v9 result

results/v9/RESULT.md

@@ -0,0 +1,25 @@
+# five-desks-v9 result (unchanged replicate of v4)
+
+Run `8eceb03347c6f626626db1c4`, submitted 2026-10-17 18:29 EDT; evaluation done 21:41 EDT.
+
+**Private score: 3/40** (industrial 2, office 1; delta -1 against the 4/40 reference).
+
+| | v4 (2026-10-12) | v9 (2026-10-17) |
+|---|---|---|
+| Request | five-desks-v4 | identical apart from id/name |
+| Training steps | 68 | 82 |
+| Steps with a gradient | 29 (43%) | 45 (55%) |
+| Private score | 5/40 (industrial 2, office 2, software 1) | 3/40 (industrial 2, office 1) |
+
+## Reading
+
+- The same request scored 5 and then 3. That two-point swing matches the field's identical-request replicates
+  (naklitechie #229: 1 then 0; Profa-QE #256: 3 then 3 with domains reshuffled) and the unchanged-model spread
+  measured earlier (SD about 1.4). v4's 5/40 was partly a good draw.
+- Industrial 2 repeated; it is the domain where our runs are most consistent.
+- No new per-domain best, so the leaderboard total is unchanged.
+
+## Training
+
+Reward by cycle (9 tasks per cycle): 0.17, 0.14, 0.36, 0.19, 0.25, 0.33, 0.31, 0.42, 0.42 over nine full cycles, plus one extra office visit. `bugfix-office-l1` rose from 0 to 1.0 by the last visits;
+`bugfix-software-l2` had a gradient on only one visit. Per-task table: `per-task.txt`.

results/v9/per-task.txt

@@ -0,0 +1,11 @@
+steps 82, steps with a gradient 45, tasks visited 9/9
+task                                     visits grad  rewards by visit
+bugfix-office-l1                             10    5  0.00 0.00 0.25 0.25 0.00 0.25 0.50 1.00 0.50 1.00
+bugfix-finance-l1                             9    7  0.00 0.00 0.75 0.50 0.50 0.50 0.25 0.50 0.75
+bugfix-science-l1                             9    4  0.00 0.25 0.00 0.00 0.00 0.75 0.25 0.50 0.00
+bugfix-industrial-l2                          9    5  0.00 0.00 0.50 0.25 0.25 0.25 0.00 0.25 0.00
+bugfix-media-l2                               9    6  0.75 0.75 0.50 0.00 0.50 0.75 0.00 0.25 0.00
+bugfix-security-l2                            9    5  0.00 0.00 0.50 0.00 0.50 0.00 0.75 0.50 0.50
+bugfix-software-l2                            9    1  0.00 0.00 0.00 0.00 0.00 0.00 0.25 0.00 1.00
+bugfix-math-l1                                9    6  0.25 0.00 0.00 0.00 0.50 0.50 0.50 0.75 0.50
+media-production-l2                           9    6  0.50 0.25 0.75 0.75 0.00 0.00 0.25 0.00 0.50