From 92f25093ebfb2f52b009ba8753a1fba6f1a5fd67 Mon Sep 17 00:00:00 2001
From: wassname <1103714+wassname@users.noreply.github.com>
Date: Fri, 18 Sep 2026 21:13:18 +0800
Subject: [PATCH] Select Qwen3-14B from measured answer-slot readability

At the 64-token resample, Qwen3-14B retained mean/min answer-token mass
0.987/0.948. Qwen3-8B collapsed to 0.567/0.009 and Qwen3-4B was smaller.
Record the evidence and retain the WVS source label on the revised figure.

Co-Authored-By: Claude <288921227+claudypoo@users.noreply.github.com>
---
 scripts/plot_wvs_steer.py            | 2 +-
 slop/plans/20260918_wvs_steer_big.md | 7 +++++--
 2 files changed, 6 insertions(+), 3 deletions(-)

diff --git a/scripts/plot_wvs_steer.py b/scripts/plot_wvs_steer.py
index 8188829..49c994d 100644
--- a/scripts/plot_wvs_steer.py
+++ b/scripts/plot_wvs_steer.py
@@ -128,7 +128,7 @@ def main() -> None:
         ("Survival", "Self-expression", "Traditional", "Secular-Rational"),
         models={f"{model.split('/')[-1]} (base)": (base["x"], base["y"])}, emphasize=emph,
         title=f"Honesty steering on the culture map\n{model.split('/')[-1]}",
-        note="Filled: pmass >= 0.90 | hollow: failed coherence gate",
+        note="World Values Survey | filled: pmass >= 0.90 | hollow: failed coherence gate",
         title_y=0.115, note_y=0.04)
     ax = fig.axes[0]
 
diff --git a/slop/plans/20260918_wvs_steer_big.md b/slop/plans/20260918_wvs_steer_big.md
index d183351..183c308 100644
--- a/slop/plans/20260918_wvs_steer_big.md
+++ b/slop/plans/20260918_wvs_steer_big.md
@@ -89,7 +89,7 @@ not the final result.
 
 ## Goals
 
-0. [/] goal: the target model's answer slot is readable, before renting anything big
+0. [x] goal: the target model's answer slot is readable, before renting anything big
    - subtle failure mode: the coordinate looks plausible while most of the answer-token mass sits
      off the digits, so every steered move is measured through mush
    - discriminator: mean pmass_allowed >= 0.95 on the unsteered battery. Qwen3-0.6B reads 1.000,
@@ -102,10 +102,13 @@ not the final result.
      - > Qwen3.5 chat template closes an empty think block by default, so the reader's own `<think>`
        > made `</think> ... <think>`. Fixed with enable_thinking=True; worth only +0.02 to +0.09 pmass,
        > so the template was not the main cause. Qwen3-0.6B unchanged at 1.000 (no regression).
+     - > `logs_modal_qwen3_readable.log`, 64-token, 8-sample resample: Qwen3-4B mean/min
+       > pmass 0.979/0.880; Qwen3-8B 0.567/0.009; Qwen3-14B 0.987/0.948. Qwen3-14B is the
+       > largest candidate that passes the preregistered mean >= 0.95 gate.
    - tasks:
      1. [x] probe think budget 1/16/64/256 on the new family
      2. [x] fix the double-think template artifact
-     3. [/] probe Qwen3.5-27B and Qwen3-32B on Modal, pick on the measured number
+     3. [x] probe candidate large models on Modal and select Qwen3-14B from measured pmass
 
 1. [x] goal: one source of truth for the WVS battery readout, importable outside `scripts/`
    - subtle failure mode: the steer script gets its own copy of the item resolution, the two