Skip to content

armgit / laya

Public Imported from GitHub
0

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.

fix(evidence): a boolean `items` is a malformed calibration entry, not a count

Aashishab6d8f2610d9
diff --git a/laya/evidence.py b/laya/evidence.py
index 69692a63..5d609c0a 100644
--- a/laya/evidence.py
+++ b/laya/evidence.py
@@ -44,7 +44,11 @@ def inspect_calibration(config: Dict[str, Any]) -> Dict[str, Any]:
return _entry(UNKNOWN, "training metadata present but no calibration block; checkpoint predates #933")
types: Dict[str, Any] = {}
for name, entry in calibration.items():
- if not isinstance(entry, dict) or not isinstance(entry.get("items"), int):
+ # `isinstance(True, int)` is True, so a JSON `true` in `items` read as a count of one and a
+ # `false` as zero. Both are type errors, unlike a negative count, which stays reported
+ # because #964 deleted the floor that judged the number.
+ if (not isinstance(entry, dict) or isinstance(entry.get("items"), bool)
+ or not isinstance(entry.get("items"), int)):
types[name] = _entry(UNKNOWN, "calibration entry is malformed")
continue
items = entry["items"]
diff --git a/tests/test_evidence.py b/tests/test_evidence.py
index 9d9a92fb..99af9bf7 100644
--- a/tests/test_evidence.py
+++ b/tests/test_evidence.py
@@ -81,6 +81,27 @@ check("mixed types overall -> INSUFFICIENT", mixed["state"], evidence.INSUFFICIE
malformed = calib({"training": {"laya_train_calibration": {"choice": {"temperature": 1.0}}}})
check("malformed entry -> UNKNOWN", malformed["types"]["choice"]["state"], evidence.UNKNOWN)
+# A JSON `true` is not a count, but `isinstance(True, int)` makes it one. Before this guard
+# `{"items": true}` answered PRESENT with "fitted with no recorded issues on 1 items" and
+# `{"items": false}` answered MISSING with "zero calibration items": two verdicts about a fit that
+# the artifact never reported, and the second one is the state a genuinely unfitted type gets.
+for _flag in (True, False):
+ _t = calib({"training": {"laya_train_calibration": {
+ "choice": {"items": _flag, "temperature": 1.1, "issues": []}}}})
+ check("items=%r is a malformed entry, not a count" % _flag, _t["types"]["choice"]["state"],
+ evidence.UNKNOWN)
+ check("items=%r keeps the malformed reason" % _flag, _t["types"]["choice"]["reason"],
+ "calibration entry is malformed")
+ check("items=%r leaves the overall undecided" % _flag, _t["state"], evidence.INSUFFICIENT)
+
+# The guard is about the type of `items`, not its value: a negative count stays reported, because
+# #964 deleted the local floor that judged it and nothing here is supposed to re-introduce one.
+_neg = calib({"training": {"laya_train_calibration": {
+ "choice": {"items": -5, "temperature": 1.1, "issues": []}}}})
+check("negative items are still reported, not judged", _neg["types"]["choice"]["state"], evidence.PRESENT)
+check("negative items keep their own count in the reason", _neg["types"]["choice"]["reason"],
+ "fitted with no recorded issues on -5 items")
+
# ------------------------------------------------------ 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