to take turns.

What else went into the mistakes log?

19 rows dated October 18. A few of them:

• The review page timed out on me. The Mac was rendering on its own GPU, streaming a 41 GB model from the same drive the review page reads. Local renders now run at night or when I'm not reviewing.
• Frames I'd already settled kept coming back. The review page showed unpicked frames of shots I'd already approved. A shot with an approved frame now hides the rest, and an Ignore button hides a frame for good.
• My rejection notes weren't reaching the prompts. The runner only passed along Claude's own check reasons, and re-saving the same prompt counted as a rewrite. Now a shot can't go back to Midjourney until my note is worked into new text.
• Lucy, 13% taller. A crop in Scene 09 wasn't 4:3, and the render stretched her. A critic caught it in the same run. The editor now refuses any crop that isn't 4:3.
• "Push it." Mike's line in Scene 11 was missing for a day, because the line checks only counted each shot's main line.
• A 15-second Runway take nobody needed. Claude prepared shot 16-10a for Runway after I'd already approved a take of it. Runway jobs now refuse any shot that has an approved clip.
• Runway downloads that looked like a bot. The automation scrolled the feed every six minutes and once grabbed a 345 MB "download all assets" zip. It now makes one look after a long wait and clicks only that clip's own download button.

What did Midjourney do?

The runner sent 42 jobs between 12:28 a.m. and 12:51 p.m., four takes each, mostly for Scene 22 and Scene 21. Claude's pre-judge ran 163 checks: 102 passed and 61 failed. Shots with every take failed were parked 12 times for rewritten prompts.

Some of the failures:

• Scene 22-06, three rounds. The first prompt came back with the staging reversed. The second had John huge in the foreground and Lucy tiny under the window. The third gave John "a giant barrel torso" with a tiny head on top.
• Scene 22-03. Four takes, and three of them had "a stray floating smoking pipe at left edge with no hand." Day 14's pipe has friends.
• Scene 21-09. The shot is a paint stroke on canvas, with no people in it. Midjourney drew an eye.

Midjourney refused one Scene 01 shot twice that morning, so it went to the local list. Under the two-refusal rule it never goes back. From 1:36 p.m. the new no-tears gate kept parking the rewritten Mike frame, and the runner drew nothing else that day.

What got posted?

Three X posts went out from earlier diary days, at 10:09 a.m., 12:04 p.m. and 10:03 p.m.: the second John standing behind John like a bodyguard, Victor's lip sync, and the ghost child in Lucy's window. One TikTok went out at 12:07 p.m., the stray hand in John's shot.

The numbers for October 18

• My verdicts: 83 in two sittings. Frames: 46 judged, 34 approved. Clips: 37 judged, 28 approved.
• RunPod: about $4.15 for 6 pods, 29 clips, 3 pods with no clips (about $2.20).
• Waste breaker: 68 batch launches turned away (55 before dawn, 13 around noon).
• Text checker: 3 batches stopped by the PARADISE sign.
• Edit room: 7 scheduled runs finished, 1 stopped before its second round.
• Midjourney runner: 42 jobs. Pre-judge: 163 checks, 102 passed.
• Mistakes logged: 19 rows.

So what did Day 15 teach me?

A check that says no is only as good as what it knows. The breaker was right to stop paying for empty pods. The text checker was wrong about a sign the script asked for, and the no-tears gate ended up blocking the very prompt it was built to fix. Both were fixed the same way: teach the check what the film wants.

The cheapest day so far was also the slowest. $4.15 bought 29 clips. Most of the waiting was us, not the GPUs. Follow the run on the Making Death Match series page.

FAQ

Why did an AI text checker block a neon sign?
The check was built to keep stray captions and garbled words out of frames. It found "PARADISE" in a Scene 06 frame and failed the whole batch. The sign is in the script, so Claude changed the check to allow text the shot's own script names.

What is a waste breaker on GPU rentals?
A rule that stops all paid batches after one rented pod finishes with no clips. On October 18 it turned away 68 launches after empty pods at 2:43 a.m. and 11:47 a.m.

How much did a day of AI animation on RunPod cost?
About $4.15 for six pods closed on October 18. They made 29 clips, and the three empty pods cost about $2.20.

What does "approved with a fix-in-post note" do?
The frame is approved for style and story, but the animation queue holds it until the problem is fixed. On October 18 six frames were held that way, including a crane hook bent back on itself.

While you wait for the next update: play Halfsies

Halfsies is our iPhone game, live now. One cut. Two halves. Each day you get three objects and one attempt at each: cut it exactly in half by eye and the scale tells you how close you got. No bent hooks. Download Halfsies on the App Store