Native agents' first films catch up with the plugin's

A native agent's first film looked worse than the plugin's. A skateboard came out as a truck, an hourglass as a light bulb. The prompts were the same. The models were not.

What changed

• The plugin draws the hero and writes the film on Sonnet 5.5, and writes scene 1 on Haiku. A native agent did all of it on its chat model.
• A kit word in the agent's own film line could beat the noun you asked for. Now your words win.
• A drawing traced in 16 points used to be thrown away whole. It is thinned to 12 and kept. That alone was why skateboard failed 3 times in 3.
• Haiku writes scene 1 while Sonnet writes the rest. The hero starts drawing the moment you ask, beside the reply.

What landed

Fresh account, a native agent, 19 nouns, first ask only, scored by the same look pass.

• Frames that pass: 26% before, 76% of 90 frames after. The goal was 70%.
• First scene, median: 4.1 s before, 3.1 s after. The goal was 3.5 s.
• The skateboard, the drum and the hourglass are now themselves.

What missed

• The strong route costs about 20 times a film on the chat model. The account that pays for it had about 62 cents left, so it ships off. Turning it on is Chris's call.
• Without it you get the guard, the outline fix and the hero drawn ahead. Those are in. The 76% needs all of it.
• The second ask after this change was not measured. The credit ran out first.
• One run per variant. The first four ran four at a time, so their seconds are inflated.