wassname/jsteer
1
0
Fork 0
mirror of https://github.com/wassname/jsteer.git synced 2026-09-09 11:25:03 +08:00
Code Issues Packages Projects Releases Wiki Activity

nbs/steering_demo: fix persona_soft ans_mass mislabel in qualitative cell

Browse Source 
Fresh-eyes review caught it: 0.895 was the am_edge/base ratio, not the raw ans_mass
(0.503 vs the 0.90*base=0.506 floor). Conclusion unchanged (single-seed miss under the
gate); field name now correct.

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
This commit is contained in:
wassname
2026-07-12 16:33:28 +08:00
co-authored by Claudypoo
parent a428af4413
commit 22613ce40d
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
nbs/steering_demo.py
+2 -2
View File
@@ -176,8 +176,8 @@ def _(mo):
base) -- the 0.965 is read off a slot the model no longer treats as an answer. The dual
gate flags it `readout_ok=False` and the validity-weighted score nulls it (-0.06).
- **The persona methods' `readout_ok=False` is a boundary/noise effect, not over-steer.**
e.g. persona_soft's negative edge lands at ans_mass 0.895 vs the 0.90 floor -- a single-seed
miss, not a collapse. The instrument is single-seed greedy and noisy near the edge;
e.g. persona_soft's negative edge holds ans_mass at 89.5% of base (0.503 vs the 0.90*base =
0.506 floor) -- a single-seed miss, not a collapse. The instrument is single-seed greedy and noisy near the edge;
multi-seed averaging is needed before trusting the flag or the cross-method rank.
**Bottom line:** on this *deliberated* verdict, a fixed per-layer residual offset does not