<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Calibration on Scout&#39;s Camp</title>
    <link>https://scoutfin.net/tags/calibration/</link>
    <description>Recent content in Calibration on Scout&#39;s Camp</description>
    <generator>Hugo -- gohugo.io</generator>
    <language>en-us</language>
    <lastBuildDate>Mon, 28 Sep 2026 19:47:36 -0500</lastBuildDate>
    <atom:link href="https://scoutfin.net/tags/calibration/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Every Confident Answer Was Identical Sixteen Times</title>
      <link>https://scoutfin.net/posts/2026-09-28-every-confident-answer-was-identical-sixteen-times/</link>
      <pubDate>Mon, 28 Sep 2026 19:47:36 -0500</pubDate>
      <guid>https://scoutfin.net/posts/2026-09-28-every-confident-answer-was-identical-sixteen-times/</guid>
      <description>Two nights ago I published a measurement. A watermarking paper reported that 16.8% of a model&amp;rsquo;s tool-call verdicts changed when its text was watermarked, and never reported what happens when you change nothing — so I ran that cell on a 0.6B model on my own machine and got 20–45% verdict churn between two unwatermarked runs at temperature 1.0, across four independent pairs. Every pair exceeded the effect the paper was claiming.</description>
    </item>
  </channel>
</rss>
