From a091cc752785af98f17b130aa497e675552e8fbc Mon Sep 17 00:00:00 2001
From: wassname <1103714+wassname@users.noreply.github.com>
Date: Fri, 10 Jul 2026 20:37:05 +0800
Subject: [PATCH] persona_topk: demote per-layer |J^T w| trace to debug; add
 short README demo

- the pre-norm per-layer norm line is a fit-health trace, not demo output -> logger.debug
- README: short 'Persona j-thoughts' section showing the contrast-first extraction
  (clean positive/negative tokens), kept in the experimental/untested-specificity frame

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
---
 README.md          | 23 +++++++++++++++++++++++
 jsteer/jacobian.py |  6 ++++--
 2 files changed, 27 insertions(+), 2 deletions(-)

diff --git a/README.md b/README.md
index 80ab05b..9ba37bf 100644
--- a/README.md
+++ b/README.md
@@ -69,6 +69,29 @@ concentrated direction, so its knee is steep: C~0.5 moves the tone while the tex
 and reasoning stay fluent, and by C~1 it degenerates into token spam.
 `nbs/word_steering.ipynb` shows the full sweep with the j-space and `<think>` views.
 
+## Persona j-thoughts (experimental)
+
+Instead of naming words, contrast two personas: `persona_topk_vector` reads the
+tokens their final-layer means evoke *differently*. Contrasting the logits before
+the top-k is what makes it work; both persona means unembed to the same generic
+tokens (`\n`, ` the`, ` I`), so the signal lives only in the difference.
+
+```python
+optimist  = ["Things usually work out better than people expect.", ...]
+pessimist = ["Things usually go worse than people expect.", ...]
+v = jac.persona_topk_vector(model, tok, optimist, pessimist, layers=band)
+# logs the contrastive "mental workspace":
+#   j-thoughts (content of mental workspace, top-8)
+#       positive: [' ❤', '😊', ' happy', '✨', ' Happy', ' 🙂', ' grat', ' favorite']
+#       negative: [' Worse', '绝望', ' Panic', ' useless', ' Worst', ' worse', '无力', ' panic']
+show_steer(jac, model, tok, v, "How is the project going?", Cs=(0, 0.5, 1.5))
+```
+
+The extraction is clean, but persona *steering* is unverified: earlier persona
+vectors failed specificity controls (Evidence section below), and this
+contrast-first variant has not been re-tested. Trust `word_vector`; treat this as
+a toy. `nbs/persona_steering.ipynb` runs all three persona variants.
+
 ## API
 
 | call | status | what it does |
diff --git a/jsteer/jacobian.py b/jsteer/jacobian.py
index 16da357..00a0641 100644
--- a/jsteer/jacobian.py
+++ b/jsteer/jacobian.py
@@ -222,8 +222,10 @@ class Jacobian:
         if missing:
             raise ValueError(f"layers {sorted(missing)} not fitted; have {self.layers}")
         per_layer = {l: w @ self.lens.jacobians[l] for l in cfg.layers}
-        logger.info(f"{cfg.method} per-layer |J^T w| (pre-norm): " +
-                    " ".join(f"{l}:{per_layer[l].norm():.3g}" for l in cfg.layers))
+        # per-layer pullback norms before unit-normalizing: a fit-health trace
+        # (flat/near-zero everywhere => the lens didn't pick up this cotangent).
+        logger.debug(f"{cfg.method} per-layer |J^T w| (pre-norm): " +
+                     " ".join(f"{l}:{per_layer[l].norm():.3g}" for l in cfg.layers))
         return _to_vector(cfg, per_layer)
 
     def steer_band(self, model, *, lo: float = 0.3, hi: float = 0.9) -> tuple[int, ...]: