<?xml version="1.0" encoding="utf-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Arcus — Research &amp; Announcements</title>
    <link>https://arcusx.ai/blog-index.html</link>
    <atom:link href="https://arcusx.ai/feed.xml" rel="self" type="application/rss+xml"/>
    <description>Protocol releases, technical thinking, and updates from Arcus Labs.</description>
    <language>en-us</language>
    <item>
      <title>Releasing irg-reference: An Open Implementation of the IRG Protocol</title>
      <link>https://arcusx.ai/blog/2026-05-04-irg-reference-release.html</link>
      <guid isPermaLink="true">https://arcusx.ai/blog/2026-05-04-irg-reference-release.html</guid>
      <pubDate>Mon, 04 May 2026 00:00:00 GMT</pubDate><category>Engineering</category>
      <description><![CDATA[Arcus Labs is releasing an open reference implementation of the Iterative Reasoning Graph protocol, including a multi-provider runtime and a trace navigator that makes reasoning literally inspectable.]]></description>
    </item>
    <item>
      <title>Two Kinds of Momentum: Why Fresh Calls Change the Reasoning Process</title>
      <link>https://arcusx.ai/blog/two-kinds-of-momentum.html</link>
      <guid isPermaLink="true">https://arcusx.ai/blog/two-kinds-of-momentum.html</guid>
      <pubDate>Tue, 14 Jul 2026 00:00:00 GMT</pubDate><category>Thinking</category>
      <description><![CDATA[A reasoning model carries two kinds of momentum: the reasoning tendencies shaped by training and encoded in its weights, and the autoregressive pull to stay consistent with whatever it said first. Chain-of-thought struggles to interrupt the second, because the check comes from the same running generation. A separate model call can reduce that commitment while keeping the capabilities in the weights — and we argue that is most of why an IRG graph differs from a long prompt.]]></description>
    </item>
    <item>
      <title>The Compression of Reasoning: What VibeThinker-3B Actually Demonstrates</title>
      <link>https://arcusx.ai/blog/compression-of-reasoning.html</link>
      <guid isPermaLink="true">https://arcusx.ai/blog/compression-of-reasoning.html</guid>
      <pubDate>Mon, 13 Jul 2026 00:00:00 GMT</pubDate><category>Technical</category>
      <description><![CDATA[A 3B-parameter model is matching flagship systems on verifiable reasoning benchmarks. The compression-coverage hypothesis behind it says reasoning and knowledge are separable — which changes how production AI systems should be architected.]]></description>
    </item>
    <item>
      <title>Reasoning Strategies, Part 1: The Epistemic Family</title>
      <link>https://arcusx.ai/blog/reasoning-strategies-epistemic.html</link>
      <guid isPermaLink="true">https://arcusx.ai/blog/reasoning-strategies-epistemic.html</guid>
      <pubDate>Thu, 02 Jul 2026 00:00:00 GMT</pubDate><category>Technical</category>
      <description><![CDATA[First in a series covering the IRG strategy inventory. This installment introduces the strategy/tactics/implementation distinction and covers the three epistemic strategies: abductive, deductive, and inductive reasoning as executable graph shapes.]]></description>
    </item>
    <item>
      <title>Active vs. Passive Governance: The Distinction That Decides What AI Can Do</title>
      <link>https://arcusx.ai/blog/active-vs-passive-governance.html</link>
      <guid isPermaLink="true">https://arcusx.ai/blog/active-vs-passive-governance.html</guid>
      <pubDate>Fri, 26 Jun 2026 00:00:00 GMT</pubDate><category>Thinking</category>
      <description><![CDATA[Traditional GRC platforms govern policy: they inventory AI assets, log outputs, and attest compliance. They do not govern outcomes. Active governance operates inside the reasoning process itself — and the difference determines what you can safely let AI decide.]]></description>
    </item>
    <item>
      <title>SR 26-2 Supersedes SR 11-7 — and the Arcus SR 26-2 Model Risk Suite Is Available Today</title>
      <link>https://arcusx.ai/blog/sr-26-2-model-risk-suite.html</link>
      <guid isPermaLink="true">https://arcusx.ai/blog/sr-26-2-model-risk-suite.html</guid>
      <pubDate>Sun, 14 Jun 2026 00:00:00 GMT</pubDate><category>Announcement</category>
      <description><![CDATA[On April 17, 2026, the Fed, OCC, and FDIC superseded SR 11-7 with risk-based interagency guidance. Arcus now offers the SR 26-2 Model Risk Management Graph Suite as an enterprise offering — same reasoning engine, new regulation layer.]]></description>
    </item>
    <item>
      <title>Measuring Reasoning Quality: Why Accuracy Benchmarks Miss the Point</title>
      <link>https://arcusx.ai/blog/measuring-reasoning-quality.html</link>
      <guid isPermaLink="true">https://arcusx.ai/blog/measuring-reasoning-quality.html</guid>
      <pubDate>Mon, 18 May 2026 00:00:00 GMT</pubDate><category>Technical</category>
      <description><![CDATA[Accuracy benchmarks reward getting the right answer, even when the system hallucinated its way there. What predicts trustworthiness in production is how the system behaves when it doesn&rsquo;t know&mdash;and that requires a different kind of measurement.]]></description>
    </item>
    <item>
      <title>From the Perimeter to the Core: Why Auditability Unlocks AI Adoption</title>
      <link>https://arcusx.ai/blog/perimeter-to-core-auditability.html</link>
      <guid isPermaLink="true">https://arcusx.ai/blog/perimeter-to-core-auditability.html</guid>
      <pubDate>Mon, 11 May 2026 00:00:00 GMT</pubDate><category>Thinking</category>
      <description><![CDATA[Enterprises are not blocked by AI capability. They are blocked by AI accountability. The path from perimeter tasks to core decisions runs through reasoning that can be inspected, not better benchmarks.]]></description>
    </item>
    <item>
      <title>The Colorado AI Act Is Live: What It Requires and What It Doesn&amp;rsquo;t</title>
      <link>https://arcusx.ai/blog/2026-05-04-colorado-ai-act.html</link>
      <guid isPermaLink="true">https://arcusx.ai/blog/2026-05-04-colorado-ai-act.html</guid>
      <pubDate>Mon, 04 May 2026 00:00:00 GMT</pubDate><category>Thinking</category>
      <description><![CDATA[Colorado&rsquo;s SB 205 went into effect in February 2026 and applies to companies that deploy AI in consequential decisions. The law requires documentation of the decision process, not just the decision.]]></description>
    </item>
    <item>
      <title>Prompt Sets as Epistemic Personalities: Same Graph, Different Reasoning</title>
      <link>https://arcusx.ai/blog/2026-04-27-prompt-sets-epistemic-personalities.html</link>
      <guid isPermaLink="true">https://arcusx.ai/blog/2026-04-27-prompt-sets-epistemic-personalities.html</guid>
      <pubDate>Mon, 27 Apr 2026 00:00:00 GMT</pubDate><category>Technical</category>
      <description><![CDATA[The same reasoning graph, run with different prompt configurations, produces measurably different epistemic behavior. Prompt engineering isn&rsquo;t UX tuning&mdash;it&rsquo;s personality configuration, and now it can be measured.]]></description>
    </item>
    <item>
      <title>What Model Validation Looks Like When the Model Is an LLM</title>
      <link>https://arcusx.ai/blog/2026-04-22-llm-model-validation.html</link>
      <guid isPermaLink="true">https://arcusx.ai/blog/2026-04-22-llm-model-validation.html</guid>
      <pubDate>Wed, 22 Apr 2026 00:00:00 GMT</pubDate><category>Thinking</category>
      <description><![CDATA[Traditional model validation rests on three assumptions&mdash;inspectable mechanics, holdout accuracy, and reproducible stress tests. LLMs break all three. Here&rsquo;s what actually works.]]></description>
    </item>
    <item>
      <title>Why AI Hallucination Rates Get Worse Where It Matters Most</title>
      <link>https://arcusx.ai/blog/2026-03-25-ai-hallucination-rates.html</link>
      <guid isPermaLink="true">https://arcusx.ai/blog/2026-03-25-ai-hallucination-rates.html</guid>
      <pubDate>Wed, 25 Mar 2026 00:00:00 GMT</pubDate><category>Thinking</category>
      <description><![CDATA[AI hallucination rates increase with input complexity and ambiguity, concentrating failures in exactly the high-stakes domains where accuracy matters most. The cause is structural, and so is the fix.]]></description>
    </item>
    <item>
      <title>EU AI Act Articles 9–15: A Technical Reading for Engineering Teams</title>
      <link>https://arcusx.ai/blog/2026-03-19-eu-ai-act-technical.html</link>
      <guid isPermaLink="true">https://arcusx.ai/blog/2026-03-19-eu-ai-act-technical.html</guid>
      <pubDate>Thu, 19 Mar 2026 00:00:00 GMT</pubDate><category>Thinking</category>
      <description><![CDATA[Most EU AI Act coverage is written by lawyers. This one is for the engineers who have to implement compliance. A technical walkthrough of Articles 9–15 and what they require architecturally.]]></description>
    </item>
    <item>
      <title>What SR 11-7 Means for AI-Driven Decision Making</title>
      <link>https://arcusx.ai/blog/2026-03-17-sr-11-7.html</link>
      <guid isPermaLink="true">https://arcusx.ai/blog/2026-03-17-sr-11-7.html</guid>
      <pubDate>Tue, 17 Mar 2026 00:00:00 GMT</pubDate><category>Thinking</category>
      <description><![CDATA[SR 11-7 was written for statistical models with inspectable coefficients. LLMs break every assumption the framework rests on. Here's what examination readiness actually requires.]]></description>
    </item>
    <item>
      <title>The Difference Between Logging and Governance in AI Systems</title>
      <link>https://arcusx.ai/blog/logging-vs-governance.html</link>
      <guid isPermaLink="true">https://arcusx.ai/blog/logging-vs-governance.html</guid>
      <pubDate>Mon, 23 Feb 2026 00:00:00 GMT</pubDate><category>Thinking</category>
      <description><![CDATA[Current AI governance platforms track which model was called. They don't trace how it reasoned. That distinction matters more than most organizations realize.]]></description>
    </item>
    <item>
      <title>Introducing EIE: A Protocol for Measuring Epistemic Integrity in AI Systems</title>
      <link>https://arcusx.ai/blog/introducing-eie.html</link>
      <guid isPermaLink="true">https://arcusx.ai/blog/introducing-eie.html</guid>
      <pubDate>Mon, 26 Jan 2026 00:00:00 GMT</pubDate><category>Protocol Release</category>
      <description><![CDATA[We're releasing the Epistemic Integrity Evaluation specification—an open protocol for measuring whether AI systems handle uncertainty honestly, consistently, and proportionally.]]></description>
    </item>
    <item>
      <title>Introducing IRG: A Protocol for Persistent, Structured AI Reasoning</title>
      <link>https://arcusx.ai/blog/introducing-irg.html</link>
      <guid isPermaLink="true">https://arcusx.ai/blog/introducing-irg.html</guid>
      <pubDate>Mon, 26 Jan 2026 00:00:00 GMT</pubDate><category>Protocol Release</category>
      <description><![CDATA[We're releasing the Iterative Reasoning Graph specification—an open protocol for AI systems that reason in explicit, persistent, revisable structures rather than ephemeral token streams.]]></description>
    </item>
  </channel>
</rss>
