Does Qwen 3.7 Flash beat OpenAI's only small open model? — live #3

I have a campaign that already has numbers on it. Swiss supplier invoices, one account code per invoice, scored by exact match — PromptPotter took it from 5% at the origin to 55% over five rounds, all of it on openai/gpt-oss-20b at reasoning_effort low, temperature 0, 20 samples a round, three variants a round. Tonight I pull that model out and drop Qwen 3.7 in behind it, live, and we watch what the same loop does with a different engine underneath. Then I swap the other model too: opencode's Qwen 3.8 Flash takes over from Claude Opus as the thing driving my editor, and I hand it a real bug from this repo. 🎯 WHAT THIS VIDEO SHOWS YOU: You cannot swap a model until you have measured the one you are replacing. That is the whole episode. Most "model X vs model Y" content is two screenshots and a vibe; this is a before with its controls attached, and then the same run with exactly one thing changed. What you'll see: 1/ The existing swiss-invoices-eval campaign, read out loud — origin, rounds, what the optimizer actually changed and why 2/ The swap: openai/gpt-oss-20b out, qwen/qwen3.7-flash in. Same dataset, same scoring, same budget, and the first rounds land live 3/ The second swap: Qwen 3.8 Flash instead of Opus driving the editor, pointed at a display bug that only appears once a campaign has stopped 🔑 THE KEY INSIGHT: A round is not won on accuracy. It is won on ability — θ — fitted across every cell the campaign has paid for, and the sample subset moves between rounds. Last stream I hit a round that elected a lower-accuracy winner on camera and said I didn't know what was going on. I know now, and I'll show you the line that settles it: the overlap row, where the origin and every winner since are read on one shared set of cells. 🕐 TIMESTAMPS 0:00 – You can't swap a model until you've measured it 3:00 – The campaign that already has numbers: 5% → 55%, and the controls 9:00 – Why a lower-accuracy round can still be the winner 14:00 – The swap: gpt-oss-20b out, Qwen 3.7 in 22:00 – First rounds on the new model, read live 30:00 – Second swap: Qwen 3.8 Flash instead of Opus, driving the editor 36:00 – The bug that only appears after a campaign has stopped 42:00 – What the numbers say, and what it would take to say more 🔗 RESOURCES: PromptPotter: https://github.com/runfish5/prompt-po... opencode: https://opencode.ai/go Claude Code: https://claude.ai/code My GitHub page: https://github.com/runfish5 My website: https://runfish.net 👋🏻 ABOUT ME My name is David. Former PhD researcher turned AI Engineer. I build practical automation systems that actually save time—not hype, just workflows that work. 📗 MORE EPISODES: ▶ EP 02: Can open models operate PromptPotter? https://www.youtube.com/live/DLhb26ppX_s ▶ opencode in VS Code — every open model on one subscription:    • AI Coding for Beginners. OpenCode Setup, 2...