From d0d97d05979d2b81522c3612bc1490341151240b Mon Sep 17 00:00:00 2001 From: ciregenz Date: Wed, 12 Aug 2026 21:51:06 -0700 Subject: [PATCH] arena: CompWoB ours 65.3 -- collapses LESS than the published specialist reference 90.9 MiniWoB -> 65.3 CompWoB beats the 95->61 specialist collapse ratio, zero false claims, clean length gradient (82% pairs, 41% triples, 0% beyond five). The 75 target needs the long-horizon mechanism the gradient points at; browser-use's identical run is next, then v23. Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_01WsbS5x2rYsMDxP2kW3qqmQ --- e2e/browser-v3/arena/ARENA.md | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/e2e/browser-v3/arena/ARENA.md b/e2e/browser-v3/arena/ARENA.md index 647af2e8..82f9a0f0 100644 --- a/e2e/browser-v3/arena/ARENA.md +++ b/e2e/browser-v3/arena/ARENA.md @@ -133,6 +133,15 @@ The 13 residual losses are the irreducible product-primitive cluster (games, con geometry, two long forms) plus a residual empty-reply subclass (model prose with no action at all -- prefill-forcing is the structural close). Seed-43 confirmation in flight. +## CompWoB — the generalization verdict (101 composed tasks, benchmark-scored) + +**Ours (v22, opus-5): 66/101 = 65.3%, 0 false claims, 17.4s median win.** Reference: the +published 95%-MiniWoB specialist collapsed to ~61% here; our 90.9→65.3 is a smaller relative +collapse, on a harness whose two instrument bugs (legacy engine, literal-URL validate) were +caught by canary before any number shipped. Gradient: 2-part 82%, 3-part 41%, 5+-part 0% -- +the long-horizon sequencing frontier, plus a traced-pending cluster of simple-pair losses. +browser-use runs the identical 101 next (their first known CompWoB number). + ## Positioning vs public generic-harness baselines (user-supplied 2026 survey) The comparable class is generic agents, NOT MiniWoB-specialized systems (HTML-T5++ 95.2 trained