add supplement_20260910/README_supplement.md

supplement_20260910/README_supplement.md

@@ -24,3 +24,21 @@ STEP=3 bash code/run_remote.sh
 ```
 Return only the new rows: `tar -czf pool_returns_<host>_<date>_sup.tar.gz results/pool/step2 results/pool/step3 logs` and upload
 under returns/ as before (the runner appends to the existing step files; sending the whole step2/step3 directories is fine).
+
+## New steps 4-6 (code/ in this supplement replaces code/ of the package)
+
+The pool owner added three removal steps so that the screen covers every level of the paper's pyramid:
+step 4  (Qwen3-VL-8B answers from the options alone, no question, no frames; 4 option permutations,
+same rule as step 1), step 5  (Qwen3-VL-8B on the same 32 frames at 448 px as step 3), step 6
+ (Qwen3-VL-2B on the step-3 frames in one fixed per-item shuffled order). Steps 4-6 run on the
+step-1 survivors that no other step has removed; they do not wait for steps 2/3, so run them after 2 and 3
+finish to avoid wasted work. Yes/no (true/false) items are skipped by every step from now on.
+```bash
+cp -r sup/supplement_20260910/code/. code/            # updated client + launcher (STEP=1..6)
+STEP=4 bash code/run_remote.sh --plan-only && STEP=4 bash code/run_remote.sh    # text only, ~1-2 h on 4 B200
+STEP=5 bash code/run_remote.sh --plan-only && STEP=5 bash code/run_remote.sh    # 8B x 32 frames, the longest step (est. 1 day)
+STEP=6 bash code/run_remote.sh --plan-only && STEP=6 bash code/run_remote.sh    # 2B x 32 shuffled frames, ~half of step 3
+```
+Step 4 is also being run on the origin cluster; if we send you its rows, drop them into results/pool/step4/ and
+ will show what is left. Return  and  as
+ under returns/ (whole step directories are fine).