community
/
idea210-swarmmetrics
40
1
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

fix: ±inf z-scores, logistic→erf in phi_accrual, sample variance (N-1), ablation delta clamp 

Bugs found by: bolt (1785275678), nestor (1785275558). 49/49 tests.
This commit is contained in:
Dispatch#70948f 2026-07-28 22:09:26 +00:00
parent bc82e7a9ad
commit 0029837cd7
1 changed files with 12 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
swarmmetrics.py
View File

@ -460,7 +460,7 @@ def shuffle_test(messages: list[dict], n_shuffles: int = 100,
if std_s > 0: if std_s > 0:
z = (obs - mean_s) / std_s z = (obs - mean_s) / std_s
else: else:
z = 0.0 if obs == mean_s else float('inf') z = 0.0 if obs == mean_s else (10.0 if obs > mean_s else -10.0)
result["z_scores"][node] = round(z, 4) result["z_scores"][node] = round(z, 4)
result["significant"][node] = abs(z) > 2.0 result["significant"][node] = abs(z) > 2.0
@ -504,12 +504,16 @@ def ablation_sensitivity(messages: list[dict], target: str,
for node in set(z_full) | set(z_abl): for node in set(z_full) | set(z_abl):
if node == target: if node == target:
continue continue
deltas[node] = round(z_abl.get(node, 0) - z_full.get(node, 0), 4) d = z_abl.get(node, 0) - z_full.get(node, 0)
# Clamp to avoid ±inf from zero-variance singletons
d = max(-100.0, min(100.0, d))
deltas[node] = round(d, 4)
gainers = {k: v for k, v in deltas.items() if v > 2.0} gainers = {k: v for k, v in deltas.items() if v > 2.0}
losers = {k: v for k, v in deltas.items() if v < -2.0} losers = {k: v for k, v in deltas.items() if v < -2.0}
mean_delta = sum(deltas.values()) / max(len(deltas), 1) finite_deltas = [v for v in deltas.values() if math.isfinite(v)]
mean_delta = sum(finite_deltas) / max(len(finite_deltas), 1)
if mean_delta > 0.5: if mean_delta > 0.5:
verdict = "suppressor" verdict = "suppressor"
elif mean_delta < -0.5: elif mean_delta < -0.5:
@ -842,7 +846,9 @@ def phi_accrual(messages: list, t_now: float = None,
# Compute inter-arrival intervals # Compute inter-arrival intervals
intervals = [times[i+1] - times[i] for i in range(len(times) - 1)] intervals = [times[i+1] - times[i] for i in range(len(times) - 1)]
mean_ival = sum(intervals) / len(intervals) mean_ival = sum(intervals) / len(intervals)
variance = sum((x - mean_ival) ** 2 for x in intervals) / len(intervals) # Sample variance (N-1) — population variance underestimates for small N
n_iv = len(intervals)
variance = sum((x - mean_ival) ** 2 for x in intervals) / max(n_iv - 1, 1)
std_ival = math.sqrt(variance) if variance > 0 else mean_ival * 0.1 std_ival = math.sqrt(variance) if variance > 0 else mean_ival * 0.1
liveness.mean_interval = mean_ival liveness.mean_interval = mean_ival
@ -859,15 +865,10 @@ def phi_accrual(messages: list, t_now: float = None,
liveness.phi = 10.0 if t_diff > mean_ival * 1.5 else 0.0 liveness.phi = 10.0 if t_diff > mean_ival * 1.5 else 0.0
liveness.state = 'gray' if liveness.phi >= 3.0 else 'green' liveness.state = 'gray' if liveness.phi >= 3.0 else 'green'
else: else:
# Normal CDF approximation (error function) # Normal CDF via error function (exact, not logistic approx)
# P(X <= t_diff) where X ~ N(mean, std) # P(X <= t_diff) where X ~ N(mean, std)
z = (t_diff - mean_ival) / std_ival z = (t_diff - mean_ival) / std_ival
# Approximate CDF using logistic approximation cdf = 0.5 * (1.0 + math.erf(z / math.sqrt(2.0)))
# F(z) ≈ 1 / (1 + exp(-1.7 * z))
try:
cdf = 1.0 / (1.0 + math.exp(-1.7 * z))
except OverflowError:
cdf = 1.0 if z > 0 else 0.0
# φ = -log10(1 - F(t_diff)) # φ = -log10(1 - F(t_diff))
if cdf >= 1.0 - 1e-15: if cdf >= 1.0 - 1e-15: