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(confidence): refuse a boolean probability in any position, not only when it is the only one
Aashishe236098d3fcf
diff --git a/tests/test_confidence.py b/tests/test_confidence.pyindex a11148b2..3ee38351 100644--- a/tests/test_confidence.py+++ b/tests/test_confidence.py@@ -2479,6 +2479,30 @@ for _name, _answer in (check("jev_confidence/none-for-%s" % _name, jev_confidence(_answer), None)+# The table above refuses a document whose probabilities are *all* boolean, which is the one shape a+# `any(isinstance(v, bool) ...)` guard and an `all(...)` one agree on. The positions below are where+# the two disagree, and they are the shapes a client actually sends: `{"spam": true, "ham": 0.4}`+# mixes a JSON boolean with real numbers. Under `all` that document is measured -- `float(True)` is+# 1.0, so the answer comes back as the most confident option the caller never scored -- and+# `/v1/systemone` publishes the number as `x_jev_confidence` for a client to threshold against.+_swept, _measured = 0, []+for _k in (2, 3, 4, 5, 8, 12):+ for _i in range(_k):+ for _flag in (True, False):+ _vals = [1.0 / _k] * _k+ _vals[_i] = _flag+ for _answer in (_choice(*_vals), _score(*_vals)):+ _got = jev_confidence(_answer)+ _swept += 1+ if _got is not None:+ _measured.append((_k, _i, _flag, _answer["type"], _got))+check_true("jev_confidence/a bool in any position of any option count is refused",+ not _measured, "%d of %d measured: %r" % (len(_measured), _swept, _measured[:3]))+# Every one of those documents differs from its all-float twin by exactly one entry, so the count+# is the witness that the sweep ran over the positions it claims.+check("jev_confidence/position sweep size", _swept, sum(2 * 2 * _k for _k in (2, 3, 4, 5, 8, 12)))++def _jev_measured(_answer):"""`jev_confidence`'s number, or the exception it raised instead, so a slip reports, not aborts."""try: