I asked for an AI micro-drama experiment for 900 RUB of GPU. The edit lost half the dialogue

I asked for an experiment with a short drama series. An agent assembled a pilot on one rented GPU: 52.2 seconds, three fictional characters, English dialogue, and a vertical format. The promo added another 22.4 seconds.

The session used four billed hours on an RTX PRO 6000, about 900 RUB.

I watched the first cut with sound and said that half the conversations could not be heard. At first, I blamed Kandinsky.

I ran transcripts of the raw clips: all 16 lines had been spoken correctly. The fault was in the edit. Kandinsky started speaking two to three seconds into a clip, while the assembler cut from zero and kept the silence.

The agent corrected the assembler. It then became clear that Kandinsky had not lost the lines.

I kept Kandinsky 6 for the pilot. I looked at its first frames next to Wan2.2: Wan can keep one voice per character, but Kandinsky had more gesture and light, and cost about 9 RUB per five-second clip, versus roughly 20–35 RUB for Wan.

What is not ready. Kandinsky generates the voice anew in every shot, so it cannot be pinned to a character. The music was chosen blind. One corridor still has a garbled exit sign. A program checked the takes, not a person: a face counter, text recognition and speech recognition.

The next step is viewers. Before paying for a season I need to see whether people finish these episodes. Withdrawing Telegram Stars money from Russia is untested.

The characters are fictional.