Live demo: runtime behavioral monitor¶
A live demo of the library watching the behaviour of an AI program while it runs. Pick one of four real recordings: the red points are the moments the library flagged as strange, and none was chosen by hand.
Four real scenarios below, computed once from dense-armor's actual detectors
(classify_segments + cusum_detector + one_sided_upper_filter, radius=5,
ref_mult=2 — the same call you would make yourself) against the real telemetry of a
Qwen2 1.8B agent (via Ollama), not synthetic noise. See
test/agent_v2/
for how it was generated.
Pick a scenario; the red points are exactly what the library flagged, nothing hand-picked.
What this does not show¶
Whether Dense-Armor catches a security attack, not just a timing anomaly. It does not: a real indirect prompt injection against the same agent succeeded 10/10 times and was flagged 0/10 times by this exact detector stack. See Experiment 40 for that real, honest negative result. Dense-Armor is a runtime behavioral-drift / glitch monitor, not a semantic security layer.
Have an agent in production with this problem?¶
Open a GitHub Discussion — two things worth knowing: does your pipeline have silent drift / glitches today, and what would you actually want a runtime monitor like this to catch that is not shown above?
Reproduce the demo locally¶
```python from dense_armor.utility.protect.arbiter import classify_segments from dense_armor.utility.drift.cusum import cusum_detector from dense_armor.utility.anomaly.one_sided import one_sided_upper_filter