From fe1b7f35ebae0c8e22f7aa50e298b24ed815a056 Mon Sep 17 00:00:00 2001
From: vaaraio <267591518+vaaraio@users.noreply.github.com>
Date: Fri, 22 May 2026 18:06:07 +0300
Subject: [PATCH] fix(docs): drop Recital 133 misattribution from conformal
 explainer

Recital 133 of the EU AI Act is about synthetic content watermarking under Article 50, not detection of non-conformities or accuracy verification. The conformal-prediction.md and v0.28.0 CHANGELOG entry both cited it incorrectly. Dropping the recital reference; the Article 15(1) anchor stands on its own and is verified against the AI Act text.
---
 CHANGELOG.md                 |  5 ++---
 docs/conformal-prediction.md | 12 ++++--------
 2 files changed, 6 insertions(+), 11 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index 5d68095..f6c0da8 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -33,9 +33,8 @@ ships both.
   `docs/formal_specification.md`. Explains why a point risk score is
   not enough, what the interval gives a reader, and how the
   distribution-free coverage guarantee maps to Article 15(1)
-  ("appropriate level of accuracy") and Recital 133 (detection of
-  non-conformities). Cross-linked from the README "Where things live"
-  table and from `VERDICTS.md` Article 15(1) discussion.
+  ("appropriate level of accuracy"). Cross-linked from the README
+  "Where things live" table.
 
 ### Changed
 - `COMPLIANCE.md` "EU AI Act Article Mapping" intro now points readers
diff --git a/docs/conformal-prediction.md b/docs/conformal-prediction.md
index 0d08561..f03e22e 100644
--- a/docs/conformal-prediction.md
+++ b/docs/conformal-prediction.md
@@ -52,8 +52,6 @@ makes the difference visible.
 
 ## Why this matters under the EU AI Act
 
-Two parts of the AI Act point at this directly.
-
 **Article 15(1)** requires high-risk AI systems to "be designed and
 developed in such a way that they achieve an appropriate level of
 accuracy, robustness and cybersecurity, and perform consistently in
@@ -65,12 +63,10 @@ deployer can publish the guarantee (e.g. "the interval covers the true
 risk at least 90% of the time") and an auditor can check it against
 observed outcomes.
 
-**Recital 133** discusses the detection of "violations or
-non-conformities" by AI systems and the need to "regularly verify the
-results obtained." A point estimate does not surface non-conformity
-until after the fact. An interval that widens or narrows is itself a
-real-time signal that the model is moving into or out of a region
-where its predictions are trustworthy.
+A widening interval is itself a real-time signal that the model is
+moving into a region where its predictions cannot be trusted. A point
+estimate does not surface that drift until ground-truth labels arrive
+and the error rate is reconstructed after the fact.
 
 In short: a conformal interval converts "trust me, the score is 0.6"
 into "the score is 0.6, here is the range we expect to be in, here is