daily notes

This is unusually concrete writing on long-running agent design: every run step is a stored task, restarts resume unfinished work from the same storage, and tool replay is explicit about what is safe to repeat. The full note is worth reading for its storage, task-ownership, extension and concurrency patterns, while its experimental status keeps the architecture separate from a blanket production claim.

OpenAI says it will add invisible watermarks to eligible ChatGPT and Codex text in the EU over coming weeks, while API customers worldwide can opt in for select models. The accompanying detector remains restricted to approved researchers and expert organizations because the company documents material false-positive, false-negative and edit-resistance limits; provenance is therefore not a test of authorship, accuracy or responsibility.

Using millions of insurance contracts matched to property exposure, disaster risk and realized claims, the authors estimate that deductibles and coverage limits leave households bearing 29% of expected losses. The full note earns its place by separating modest insurer moral-hazard costs from unequal residual risk: lower-FICO households and properties exposed to tail disasters retain more uninsured exposure, while the paper is clear that those patterns are not causal estimates.

The October 2 model update puts October CPI at 0.22% month over month and core CPI at 0.20%; its year-over-year readings are 3.58% and 2.30%, respectively. The modest headline revision does not settle the next official CPI print: this is a nowcast built from higher-frequency inputs, not released inflation data.

The October 2 chair statement says the G20 reached consensus on a statement about food weaponization through coercive trade actions and discussed adding diesel to global supply. It is evidence of ministerial alignment and an energy-policy discussion, not a quantified or binding coordinated release; execution details remain the relevant monitor.