fix(v1): the proposer contract tolerates LaTeX escapes

First live multi-agent training run: Qwen3-4B proposers reliably end with a
well-formed contract whose problem text carries LaTeX (\( S \), \frac) —
off-spec JSON escapes that failed every such episode. One-retry fixup
doubles any backslash that isn't a valid JSON escape (valid escapes are
consumed whole, so \\ stays itself).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: hallerite

environments/proposer_solver_v1/proposer_solver_v1/taskset.py

@@ -53,6 +53,22 @@ class SolveData(vf.TaskData):
     """The proposer's verified answer — the minted task's ground truth."""
 
 
+def _loads_contract(chunk: str) -> object:
+    """``json.loads`` with one retry that doubles any backslash not part of a
+    valid JSON escape — a math proposer writes LaTeX (``\\(``, ``\\frac``) inside
+    the contract, which is off-spec JSON but an unambiguous fixup. Valid escapes
+    are consumed whole so ``\\\\`` stays itself."""
+    try:
+        return json.loads(chunk)
+    except json.JSONDecodeError:
+        fixed = re.sub(
+            r'\\(?:[\\"/bfnrtu]|u[0-9a-fA-F]{4})|\\',
+            lambda m: m.group(0) if len(m.group(0)) > 1 else "\\\\",
+            chunk,
+        )
+        return json.loads(fixed)
+
+
 class SolveTask(vf.Task[SolveData]):
     @classmethod
     def from_trace(cls, proposer: vf.Trace) -> "SolveTask":
@@ -69,7 +85,7 @@ def from_trace(cls, proposer: vf.Trace) -> "SolveTask":
             if not (chunk.startswith("{") and chunk.endswith("}")):
                 continue
             try:
-                parsed = json.loads(chunk)
+                parsed = _loads_contract(chunk)
             except json.JSONDecodeError:
                 continue
             if isinstance(parsed, dict) and {"problem", "answer"} <= parsed.keys():

tests/v1/test_bundled_envs.py

@@ -351,3 +351,16 @@ def score(hits: int, n: int = 4) -> Trace:
     assert score(2).rewards["learnability"] == 1.0
     assert score(4).rewards["learnability"] == 0.0
     assert score(1).metrics["solve_rate"] == 0.25
+
+
+def test_solve_task_contract_tolerates_latex_escapes():
+    """Math proposers write LaTeX inside the JSON contract; the parse doubles the
+    off-spec backslashes instead of failing the episode."""
+    from proposer_solver_v1.taskset import SolveTask
+
+    trace = _reply_trace(
+        'Here it is:\n{"problem": "Let \\( S \\) be a set with \\\\leq 5 elements.", "answer": 3}'
+    )
+    task = SolveTask.from_trace(trace)
+    assert task.data.answer == "3"
+    assert "\\( S \\)" in task.data.prompt