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(evidence): pin the entry shapes the malformed door already refuses
Aashishf9e7dfda4938
diff --git a/tests/test_evidence.py b/tests/test_evidence.pyindex 99af9bf7..11fc65b3 100644--- a/tests/test_evidence.py+++ b/tests/test_evidence.py@@ -102,6 +102,18 @@ check("negative items are still reported, not judged", _neg["types"]["choice"]["check("negative items keep their own count in the reason", _neg["types"]["choice"]["reason"],"fitted with no recorded issues on -5 items")+# The rest of the same door, which main already refuses but nothing pinned: a string or a float is+# not a count, and an entry that is not an object cannot carry one. All three land on the same+# `UNKNOWN` as a missing key, because none of them states a number the fit can be checked against.+for _bad, _label in (("5", "a string"), (2.0, "a float")):+ _t = calib({"training": {"laya_train_calibration": {+ "choice": {"items": _bad, "temperature": 1.1, "issues": []}}}})+ check("items=%s is not a count -> UNKNOWN" % _label, _t["types"]["choice"]["state"], evidence.UNKNOWN)+_not_entry = calib({"training": {"laya_train_calibration": {"choice": "5"}}})+check("entry that is not an object -> UNKNOWN", _not_entry["types"]["choice"]["state"], evidence.UNKNOWN)+check("entry that is not an object keeps the malformed reason", _not_entry["types"]["choice"]["reason"],+ "calibration entry is malformed")+# ------------------------------------------------------ one numeric source of truth# `items` is reported here, never judged: #964 deleted the local `_thresholds()` that compared it# against copies of `MIN_TYPE_N`/`CALIB_WARN_N`, and the comment describing that mechanism stayed