diff --git a/scripts/run_modal_wvs.py b/scripts/run_modal_wvs.py
index 3b993dc..c31c1b2 100644
--- a/scripts/run_modal_wvs.py
+++ b/scripts/run_modal_wvs.py
@@ -77,7 +77,9 @@ def main(model: str = MODEL, methods: str = ",".join(METHODS), seeds: str = ",".
             res = json.loads(handle.get())
             base = res["doses"][0]
             far = max(res["doses"], key=lambda d: abs(d["mult"]))
+            check = res["manipulation_check"]["scored"]
             print(f"{method}\ts{seed}\tC={res['calibrated_C']:+.3f}\t"
+                  f"honesty={check['effect_logodds']:+.3f}\t"
                   f"dx={far['x'] - base['x']:+.4f}\tdy={far['y'] - base['y']:+.4f}\t"
                   f"pmass={far['mean_pmass']:.3f}")
         except Exception as error:
diff --git a/slop/plans/20260918_wvs_steer_big.md b/slop/plans/20260918_wvs_steer_big.md
index f4a14c7..d183351 100644
--- a/slop/plans/20260918_wvs_steer_big.md
+++ b/slop/plans/20260918_wvs_steer_big.md
@@ -70,6 +70,23 @@ leave-one-out is one item reacting to the persona wording, not a cultural shift.
 This mattered most for credulity, which nearly paraphrases the X-axis trust item
 (`"Most people can be trusted"`). Credulity is now out of scope, but the check is cheap and stays.
 
+## Corrected-run decision rule (recorded before the next full sweep)
+
+The first 27B sweep found two evaluation bugs: sampled think traces were not seeded before a
+"paired" bootstrap, and VJP's behavioral sign was opposite its configured sign. Its baseline
+pmass was also 0.75, below the preregistered 0.95 readability gate. Those outputs are diagnostic,
+not the final result.
+
+1. Use the largest probed model with base mean pmass >= 0.95. A plotted dose must keep mean pmass
+   >= 0.90; do not replace this with a post-hoc relative threshold.
+2. Orient every vector, including random, on four held-out true-vs-welcome questions. A method is
+   honesty-specific only if its log-odds effect exceeds the 95th percentile over random directions.
+3. A culture-map effect must beat the matched-dose random displacement and retain its direction
+   after the single most influential WVS item is removed. Opposite doses should point to opposite
+   half-planes locally; a non-monotonic or same-direction trajectory is evidence against an axis.
+4. If no implementation passes these checks, conclude only that these implementations are
+   inconclusive. Do not conclude that honesty has no cultural effect.
+
 ## Goals
 
 0. [/] goal: the target model's answer slot is readable, before renting anything big