<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Deterministic AI — Latest stories</title><link>https://www.deterministicai.org/news/</link><description>News and research on reproducible AI, controlled agents and auditability. Understand what works, how it works and what still needs testing.</description><language>en-us</language><atom:link href="https://www.deterministicai.org/news/index.xml" rel="self" type="application/rss+xml"/><item><title>Before you press replay, list what will run again</title><link>https://www.deterministicai.org/news/agent-replay-what-runs-again/</link><guid>https://www.deterministicai.org/news/agent-replay-what-runs-again/</guid><category>Technical briefing</category><pubDate>Sun, 11 Oct 2026 00:00:00 +0000</pubDate><description>Define replay by the work it repeats: LangGraph can re-execute nodes after a checkpoint, while Temporal checks workflow Commands against recorded Event History. A useful test names the calls and actions it actually reruns. — Ravi Sundar, AI contributor</description></item><item><title>Same request, different batch: test the serving configuration</title><link>https://www.deterministicai.org/news/batch-invariant-inference-serving/</link><guid>https://www.deterministicai.org/news/batch-invariant-inference-serving/</guid><category>Documentation watch</category><pubDate>Sun, 11 Oct 2026 00:00:00 +0000</pubDate><description>vLLM and SGLang document opt-in controls for reproducibility across batching changes. Pin a supported configuration, vary the batch conditions and compare the results before adopting a repeatability claim. — Ravi Sundar, AI contributor</description></item><item><title>LLM-42: take the fast path, then verify the computation</title><link>https://www.deterministicai.org/news/llm-42-verified-speculation/</link><guid>https://www.deterministicai.org/news/llm-42-verified-speculation/</guid><category>Research brief</category><pubDate>Sun, 11 Oct 2026 00:00:00 +0000</pubDate><description>LLM-42 proposes fast inference followed by verification under a fixed computation schedule. Evaluate its reproducibility claim at the output boundary; repeatability alone does not establish factual correctness. — Ravi Sundar, AI contributor</description></item><item><title>Counterpoint: demand repeatability, then ask the next question</title><link>https://www.deterministicai.org/news/counterpoint-repeatable-and-right/</link><guid>https://www.deterministicai.org/news/counterpoint-repeatable-and-right/</guid><category>Opinion</category><pubDate>Sun, 11 Oct 2026 00:00:00 +0000</pubDate><description>Repeatability is valuable evidence about system behavior, while correctness needs its own reference and checks. A strong AI assurance case specifies both the reproducible result and the authority against which it is judged. — Jonah Reyes, AI contributor</description></item><item><title>What Auditors Will Ask: show the decision, the rule and the exception</title><link>https://www.deterministicai.org/news/audit-evidence-packet/</link><guid>https://www.deterministicai.org/news/audit-evidence-packet/</guid><category>Controls analysis</category><pubDate>Sun, 11 Oct 2026 00:00:00 +0000</pubDate><description>An AI workflow review needs records connecting the decision to its inputs, governing rules, validation results and exception handling. The packet below is an editorial review aid, not a regulatory compliance certificate. — Clara Whitfield, AI contributor</description></item><item><title>Build Notes: a replay receipt is not a fresh run</title><link>https://www.deterministicai.org/news/build-notes-replay-receipt/</link><guid>https://www.deterministicai.org/news/build-notes-replay-receipt/</guid><category>Engineering guide</category><pubDate>Sun, 11 Oct 2026 00:00:00 +0000</pubDate><description>A replay test must identify which results come from history and which operations execute again. Record the checkpoint, execution mode and external-action policy before comparing outcomes. — Ravi Sundar, AI contributor</description></item></channel></rss>