loopgain 0.4.1__tar.gz → 0.4.2__tar.gz

{loopgain-0.4.1 → loopgain-0.4.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: loopgain
-Version: 0.4.1
+Version: 0.4.2
 Summary: An open-source cost controller for AI agent loops. Stops a loop when it has actually converged and rolls back before it degrades — replacing the max_iterations guess with a real-time loop-gain (Aβ) monitor with five named threshold bands and best-so-far rollback.
 Author-email: Dave Fitzsimmons <hello@loopgain.ai>
 License: Apache-2.0
@@ -58,7 +58,7 @@ AI agent loops waste time and money when they don't know when to stop. LoopGain
 [![PyPI](https://img.shields.io/pypi/v/loopgain.svg)](https://pypi.org/project/loopgain/)
 [![Python](https://img.shields.io/pypi/pyversions/loopgain.svg)](https://pypi.org/project/loopgain/)
 [![License](https://img.shields.io/badge/license-Apache_2.0-blue.svg)](LICENSE)
-[![Tests](https://img.shields.io/badge/tests-202_passing-brightgreen.svg)](tests/)
+[![Tests](https://img.shields.io/badge/tests-200%2B_passing-brightgreen.svg)](tests/)
 
 **Home:** [loopgain.ai](https://loopgain.ai)
 

{loopgain-0.4.1 → loopgain-0.4.2}/README.md

@@ -9,7 +9,7 @@ AI agent loops waste time and money when they don't know when to stop. LoopGain
 [![PyPI](https://img.shields.io/pypi/v/loopgain.svg)](https://pypi.org/project/loopgain/)
 [![Python](https://img.shields.io/pypi/pyversions/loopgain.svg)](https://pypi.org/project/loopgain/)
 [![License](https://img.shields.io/badge/license-Apache_2.0-blue.svg)](LICENSE)
-[![Tests](https://img.shields.io/badge/tests-202_passing-brightgreen.svg)](tests/)
+[![Tests](https://img.shields.io/badge/tests-200%2B_passing-brightgreen.svg)](tests/)
 
 **Home:** [loopgain.ai](https://loopgain.ai)
 

{loopgain-0.4.1 → loopgain-0.4.2}/loopgain/_version.py

@@ -7,4 +7,4 @@ from here so the value never drifts between ``__version__`` and the
 ``pyproject.toml``) for each release.
 """
 
-__version__ = "0.4.1"
+__version__ = "0.4.2"

{loopgain-0.4.1 → loopgain-0.4.2}/loopgain/classifier.py

@@ -184,9 +184,14 @@ def _two_sided_t_p(t_abs: float, df: int) -> float:
         # exact: cdf_t(t,1) = 0.5 + arctan(t)/pi
         return 2.0 * (0.5 - math.atan(t_abs) / math.pi)
     if df == 2:
-        # exact one-sided survival: 1 - (1 + t²/2)^(-1) doubled
-        return min(1.0, 2.0 * (1.0 - t_abs / math.sqrt(2.0 + t_abs * t_abs) / 1.0) * 0.5
-                   + 2.0 * (0.5 - 0.5 * t_abs / math.sqrt(2.0 + t_abs * t_abs)))
+        # Exact two-sided p-value for Student-t with df=2. The df=2 CDF is
+        # F(t) = 1/2 + t / (2·√(2 + t²)), so the one-sided survival is
+        # P(T > t) = 1/2 − t / (2·√(2 + t²)) and the two-sided p is
+        #   2·P(T > |t|) = 1 − |t| / √(2 + t²).
+        # (The previous implementation returned twice this — it required
+        # |t| > 6.21 for p<0.05 instead of the correct |t| > 4.30, making
+        # the n=4 classifier far too conservative. See test_classifier.)
+        return max(0.0, 1.0 - t_abs / math.sqrt(2.0 + t_abs * t_abs))
     # Wilson-Hilferty: transform t² ~ F(1, df), then F → chi-square via
     # cube-root approximation. For our purposes the simpler normal-approx
     # to the t with the Hill / Abramowitz adjustment is enough.

{loopgain-0.4.1 → loopgain-0.4.2}/loopgain/telemetry.py

@@ -178,6 +178,11 @@ def build_payload(
             "savings_vs_fixed_cap": result.savings_vs_fixed_cap,
             "convergence_profile_summary": profile_summary,
             "rollback_triggered": result.outcome in ("oscillating", "diverged"),
+            # Index (0-based) of the lowest-error iteration. Lets the receiver
+            # derive iterations-to-best (best_index+1) and iterations-past-best
+            # (iterations_used-1-best_index) — the "Iteration Waste" view.
+            # Privacy-safe: an integer position, no output/error content.
+            "best_index": result.best_index,
             # v2: first computable eta snapshot, for ETA calibration dashboard.
             # Predicted total iterations = first_eta_at_iteration +
             # first_eta_prediction; compare to iterations_used to compute the

{loopgain-0.4.1 → loopgain-0.4.2}/loopgain.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: loopgain
-Version: 0.4.1
+Version: 0.4.2
 Summary: An open-source cost controller for AI agent loops. Stops a loop when it has actually converged and rolls back before it degrades — replacing the max_iterations guess with a real-time loop-gain (Aβ) monitor with five named threshold bands and best-so-far rollback.
 Author-email: Dave Fitzsimmons <hello@loopgain.ai>
 License: Apache-2.0
@@ -58,7 +58,7 @@ AI agent loops waste time and money when they don't know when to stop. LoopGain
 [![PyPI](https://img.shields.io/pypi/v/loopgain.svg)](https://pypi.org/project/loopgain/)
 [![Python](https://img.shields.io/pypi/pyversions/loopgain.svg)](https://pypi.org/project/loopgain/)
 [![License](https://img.shields.io/badge/license-Apache_2.0-blue.svg)](LICENSE)
-[![Tests](https://img.shields.io/badge/tests-202_passing-brightgreen.svg)](tests/)
+[![Tests](https://img.shields.io/badge/tests-200%2B_passing-brightgreen.svg)](tests/)
 
 **Home:** [loopgain.ai](https://loopgain.ai)
 

{loopgain-0.4.1 → loopgain-0.4.2}/tests/test_classifier_mock_validation.py

@@ -223,11 +223,13 @@ def test_loop_length_robustness():
     - n=8 (df=6): ≥ 90% (the default real-loop length)
     - n=12 (df=10): ≥ 95%
     """
-    # n=4 is intentionally excluded: with df=2 the t-test requires |t|>4.3
-    # for p<0.05, which is a fundamental statistical-power floor. The
-    # classifier correctly falls back to STALLING (insufficient evidence)
-    # for most convergent trajectories at n=4. Documented as a
-    # min-recommended-iterations limit, not a bug.
+    # n=4 is intentionally excluded from the high-accuracy thresholds below:
+    # with df=2 the t-test correctly requires |t|>4.30 for p<0.05 (see
+    # test_two_sided_t_p_df2_exact), a fundamental statistical-power floor at
+    # this length. The classifier falls back to cumulative E_ratio when the
+    # slope test is underpowered. This is a min-recommended-iterations limit,
+    # not a bug. (Historically the df=2 p-value was computed at 2x its true
+    # value, requiring |t|>6.21 and worsening this floor — now fixed.)
     LEN_THRESHOLDS = {6: 0.80, 8: 0.90, 12: 0.95}
     for n, threshold in LEN_THRESHOLDS.items():
         for gen, expected in [

{loopgain-0.4.1 → loopgain-0.4.2}/tests/test_classifier_synthetic.py

@@ -33,7 +33,42 @@ from loopgain import (
     classify_trajectory,
     extract_features,
 )
-from loopgain.classifier import _ols_slope_and_p
+from loopgain.classifier import _ols_slope_and_p, _two_sided_t_p
+
+
+# ----- Two-sided t p-value closed forms -----
+
+
+def test_two_sided_t_p_df1_exact():
+    """df=1 is the Cauchy distribution: two-sided p = 1 - 2·atan(t)/pi."""
+    for t in (0.0, 0.5, 1.0, 2.0, 5.0, 12.706):
+        expected = 1.0 - 2.0 * math.atan(t) / math.pi
+        assert _two_sided_t_p(t, 1) == pytest.approx(expected, abs=1e-9)
+    # t=1 is the median of |T| for df=1 → two-sided p = 0.5.
+    assert _two_sided_t_p(1.0, 1) == pytest.approx(0.5, abs=1e-9)
+
+
+def test_two_sided_t_p_df2_exact():
+    """df=2 closed form: two-sided p = 1 - |t|/sqrt(2 + t^2).
+
+    Regression guard for the doubled-p bug: the critical value for p=0.05
+    at df=2 is t=4.302653. The previous implementation returned ~0.10 here
+    (2x too large), which forced |t|>6.21 for significance and made the n=4
+    classifier far too conservative.
+    """
+    for t in (0.0, 0.5, 1.0, 2.0, 5.0):
+        expected = 1.0 - t / math.sqrt(2.0 + t * t)
+        assert _two_sided_t_p(t, 2) == pytest.approx(expected, abs=1e-9)
+    # The exact 5% two-sided critical value for df=2.
+    assert _two_sided_t_p(4.302653, 2) == pytest.approx(0.05, abs=1e-4)
+    # p is a probability: monotone non-increasing in t, bounded to [0, 1].
+    assert _two_sided_t_p(0.0, 2) == pytest.approx(1.0, abs=1e-9)
+    prev = 1.1
+    for t in (0.0, 0.5, 1.0, 2.0, 4.0, 8.0, 50.0):
+        p = _two_sided_t_p(t, 2)
+        assert 0.0 <= p <= 1.0
+        assert p <= prev + 1e-12
+        prev = p
 
 
 # ----- OLS slope / p-value building blocks -----