Non-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per request.
Python calibrationclassificationdecision-modelhuggingfacejevmodernbertmultilingualnlppythonpytorchroutingtyped-decisionszero-shot
test(predict_long): pin the exact-equality boundary of the budget check
Adeshina Lasisi8d5b9a97f368
diff --git a/tests/test_predict_long.py b/tests/test_predict_long.pyindex 30f10c92..f0960d82 100644--- a/tests/test_predict_long.py+++ b/tests/test_predict_long.py@@ -897,6 +897,13 @@ check_raises("hook budget/a hook that shrinks the room is refused", ValueError,check_raises("hook budget/rewriting ctx.questions is caught too", ValueError,lambda: _check_scan_budget(_probe_agent, _evidence(questions={"a": q_many(16)}),72, *_CFG, asked={"a": q_many(2)}))+# The boundary is `>=`, not `>`: a re-budget that leaves the room EXACTLY the size the scan used+# is harmless and must still answer. At cfg 100/20 a 4-option question leaves exactly 72, which is+# what a 2-option scan sized to, so this is the one shape that separates the two comparisons.+check_true("hook budget/a room exactly equal to the window is allowed",+ _check_scan_budget(_probe_agent, _evidence(questions={"a": q_many(4)}), 72, *_CFG,+ asked={"a": q_many(2)}) is None, "room == sized")+check_true("hook budget/questions untouched skips the recompute entirely",_check_scan_budget(_probe_agent, _evidence(questions={"a": q_many(16)}), 72, *_CFG,asked={"a": q_many(16)}) is None, "nothing moved")