Kristian Fagerlie gives GPT-5.6 and GPT-5.5 the same research brief over a collected Bitcoin prediction-market dataset after separating prior strategy material. He watches the agents investigate, backtest and encounter additional safety checks, without treating the workflow as a controlled scientific comparison.
Kristian Fagerlie then asks GPT-5.6 to assess both proposals. It favors its own model's hypothesis but scores it only 47 out of 100 and calls for more research; the competing proposal also lacks sufficient evidence. This self-evaluation is not independent validation or proof of a live trading edge.
Kristian Fagerlie plans further investigation rather than deploying the new strategy during this video. Earlier reported profits do not establish the new hypothesis's profitability, and the incomplete dataset, different reasoning settings and unverified execution costs limit what this example can show.
Watch on YouTube
