From d56d41279bd379a9827d7feffa1204821e453f0b Mon Sep 17 00:00:00 2001
From: rustfuture <121799572+rustfuture@users.noreply.github.com>
Date: Wed, 30 Sep 2026 07:46:10 +0000
Subject: [PATCH] docs: describe held-out scoring as the code performs it

The held-out split is scored twice, both after selection ends: once for the baseline policy and once for the final policy (rsi_framework/core.py). README and DESIGN.md said "measured once".
---
 DESIGN.md | 3 ++-
 README.md | 2 +-
 2 files changed, 3 insertions(+), 2 deletions(-)

diff --git a/DESIGN.md b/DESIGN.md
index 2f91e95..ddd9726 100644
--- a/DESIGN.md
+++ b/DESIGN.md
@@ -69,7 +69,8 @@ The candidate loop implements monotonic greedy local search:
    ever reverted: this is candidate rejection, not rollback of an applied state. The
    historical label `rollback_regression` appears only in v1 archives. The rule is a
    fixed-tolerance comparison, not a confidence-interval or trust-region barrier.
-5. **Held-out Isolation**: the held-out split is measured once, after selection stops,
+5. **Held-out Isolation**: the held-out split is scored only after selection stops (once
+   for the baseline policy and once for the final policy, both in `rsi_framework/core.py`),
    and never enters generation, scoring, or the rejection rule.
 
 ## Scoped Data-Isolation Check
diff --git a/README.md b/README.md
index 4c99637..cd2e661 100644
--- a/README.md
+++ b/README.md
@@ -13,7 +13,7 @@ RSI (recursive self-improvement) is the research question behind the name. This
 - Generates, evaluates, and selects keyword-rule changes on synthetic text.
 - Checks each proposed change against vocabulary, polarity, and bias rules.
 - Rejects changes that reduce accuracy on development data.
-- Measures held-out data once after selection ends.
+- Scores held-out data only after selection ends (once for the baseline policy and once for the final policy).
 - Writes reproducible run records and checked result summaries.
 
 ## Quick start