https://miaomind.com/#notes Low-ceremony research log from MiaoMind. Dates, code, traces, honest findings. We read the paper so you don't have to. en https://miaomind.com/notes/the-runtime-is-the-product.html https://miaomind.com/notes/the-runtime-is-the-product.html Mon, 06 Jul 2026 09:00:00 +0000 Harness engineering was the right correction — and it's table stakes now. What clients pay for is runtime design that speaks their vertical. Models are rented, harnesses are copied; the runtime compounds. https://miaomind.com/notes/dynamic-workflow-is-superpower.html https://miaomind.com/notes/dynamic-workflow-is-superpower.html Thu, 02 Jul 2026 09:00:00 +0000 We split all 32 marketing-agent jobs into intake + run: one batch of questions up front, then an autonomous pipeline with three legal endings — staged, escalated, failed. A mid-run question is now a bug, not a feature. https://miaomind.com/notes/our-daily-driver-recalled.html https://miaomind.com/notes/our-daily-driver-recalled.html Sat, 13 Jun 2026 09:00:00 +0000 Eighteen days of a global Fable 5 blackout, triggered by a verbal report of one narrow jailbreak. Where we stand, what we did about it, and why your toolchain is a geopolitical variable. https://miaomind.com/notes/2026-ai-index.html https://miaomind.com/notes/2026-ai-index.html Tue, 14 Apr 2026 09:00:00 +0000 SWE-bench Verified went 60% to ~100% in a year — the headline chart. The one we're watching: the Foundation Model Transparency Index dropped 18 points. What that means for anyone building honest evals. https://miaomind.com/notes/mythos-preview-system-card.html https://miaomind.com/notes/mythos-preview-system-card.html Tue, 07 Apr 2026 09:00:00 +0000 Anthropic built a 93.9%-on-SWE-bench model, measured it, then decided not to ship it. Three moves worth copying: the welfare section, the refusal framing, and the eval appendix. https://miaomind.com/notes/harness-engineering-is-a-discipline.html https://miaomind.com/notes/harness-engineering-is-a-discipline.html Sat, 21 Mar 2026 09:00:00 +0000 We read the 60-page externalization survey so you don't have to. Same model, rebuilt harness: our 50-step cliff moved to step ~82. Zero weights were harmed. https://miaomind.com/notes/the-50-step-cliff.html https://miaomind.com/notes/the-50-step-cliff.html Tue, 10 Mar 2026 09:00:00 +0000 Four agent frameworks, one 100-step ops workflow. All of them nose-dive between steps 40 and 60. Which failure modes are avoidable, which aren't — with traces. https://miaomind.com/notes/process-rewards-without-labels.html https://miaomind.com/notes/process-rewards-without-labels.html Sat, 28 Feb 2026 09:00:00 +0000 Following the DPO-PRM hybrid thread on our code-review task: 71% to 84% pass rate with ~2k unlabeled agent traces. Method, numbers, honest caveats.