loopgain 0.3.0__tar.gz → 0.4.0__tar.gz

{loopgain-0.3.0 → loopgain-0.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: loopgain
-Version: 0.3.0
+Version: 0.4.0
 Summary: Barkhausen stability monitor for AI agent loops. Real-time loop-gain (Aβ) monitoring with five named threshold bands, best-so-far rollback, and ETA prediction.
 Author-email: Dave Fitzsimmons <hello@loopgain.ai>
 License: Apache-2.0
@@ -108,6 +108,28 @@ print(result.savings_vs_fixed_cap)
 
 ---
 
+## Defining your error signal
+
+The one thing you provide is the **error signal**: a single non-negative number, every iteration, that says how wrong the current output is. **Lower is better; zero means done.** LoopGain doesn't know what your loop does — it just watches that number's trajectory and decides whether to keep going, stop, or roll back.
+
+Your loop already has some way of knowing the output isn't good yet (or it wouldn't keep revising). Turn that into a number:
+
+| Loop | Error signal = |
+| --- | --- |
+| Agentic coding (write code → run tests) | number of **failing tests** (10 → 3 → 0) |
+| JSON / structured extraction | number of **schema violations** |
+| RAG with self-correction | number of **required facts still missing** |
+| Self-refinement with an LLM judge | judge's **gap to target** (e.g. `10 − quality_score`) |
+| Lint / format loop | **lint error count** |
+
+The only rules: non-negative, and **smaller as the output gets better**. Returning the raw list of problems works directly — `observe()` uses its length as the magnitude (e.g. hand it the list of failing tests).
+
+If your quality is fuzzy and has no natural "zero," run with `target_error=None`: LoopGain then stops when the number **stops improving**, wherever that plateau is, instead of waiting for an exact target.
+
+Every stop/continue decision is made from this one number, so **LoopGain is only as good as the error signal you give it** — pick one that genuinely tracks output quality.
+
+---
+
 ## How it works
 
 LoopGain measures empirical loop gain (`Aβ = E(n) / E(n-1)`) at every iteration and exposes it as a smoothed time series for visualization. The decision engine, however, classifies the **full error trajectory** using four features:
@@ -165,14 +187,23 @@ This transforms divergence detection from "abort with garbage" into "abort with
 
 ---
 
+## What LoopGain does and doesn't guarantee
+
+LoopGain saves money by stopping a loop once it stops improving — fewer iterations, fewer tokens. In our [public benchmark](https://github.com/loopgain-ai/loopgain-bench), that was a **93.5% median cut in API spend** vs `max_iterations=20`, with output quality preserved. Two honest limits:
+
+- **Savings depend on your workload.** Loops that usually succeed fast save the most (~96%); adversarial, failure-prone loops save less (~84%). The headline is a blend — run the benchmark on your own loops before quoting a number.
+- **LoopGain detects convergence, not correctness.** It stops when your error signal stops improving — which means more iterations won't help, *not* that the loop succeeded. On the benchmark this preserved quality (it rarely stopped early on a worse output; false-stop rate ≤3.5%), but a loop can stall with the error still above zero — a plateau at, say, 2 failing tests. So check `result.best_error` (or your own pass/fail) before you trust the output: if it plateaued short of your target, that's a quality gap LoopGain can't see, and a false stop that forces a rerun is the one way it eats into the savings. LoopGain decides *when to stop*; you decide *whether the answer is good enough*.
+
+---
+
 ## API reference
 
-### `LoopGain(target_error=0.0, max_iterations=None, thresholds=None, trajectory_thresholds=None, classifier='trajectory', smoothing_window=3, assumed_fixed_cap=10)`
+### `LoopGain(target_error=0.0, max_iterations=50, thresholds=None, trajectory_thresholds=None, classifier='trajectory', smoothing_window=3, assumed_fixed_cap=10)`
 
 Construct the monitor.
 
 - `target_error` — Stop when an observed error drops at or below this. Default `0.0` short-circuits on exactly zero error (the natural completion signal for verifier-driven loops). Pass `None` to disable the short-circuit entirely.
-- `max_iterations` — Hard safety cap. Default `None` (rely on stability detection). Recommended ~20–50 for production.
+- `max_iterations` — Hard safety backstop. Default `50` so the loop can never run unbounded; a stability verdict normally terminates it well before this. Pass `None` to opt into a fully unbounded loop (only safe if your loop is guaranteed to reach `target_error` or a stop-state), or a smaller integer to cap tighter.
 - `thresholds` — Custom `ThresholdBands` for the legacy single-Aβ-band classifier. Ignored when `classifier='trajectory'`.
 - `trajectory_thresholds` — Custom `TrajectoryThresholds` for the multi-feature classifier (the default). Override only with workload-specific evidence.
 - `classifier` — `'trajectory'` (default, v0.2 multi-feature classifier) or `'legacy_bands'` (v0.1 single-Aβ-band classifier).

{loopgain-0.3.0 → loopgain-0.4.0}/README.md

@@ -59,6 +59,28 @@ print(result.savings_vs_fixed_cap)
 
 ---
 
+## Defining your error signal
+
+The one thing you provide is the **error signal**: a single non-negative number, every iteration, that says how wrong the current output is. **Lower is better; zero means done.** LoopGain doesn't know what your loop does — it just watches that number's trajectory and decides whether to keep going, stop, or roll back.
+
+Your loop already has some way of knowing the output isn't good yet (or it wouldn't keep revising). Turn that into a number:
+
+| Loop | Error signal = |
+| --- | --- |
+| Agentic coding (write code → run tests) | number of **failing tests** (10 → 3 → 0) |
+| JSON / structured extraction | number of **schema violations** |
+| RAG with self-correction | number of **required facts still missing** |
+| Self-refinement with an LLM judge | judge's **gap to target** (e.g. `10 − quality_score`) |
+| Lint / format loop | **lint error count** |
+
+The only rules: non-negative, and **smaller as the output gets better**. Returning the raw list of problems works directly — `observe()` uses its length as the magnitude (e.g. hand it the list of failing tests).
+
+If your quality is fuzzy and has no natural "zero," run with `target_error=None`: LoopGain then stops when the number **stops improving**, wherever that plateau is, instead of waiting for an exact target.
+
+Every stop/continue decision is made from this one number, so **LoopGain is only as good as the error signal you give it** — pick one that genuinely tracks output quality.
+
+---
+
 ## How it works
 
 LoopGain measures empirical loop gain (`Aβ = E(n) / E(n-1)`) at every iteration and exposes it as a smoothed time series for visualization. The decision engine, however, classifies the **full error trajectory** using four features:
@@ -116,14 +138,23 @@ This transforms divergence detection from "abort with garbage" into "abort with
 
 ---
 
+## What LoopGain does and doesn't guarantee
+
+LoopGain saves money by stopping a loop once it stops improving — fewer iterations, fewer tokens. In our [public benchmark](https://github.com/loopgain-ai/loopgain-bench), that was a **93.5% median cut in API spend** vs `max_iterations=20`, with output quality preserved. Two honest limits:
+
+- **Savings depend on your workload.** Loops that usually succeed fast save the most (~96%); adversarial, failure-prone loops save less (~84%). The headline is a blend — run the benchmark on your own loops before quoting a number.
+- **LoopGain detects convergence, not correctness.** It stops when your error signal stops improving — which means more iterations won't help, *not* that the loop succeeded. On the benchmark this preserved quality (it rarely stopped early on a worse output; false-stop rate ≤3.5%), but a loop can stall with the error still above zero — a plateau at, say, 2 failing tests. So check `result.best_error` (or your own pass/fail) before you trust the output: if it plateaued short of your target, that's a quality gap LoopGain can't see, and a false stop that forces a rerun is the one way it eats into the savings. LoopGain decides *when to stop*; you decide *whether the answer is good enough*.
+
+---
+
 ## API reference
 
-### `LoopGain(target_error=0.0, max_iterations=None, thresholds=None, trajectory_thresholds=None, classifier='trajectory', smoothing_window=3, assumed_fixed_cap=10)`
+### `LoopGain(target_error=0.0, max_iterations=50, thresholds=None, trajectory_thresholds=None, classifier='trajectory', smoothing_window=3, assumed_fixed_cap=10)`
 
 Construct the monitor.
 
 - `target_error` — Stop when an observed error drops at or below this. Default `0.0` short-circuits on exactly zero error (the natural completion signal for verifier-driven loops). Pass `None` to disable the short-circuit entirely.
-- `max_iterations` — Hard safety cap. Default `None` (rely on stability detection). Recommended ~20–50 for production.
+- `max_iterations` — Hard safety backstop. Default `50` so the loop can never run unbounded; a stability verdict normally terminates it well before this. Pass `None` to opt into a fully unbounded loop (only safe if your loop is guaranteed to reach `target_error` or a stop-state), or a smaller integer to cap tighter.
 - `thresholds` — Custom `ThresholdBands` for the legacy single-Aβ-band classifier. Ignored when `classifier='trajectory'`.
 - `trajectory_thresholds` — Custom `TrajectoryThresholds` for the multi-feature classifier (the default). Override only with workload-specific evidence.
 - `classifier` — `'trajectory'` (default, v0.2 multi-feature classifier) or `'legacy_bands'` (v0.1 single-Aβ-band classifier).

{loopgain-0.3.0 → loopgain-0.4.0}/loopgain/_version.py

@@ -7,4 +7,4 @@ from here so the value never drifts between ``__version__`` and the
 ``pyproject.toml``) for each release.
 """
 
-__version__ = "0.3.0"
+__version__ = "0.4.0"

{loopgain-0.3.0 → loopgain-0.4.0}/loopgain/classifier.py

@@ -66,6 +66,20 @@ DEFAULT_OSC_STD_THRESHOLD = 0.30
 # for the oscillation gate.
 DEFAULT_SLOPE_TOL = 0.05
 
+# Liveness gate: number of iterations a loop may go without achieving a new
+# best (lowest) error before its "continue" verdicts (FAST_CONVERGE /
+# CONVERGING) are withdrawn so it can reach STALLING / OSCILLATING and
+# terminate. Without this, a loop that drops a lot and then plateaus or
+# oscillates *below* the cumulative thresholds keeps its historical win
+# forever and never terminates. Derivation: the continue-states are claims
+# about *ongoing* progress; cumulative reduction (E_current/E_first) and a
+# whole-history slope are claims about the *past* and do not expire. We treat
+# "no new low in N steps" as the loop having stopped improving. N is small
+# (3) so a sustained plateau is caught quickly, but the consecutive-STALLING
+# termination rule (2 readings) still protects a loop that briefly stalls and
+# then resumes hitting new lows.
+DEFAULT_STALL_PATIENCE = 3
+
 # Numerical floor to avoid log(0).
 _EPS = 1e-12
 
@@ -85,6 +99,7 @@ class TrajectoryThresholds:
     div_margin: float = DEFAULT_DIV_MARGIN
     osc_std_threshold: float = DEFAULT_OSC_STD_THRESHOLD
     slope_tol: float = DEFAULT_SLOPE_TOL
+    stall_patience: int = DEFAULT_STALL_PATIENCE
 
 
 @dataclass(frozen=True)
@@ -276,6 +291,18 @@ def classify_trajectory(
 
     f = extract_features(error_history)
 
+    # Liveness signal: how many iterations since the loop last achieved a new
+    # best (lowest) error. A genuinely converging loop keeps hitting new lows,
+    # so this stays small; a loop that dropped a lot and then plateaued (or is
+    # oscillating below the cumulative thresholds) has a large value. We use it
+    # to withdraw the "continue" verdicts (FAST_CONVERGE / CONVERGING) once a
+    # loop has stopped improving, so it can reach STALLING / OSCILLATING and
+    # terminate instead of riding its historical cumulative win forever. See
+    # DEFAULT_STALL_PATIENCE.
+    hist = list(error_history)
+    iters_since_best = (n - 1) - hist.index(min(hist))
+    still_improving = iters_since_best < th.stall_patience
+
     # n == 2 special case: with two observations, the slope is well defined
     # but its p-value is not (zero residual degrees of freedom). Fall back to
     # the sign of the change. This is the same conservatism as a Wilcoxon
@@ -291,13 +318,20 @@ def classify_trajectory(
         return STALLING
 
     # Order matters: FAST_CONVERGE precedes CONVERGING; both precede the
-    # remaining gates.
-    if f.e_ratio <= th.e_ratio_fast:
+    # remaining gates. Both continue-verdicts are gated on `still_improving`:
+    # a loop that has stopped hitting new lows is no longer "converging" no
+    # matter how large its historical cumulative reduction was, and must be
+    # allowed to fall through to STALLING / OSCILLATING so it can terminate.
+    if f.e_ratio <= th.e_ratio_fast and still_improving:
         return FAST_CONVERGE
 
     slope_significant = f.slope_p < th.p_sig
 
-    if f.slope_log < 0 and (slope_significant or f.e_ratio <= th.e_ratio_conv):
+    if (
+        f.slope_log < 0
+        and still_improving
+        and (slope_significant or f.e_ratio <= th.e_ratio_conv)
+    ):
         return CONVERGING
 
     if f.slope_log > 0 and slope_significant and f.e_ratio > 1.0 + th.div_margin:

{loopgain-0.3.0 → loopgain-0.4.0}/loopgain/core.py

@@ -40,6 +40,16 @@ DEFAULT_STALLING = 0.95
 DEFAULT_OSCILLATING_UPPER = 1.05
 
 
+# Bounded-by-default safety backstop. The loop should normally terminate on a
+# stability verdict (target met / oscillating / diverging / stalled) long
+# before this; it exists only so the library can never run truly unbounded if
+# a loop never converges and never stalls (e.g. infinitesimal-but-real progress
+# with target_error=None). Generous relative to typical loop lengths (the
+# bench capped at 20). Pass max_iterations=None to opt into a fully unbounded
+# loop, or a smaller integer to cap tighter.
+DEFAULT_MAX_ITERATIONS = 50
+
+
 # State names. Exported for use in switch/case in user code.
 INIT = "INIT"
 FAST_CONVERGE = "FAST_CONVERGE"
@@ -165,8 +175,11 @@ class LoopGain:
             tests, no validation errors, etc.). Pass ``None`` to disable
             the short-circuit entirely and rely only on stability
             detection and ``max_iterations``.
-        max_iterations: Hard safety cap. Default ``None`` (rely on
-            stability detection). Recommended ~20-50 for production.
+        max_iterations: Hard safety backstop. Default
+            ``DEFAULT_MAX_ITERATIONS`` (50) so the loop can never run
+            unbounded; normally a stability verdict terminates it long
+            before this. Pass ``None`` to opt into a fully unbounded loop,
+            or a smaller integer to cap tighter.
         thresholds: Custom ``ThresholdBands`` (legacy single-feature
             classifier only). Default is the canonical 0.3 / 0.85 / 0.95 /
             1.05. Ignored when ``classifier='trajectory'``.
@@ -190,7 +203,7 @@ class LoopGain:
     def __init__(
         self,
         target_error: Optional[float] = 0.0,
-        max_iterations: Optional[int] = None,
+        max_iterations: Optional[int] = DEFAULT_MAX_ITERATIONS,
         thresholds: Optional[ThresholdBands] = None,
         trajectory_thresholds: Optional[TrajectoryThresholds] = None,
         classifier: str = "trajectory",

{loopgain-0.3.0 → loopgain-0.4.0}/loopgain.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: loopgain
-Version: 0.3.0
+Version: 0.4.0
 Summary: Barkhausen stability monitor for AI agent loops. Real-time loop-gain (Aβ) monitoring with five named threshold bands, best-so-far rollback, and ETA prediction.
 Author-email: Dave Fitzsimmons <hello@loopgain.ai>
 License: Apache-2.0
@@ -108,6 +108,28 @@ print(result.savings_vs_fixed_cap)
 
 ---
 
+## Defining your error signal
+
+The one thing you provide is the **error signal**: a single non-negative number, every iteration, that says how wrong the current output is. **Lower is better; zero means done.** LoopGain doesn't know what your loop does — it just watches that number's trajectory and decides whether to keep going, stop, or roll back.
+
+Your loop already has some way of knowing the output isn't good yet (or it wouldn't keep revising). Turn that into a number:
+
+| Loop | Error signal = |
+| --- | --- |
+| Agentic coding (write code → run tests) | number of **failing tests** (10 → 3 → 0) |
+| JSON / structured extraction | number of **schema violations** |
+| RAG with self-correction | number of **required facts still missing** |
+| Self-refinement with an LLM judge | judge's **gap to target** (e.g. `10 − quality_score`) |
+| Lint / format loop | **lint error count** |
+
+The only rules: non-negative, and **smaller as the output gets better**. Returning the raw list of problems works directly — `observe()` uses its length as the magnitude (e.g. hand it the list of failing tests).
+
+If your quality is fuzzy and has no natural "zero," run with `target_error=None`: LoopGain then stops when the number **stops improving**, wherever that plateau is, instead of waiting for an exact target.
+
+Every stop/continue decision is made from this one number, so **LoopGain is only as good as the error signal you give it** — pick one that genuinely tracks output quality.
+
+---
+
 ## How it works
 
 LoopGain measures empirical loop gain (`Aβ = E(n) / E(n-1)`) at every iteration and exposes it as a smoothed time series for visualization. The decision engine, however, classifies the **full error trajectory** using four features:
@@ -165,14 +187,23 @@ This transforms divergence detection from "abort with garbage" into "abort with
 
 ---
 
+## What LoopGain does and doesn't guarantee
+
+LoopGain saves money by stopping a loop once it stops improving — fewer iterations, fewer tokens. In our [public benchmark](https://github.com/loopgain-ai/loopgain-bench), that was a **93.5% median cut in API spend** vs `max_iterations=20`, with output quality preserved. Two honest limits:
+
+- **Savings depend on your workload.** Loops that usually succeed fast save the most (~96%); adversarial, failure-prone loops save less (~84%). The headline is a blend — run the benchmark on your own loops before quoting a number.
+- **LoopGain detects convergence, not correctness.** It stops when your error signal stops improving — which means more iterations won't help, *not* that the loop succeeded. On the benchmark this preserved quality (it rarely stopped early on a worse output; false-stop rate ≤3.5%), but a loop can stall with the error still above zero — a plateau at, say, 2 failing tests. So check `result.best_error` (or your own pass/fail) before you trust the output: if it plateaued short of your target, that's a quality gap LoopGain can't see, and a false stop that forces a rerun is the one way it eats into the savings. LoopGain decides *when to stop*; you decide *whether the answer is good enough*.
+
+---
+
 ## API reference
 
-### `LoopGain(target_error=0.0, max_iterations=None, thresholds=None, trajectory_thresholds=None, classifier='trajectory', smoothing_window=3, assumed_fixed_cap=10)`
+### `LoopGain(target_error=0.0, max_iterations=50, thresholds=None, trajectory_thresholds=None, classifier='trajectory', smoothing_window=3, assumed_fixed_cap=10)`
 
 Construct the monitor.
 
 - `target_error` — Stop when an observed error drops at or below this. Default `0.0` short-circuits on exactly zero error (the natural completion signal for verifier-driven loops). Pass `None` to disable the short-circuit entirely.
-- `max_iterations` — Hard safety cap. Default `None` (rely on stability detection). Recommended ~20–50 for production.
+- `max_iterations` — Hard safety backstop. Default `50` so the loop can never run unbounded; a stability verdict normally terminates it well before this. Pass `None` to opt into a fully unbounded loop (only safe if your loop is guaranteed to reach `target_error` or a stop-state), or a smaller integer to cap tighter.
 - `thresholds` — Custom `ThresholdBands` for the legacy single-Aβ-band classifier. Ignored when `classifier='trajectory'`.
 - `trajectory_thresholds` — Custom `TrajectoryThresholds` for the multi-feature classifier (the default). Override only with workload-specific evidence.
 - `classifier` — `'trajectory'` (default, v0.2 multi-feature classifier) or `'legacy_bands'` (v0.1 single-Aβ-band classifier).

{loopgain-0.3.0 → loopgain-0.4.0}/loopgain.egg-info/SOURCES.txt

@@ -28,4 +28,5 @@ tests/test_core.py
 tests/test_funnel.py
 tests/test_integrations.py
 tests/test_stress.py
-tests/test_telemetry.py
+tests/test_telemetry.py
+tests/test_termination_safety.py

{loopgain-0.3.0 → loopgain-0.4.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "loopgain"
-version = "0.3.0"
+version = "0.4.0"
 description = "Barkhausen stability monitor for AI agent loops. Real-time loop-gain (Aβ) monitoring with five named threshold bands, best-so-far rollback, and ETA prediction."
 authors = [{name = "Dave Fitzsimmons", email = "hello@loopgain.ai"}]
 readme = "README.md"

{loopgain-0.3.0 → loopgain-0.4.0}/tests/test_classifier_synthetic.py

@@ -158,12 +158,22 @@ def test_pure_stall_no_trend():
     )
 
 
-def test_floor_convergence_already_at_target():
-    """If error is already ≈ 0 at observation 1, classifier returns
-    FAST_CONVERGE (cumulative reduction to floor)."""
+def test_floor_convergence_already_flat_at_floor_stalls():
+    """A loop already pinned at the numerical floor from iteration 0, flat,
+    classifies as STALLING — not FAST_CONVERGE.
+
+    Updated 2026-06 with the liveness-gate fix (see DEFAULT_STALL_PATIENCE).
+    Previously this returned FAST_CONVERGE on the strength of cumulative
+    reduction alone — but FAST_CONVERGE is a *continue* verdict, so an
+    at-floor flat loop would have continued (and, with no max_iterations,
+    run unbounded) instead of stopping. STALLING is the correct verdict: the
+    loop has made no progress for `stall_patience` iterations, so it
+    terminates via the consecutive-stall rule and returns best-so-far (the
+    floor value — a fine answer). In real use the `target_error`
+    short-circuit (next test) handles the at-target case directly."""
     trajectory = [1e-15] * 5
     state = classify_trajectory(trajectory)
-    assert state == FAST_CONVERGE
+    assert state == STALLING
 
 
 def test_target_met_short_circuit():

loopgain-0.4.0/tests/test_termination_safety.py

@@ -0,0 +1,115 @@
+"""Termination-safety tests: a loop must not run unbounded.
+
+Regression coverage for the FAST_CONVERGE/CONVERGING liveness bug (2026-06):
+the trajectory classifier used *cumulative* reduction (E_current/E_first) and a
+*whole-history* slope to emit the "continue" verdicts FAST_CONVERGE and
+CONVERGING. A loop that reduced its error and then plateaued (or oscillated)
+*below* the cumulative thresholds kept its historical win forever — it was
+pinned in a continue-state, never reached STALLING/OSCILLATING, and with the
+(then-default) max_iterations=None it ran forever.
+
+The fix has two independent layers, each tested here:
+  1. A liveness gate on the continue-verdicts: a loop that has not achieved a
+     new best error in `stall_patience` iterations is no longer treated as
+     "improving", so it can reach STALLING/OSCILLATING and terminate.
+  2. A bounded default max_iterations backstop, so the library can never run
+     truly unbounded even if a future classifier path regresses.
+
+Output quality was never at risk (best-so-far rollback held the good answer);
+the bug was a *liveness* failure — the loop never returned to hand it back.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from loopgain import CONVERGING, FAST_CONVERGE, LoopGain, classify_trajectory
+
+# Hard test guard: large enough that a *correctly* terminating loop never hits
+# it, small enough that a regression (unbounded loop) fails fast instead of
+# hanging the suite.
+GUARD = 500
+
+
+def _run_to_termination(lg: LoopGain, errors, guard: int = GUARD):
+    """Drive a loop, plateauing/repeating the last error, until it terminates
+    or hits the guard. Returns (iterations_run, hit_guard)."""
+    i = 0
+    while lg.should_continue():
+        e = errors[i] if i < len(errors) else errors[-1]
+        lg.observe(e, output=f"o{i}")
+        i += 1
+        if i >= guard:
+            return i, True
+    return i, False
+
+
+# ----- Layer 1: classifier liveness gate -----
+
+
+def test_plateau_below_fast_floor_terminates_without_max_iter():
+    """Error drops to 8% of initial then plateaus. e_ratio<=0.1 used to pin
+    FAST_CONVERGE forever. Must now terminate via STALLING."""
+    lg = LoopGain(max_iterations=None, target_error=None)
+    n, hit_guard = _run_to_termination(lg, [100, 8, 8, 8, 8, 8, 8, 8])
+    assert not hit_guard, f"loop did not terminate within {GUARD} iters (unbounded)"
+    assert not lg.should_continue()
+    assert lg.result.best_error == 8.0  # best-so-far still returned
+
+
+def test_plateau_above_fast_floor_terminates_without_max_iter():
+    """Error drops to 30% of initial (below E_RATIO_CONV=0.5) then plateaus.
+    e_ratio<=0.5 with a whole-history negative slope used to pin CONVERGING
+    forever. Must now terminate."""
+    lg = LoopGain(max_iterations=None, target_error=None)
+    n, hit_guard = _run_to_termination(lg, [100, 30, 30, 30, 30, 30, 30, 30])
+    assert not hit_guard, f"loop did not terminate within {GUARD} iters (unbounded)"
+    assert not lg.should_continue()
+
+
+def test_oscillation_below_floor_terminates_without_max_iter():
+    """Oscillation entirely below the 10% cumulative floor used to be shadowed
+    by FAST_CONVERGE. Must now terminate (OSCILLATING or STALLING)."""
+    lg = LoopGain(max_iterations=None, target_error=None)
+    n, hit_guard = _run_to_termination(lg, [100, 5, 8, 5, 8, 5, 8, 5, 8])
+    assert not hit_guard, f"loop did not terminate within {GUARD} iters (unbounded)"
+    assert not lg.should_continue()
+
+
+def test_classifier_flags_plateau_after_big_drop_as_terminable():
+    """Direct classifier check: a big drop followed by a flat tail must NOT be
+    reported as a continue-state (FAST_CONVERGE/CONVERGING)."""
+    plateau_low = [100, 8, 8, 8, 8, 8]
+    plateau_mid = [100, 30, 30, 30, 30, 30]
+    assert classify_trajectory(plateau_low) not in (FAST_CONVERGE, CONVERGING)
+    assert classify_trajectory(plateau_mid) not in (FAST_CONVERGE, CONVERGING)
+
+
+def test_genuine_fast_converge_still_continues():
+    """Guard against over-correction: a monotone steep decline that keeps
+    hitting new lows must still read FAST_CONVERGE (continue), not be
+    prematurely stalled."""
+    monotone = [100, 25, 6, 1.5, 0.4, 0.1]  # new low every step
+    assert classify_trajectory(monotone) == FAST_CONVERGE
+
+
+def test_genuine_converging_still_continues():
+    """A steady decline landing between the two cumulative thresholds must
+    still read CONVERGING while it is still hitting new lows."""
+    converging = [10.0, 8.0, 6.4, 5.1, 4.1, 3.3]  # ~0.8x/step, new low every step
+    assert classify_trajectory(converging) == CONVERGING
+
+
+# ----- Layer 2: bounded default backstop -----
+
+
+def test_default_max_iterations_is_a_bounded_backstop():
+    """The default config must not be able to run unbounded. A never-improving
+    loop under all-default construction must terminate at the backstop."""
+    lg = LoopGain()  # all defaults
+    assert lg.max_iterations is not None, "default max_iterations must be bounded"
+    # A strictly increasing error never converges/stalls into best-so-far early
+    # under every classifier path; the backstop must still stop it.
+    i, hit_guard = _run_to_termination(lg, list(range(1, GUARD + 5)))
+    assert not hit_guard, "default backstop failed to bound the loop"
+    assert not lg.should_continue()