<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom">
  <id>https://gaige.dev/feed.xml</id>
  <title>gaige</title>
  <subtitle>new pages and updated guides from gaige.dev, the instrument you check other instruments against</subtitle>
  <updated>2026-08-12T00:00:00Z</updated>
  <author><name>gaige</name></author>
  <link rel="self" type="application/atom+xml" href="https://gaige.dev/feed.xml"/>
  <link rel="alternate" type="text/html" href="https://gaige.dev/"/>
  <entry>
    <id>https://gaige.dev/observatory/</id>
    <title>drift observatory: the committed receipts, rendered live</title>
    <link rel="alternate" type="text/html" href="https://gaige.dev/observatory/"/>
    <updated>2026-08-04T00:00:00Z</updated>
    <summary>every committed receipt on one static page, rebuilt nightly as a pure function of the export directory: aggregates beside the slices they hide, refusals rendered as refusals, zero javascript.</summary>
  </entry>
  <entry>
    <id>https://gaige.dev/refusals.html</id>
    <title>refusals: the honest no, collected</title>
    <link rel="alternate" type="text/html" href="https://gaige.dev/refusals.html"/>
    <updated>2026-08-17T00:00:00Z</updated>
    <summary>every refusal gaige ships, collected: the conformal floor, the memory floor, the quantization gate, admission floors, the absence rule, and why refusing is never priced.</summary>
  </entry>
  <entry>
    <id>https://gaige.dev/receipt-anatomy.html</id>
    <title>receipt anatomy: the reference receipt, read line by line</title>
    <link rel="alternate" type="text/html" href="https://gaige.dev/receipt-anatomy.html"/>
    <updated>2026-08-04T00:00:00Z</updated>
    <summary>every field of a gaige receipt explained from the committed reference record: corpus, instrument, result, thresholds, refusal, fingerprint, and the command that reproduces it.</summary>
  </entry>
  <entry>
    <id>https://gaige.dev/false-positives.html</id>
    <title>false positives: the arithmetic of a wrong flag</title>
    <link rel="alternate" type="text/html" href="https://gaige.dev/false-positives.html"/>
    <updated>2026-08-07T00:00:00Z</updated>
    <summary>false-positive rates, base rates, and the arithmetic that decides how many flags land on humans. cited and committed numbers, worked in the open; measurements, not advice.</summary>
  </entry>
  <entry>
    <id>https://gaige.dev/corpus.html</id>
    <title>corpus: the pile every number stands on</title>
    <link rel="alternate" type="text/html" href="https://gaige.dev/corpus.html"/>
    <updated>2026-08-07T00:00:00Z</updated>
    <summary>what a detection corpus is, how to build one you can vouch for, how to keep it yours, and why the same detector reads 86.0% on one pile and 61.5% on another. ground-up track step 1.</summary>
  </entry>
  <entry>
    <id>https://gaige.dev/why-gaige.html</id>
    <title>why gaige: a $1.8M lesson in unmeasured AI</title>
    <link rel="alternate" type="text/html" href="https://gaige.dev/why-gaige.html"/>
    <updated>2026-08-04T00:00:00Z</updated>
    <summary>a reported $1.8M overrun on a menial AI task, undetected for months: a case study in the failure class calibrated pilots and drift gates exist to catch.</summary>
  </entry>
  <entry>
    <id>https://gaige.dev/instrument.html</id>
    <title>the instrument: ai metrology in practice</title>
    <link rel="alternate" type="text/html" href="https://gaige.dev/instrument.html"/>
    <updated>2026-08-12T00:00:00Z</updated>
    <summary>what the gaige instrument is and commits to: calibration before trust, error bars on every number, refusals where the data cannot support the ask, and the reproduce command that makes every receipt checkable.</summary>
  </entry>
  <entry>
    <id>https://gaige.dev/ai-marking-dates.html</id>
    <title>what changed on august 2, and what did not</title>
    <link rel="alternate" type="text/html" href="https://gaige.dev/ai-marking-dates.html"/>
    <updated>2026-08-03T00:00:00Z</updated>
    <summary>the eu ai act article 50 marking obligations and california ab 853, read against the primary sources: a maintained date table and what detectable and reliable look like as evidence. not legal advice.</summary>
  </entry>
  <entry>
    <id>https://gaige.dev/thresholds.html</id>
    <title>thresholds, worked</title>
    <link rel="alternate" type="text/html" href="https://gaige.dev/thresholds.html"/>
    <updated>2026-08-07T00:00:00Z</updated>
    <summary>empirical versus conformal thresholds, what each one promises, and the base-rate arithmetic that decides how many of your flags are wrong. worked from committed receipts.</summary>
  </entry>
  <entry>
    <id>https://gaige.dev/drift.html</id>
    <title>drift practice</title>
    <link rel="alternate" type="text/html" href="https://gaige.dev/drift.html"/>
    <updated>2026-08-17T00:00:00Z</updated>
    <summary>instrument drift monitoring with registered run series, replicate-measured variance bounds, and per-interval conformal alarms that stay quiet when movement is only noise.</summary>
  </entry>
  <entry>
    <id>https://gaige.dev/fast-detectgpt.html</id>
    <title>fast-detectgpt in practice</title>
    <link rel="alternate" type="text/html" href="https://gaige.dev/fast-detectgpt.html"/>
    <updated>2026-08-04T00:00:00Z</updated>
    <summary>fast-detectgpt in practice: what the curvature criterion measures, how to run it through gaige on your own corpus, and what the committed receipts show, error bars included.</summary>
  </entry>
  <entry>
    <id>https://gaige.dev/binoculars.html</id>
    <title>binoculars in practice</title>
    <link rel="alternate" type="text/html" href="https://gaige.dev/binoculars.html"/>
    <updated>2026-08-04T00:00:00Z</updated>
    <summary>binoculars in practice: how the observer and performer perplexity ratio works, how to run it through gaige, and what the committed receipt shows next to the independent out-of-domain numbers.</summary>
  </entry>
</feed>
