Leaderboard analysis 2026-10-19 (131 runs)

analysis/2026-10-19/leaderboard-analysis.md

@@ -0,0 +1,54 @@
+# Leaderboard analysis, 2026-10-19 14:00 EDT
+
+Source: `GET /api/leaderboard` (snapshot `leaderboard-2026-10-19-1400.json`), 131 heldout-v2 runs, 33 users.
+Untrained reference (published in the same response): Qwen3.8-27B at revision 1d4bf0f, one attempt per task, 4/40
+(industrial 2, office 2).
+
+## Score per run
+
+| Total solved | 0 | 1 | 2 | 3 | 4 | 5 |
+|---|---|---|---|---|---|---|
+| Runs | 4 | 12 | 32 | 48 | 28 | 7 |
+
+Mean 2.80. No run has exceeded 5.
+
+## Per domain (solved out of 5)
+
+| Domain | Mean | Runs with ≥1 | Max | Runs at max |
+|---|---|---|---|---|
+| office-white-collar | 1.37 | 94% | 3 | 3 |
+| industrial-physical-systems | 1.08 | 78% | 3 | 1 |
+| software-engineering | 0.18 | 18% | 1 | 23 |
+| finance-economics | 0.10 | 10% | 1 | 13 |
+| natural-science | 0.03 | 3% | 1 | 4 |
+| media-content-production | 0.02 | 2% | 1 | 3 |
+| mathematics-or-formal-reasoning | 0.02 | 2% | 1 | 2 |
+| cybersecurity | 0.01 | 1% | 1 | 1 |
+
+The leaderboard keeps each user's best per domain, so a run adds to a user's score only if it beats one of those bests.
+For a user holding industrial 2, office 2, software 1, finance 1 and media 1 (rca-umb), the historical chance that a single
+run raises the score is about 8%: office 3 (2%), industrial 3 (1%), science (3%), math (2%), cybersecurity (1%). No run has
+scored 2 in software, finance or media.
+
+The rare-domain hits (science, math, cybersecurity) come from general multi-domain environments, not from environments
+aimed at those domains, so there is no sign that targeting a domain causes them.
+
+## Score by training steps
+
+| Steps | Runs | Mean solved |
+|---|---|---|
+| 0–25 | 4 | 3.75 |
+| 26–75 | 11 | 2.91 |
+| 76–150 | 44 | 2.80 |
+| 151+ | 72 | 2.74 |
+
+The trend favours fewer steps, but the 0–25 bucket has four runs. One of them (10 steps) matches the untrained reference
+exactly. Step count is also mixed up with other differences (short task lists, failures). The untrained reference is a
+single draw, so its 4/40 is itself uncertain; if its true mean is near 3, most of the gap to trained runs is noise.
+
+## Reading
+
+Across 131 runs, no environment has shown training that carries over to held-out SkillsBench beyond run-to-run noise.
+Identical requests have scored 5/3/2 (rca-umb v4), 3/3/1 (dnakhla terminal-desk-v5), 3/3/2 (Profa-QE) and 3/4
+(ajinkyamulay). A light-touch run, one whose tasks are mostly solved so few steps carry a gradient, is a cheap test of
+whether "change the model less" beats the field mean; it cannot exceed the untrained model's own ability.