arena: v33 verdict (fastpath is the reverse-task killer) + v34 inversion gate (pre-registered)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WsbS5x2rYsMDxP2kW3qqmQ

e2e/browser-v3/arena/ARENA.md

@@ -407,6 +407,20 @@ history. Targets: exactly the 12 tasks they win and we lose (v33_pilot.json) + 1
 controls. Prediction: >=5 of 12 targets (the 7 reverse-order ones are the core candidates),
 controls 12/12. Pilot runs only after the MiniWoB regression frees the LLM lane.
 
+VERDICT v33 (2026-08-14): **fail — targets 1/12, controls 12/12.** Deferral never fires on
+reverse tasks because nothing ERRORS: the fastpath instantly clicks the goal's first-QUOTED
+target, which under 'X, after doing Y' grammar is the last-executed step — 1-step terminal
+losses, no model call involved. The model, when consulted, parses the inversion correctly
+(it opened the dialog case in the right order unaided). Mechanism withdrawn.
+
+## PRE-REGISTERED (2026-08-14, before any v34 episode): fastpath inversion gate
+
+Mechanism: v34 = v32 + one gate — fastpath stands down when the goal contains subordinate
+order conjunctions (after/before/once), sending those sentences to the model. Linguistic
+feature-trigger, no task names, no new scaffolding; normal goals keep the scripted speed path
+(unit-checked both ways). Prediction: >=3 of the 7 reverse-order targets flip (the two 1-step
+losses at minimum), controls 12/12 (no control goal contains the conjunctions).
+
 ## Benchmark roadmap (2026 landscape survey, method-filtered)
 
 Rules: third-party scoring, reproducible from a committed artifact, no LLM-judge (or deterministic

e2e/browser-v3/arena/llm_policy.py

@@ -424,6 +424,12 @@ class OpenSwarmLlmPolicy(LlmPolicy):
         targets = [t for t in quoted(goal) if t and t not in self.fastpath_used]
         if not targets:
             return ""
+        # v34: subordinate order clauses ('X, after doing Y' / 'before' / 'once') make stated
+        # order != execution order. A scripted first-quoted click is then a coin flip at best and
+        # an instant terminal loss at worst (measured: 1-step losses on every reverse variant).
+        # The model parses these sentences correctly when consulted -- so consult it.
+        if re_.search(r"\b(after|before|once)\b", goal, re_.I):
+            return ""
         # ORDER IS PART OF THE TASK: composed goals ("click Ok, then the link") enforce sequence,
         # and skipping to a later quoted target because the first is ambiguous clicked things out
         # of order -- an instant, unrecoverable loss (measured on ordered pairs). The fastpath may
@@ -924,6 +930,15 @@ def build(name: str, model: str = "", endpoint: str = "", **_: Any) -> Any:
                                   scripted_drag=True, auto_complete=True, som=False,
                                   native_pickers=True, verify_terminal=True, post_mouse_vision=True,
                                   multi_cap=6, fill_verify=True, **v17)
+    if name == "osw-llm-v34":  # v32 + fastpath inversion gate (subordinate-clause goals go to the model)
+        v34 = dict(v7, system=OSW_SYSTEM_V8 + OSW_SYSTEM_V9_WIDGETS + OSW_SYSTEM_V16 + OSW_SYSTEM_V30,
+                   max_tokens=800)
+        return OpenSwarmLlmPolicy(name=name, multi=True, vision="progressive", fastpath=True,
+                                  scripted_drag=True, auto_complete=True, som=False,
+                                  native_pickers=True, verify_terminal=True, post_mouse_vision=True,
+                                  multi_cap=6, fill_verify=True, dispatch=True, offscreen=True,
+                                  local_ctx=True, blocker_probe=True, suppress_wrappers=True,
+                                  force_unblock=True, **v34)
     if name == "osw-llm-v33":  # v32 + opportunistic ordering (deferral doctrine + defer-nudge)
         v33 = dict(v7, system=OSW_SYSTEM_V8 + OSW_SYSTEM_V9_WIDGETS + OSW_SYSTEM_V16 + OSW_SYSTEM_V30
                    + OSW_SYSTEM_V33, max_tokens=800)