Active Protocol Discovery (APD)

A lightweight framework for detecting hidden adaptive structure in noisy environments via sequential, KL-optimal probing.

Most detection systems are passive — they watch and wait. APD is active — it probes, measures, adapts, and decides. This makes it fundamentally faster at detecting hidden adaptive behaviour.

Rather than passively monitoring a system, APD:

1. Injects structured perturbations into the target environment
2. Measures statistical deviation between expected and observed responses
3. Updates posterior belief via sequential likelihood ratios
4. Selects the next probe to maximise expected evidence gain (KL-optimal)
5. Decides when statistical confidence is reached — no fixed sample size required

Why APD Over Existing Approaches?

Approach	Method	Weakness	APD Advantage
Passive monitoring	Watch logs/traffic for anomalies	Fails against agents that stay quiet until activated	APD forces adaptive agents to reveal themselves
Fixed probing (honeypots, canaries)	Deploy known traps	Predictable — sophisticated agents learn to avoid them	APD adapts its probes based on what it learns
Threshold detectors	Alert on metric > threshold	Requires pre-set thresholds, high false positive rate	APD uses sequential testing — mathematically controlled error rates
Batch hypothesis testing	Collect N samples, then decide	Wastes samples, fixed sample size regardless of signal strength	APD stops early when evidence is sufficient — up to 45× faster
Anomaly detection (ML)	Train model on "normal", flag outliers	Needs training data, can't detect novel adaptive behaviour	APD is model-free — works with any response distribution

Key Strengths

• 45× faster than passive monitoring — empirically demonstrated (see demo results below)
• 2× faster than best fixed-probe methods — KL-optimal selection always finds the most informative probe
• Mathematically guaranteed error rates — false positive and false negative rates are configurable via Wald SPRT, not heuristic
• No training data required — works from first observation, no labelled dataset needed
• Probe-agnostic — plug in any environment, any probe set, any response distribution
• Lightweight — pure NumPy/SciPy, no deep learning frameworks, runs on a Raspberry Pi

Applications

AI Safety & Adversarial Detection

• Deceptive AI agents — detect if an autonomous agent is strategically hiding capabilities or intentions
• LLM prompt injection — inject structured perturbations, monitor output entropy shifts to detect hidden tool-use loops
• Sleeper agents — probe for trigger-activated behaviour in fine-tuned models
• Multi-agent deception — detect if agents in a sandbox are colluding or behaving adaptively

Network Security

• Bot detection — send structured timing probes, observe latency distribution shifts to distinguish bots from humans
• Adversarial infrastructure — detect command-and-control servers that adapt responses based on probe patterns
• DDoS source identification — probe suspect sources with varied request patterns to detect coordinated adaptive behaviour

Distributed Systems

• Hidden node discovery — detect undeclared adaptive nodes in a distributed network
• Byzantine fault detection — identify nodes that respond strategically rather than honestly
• Protocol compliance testing — probe services to detect non-standard adaptive behaviour

Research

• Multi-agent systems — test whether agents develop emergent adaptive strategies
• Reinforcement learning — detect if an RL agent has learned to game its environment
• Cognitive science — sequential optimal experiment design for detecting adaptive structure in human/animal behaviour

Installation

pip install active-protocol-discovery

Or from source:

git clone https://github.com/DarrenEdwards111/active-protocol-discovery.git
cd active-protocol-discovery
pip install -e ".[dev]"

Or clone and install locally:

git clone https://github.com/DarrenEdwards111/active-protocol-discovery.git
cd active-protocol-discovery
pip install -e ".[dev]"

Quick Start

from apd import APDEngine, GaussianWorld, KLOptimalPolicy, WaldSPRT

world = GaussianWorld(sigma=1.0, adaptive=True)
policy = KLOptimalPolicy(probes=[0.2, 0.5, 1.0])
sprt = WaldSPRT(alpha=0.01, beta=0.01)

engine = APDEngine(world, policy, sprt)
result = engine.run(max_steps=1000)

print(f"Decision: {'Adaptive' if result.decision else 'Null'}")
print(f"Steps: {result.steps}")

How It Works

APD implements a sequential likelihood ratio framework (Wald SPRT) with probe selection based on KL-separability. At each step the engine:

1. Selects a probe — the perturbation that maximises expected KL divergence between H0 and H1
2. Observes the response — samples from the true environment
3. Updates the log-likelihood ratio — accumulates evidence for/against adaptive structure
4. Decides or continues — stops when the SPRT threshold is crossed

Demo Results

Results from experiments/gaussian_demo.py (1000 trials, σ=1.0, α=β=0.01):

H1: Adaptive world (adversary present)

Method	Mean Steps	Detection Rate	Undecided
Passive (u=0)	10000.0	0.000	1.000
Weak beacon (u=0.2)	231.5	0.987	0.000
Strong beacon (u=1.0)	10.4	0.996	0.000
APD KL-optimal {0.2, 1.0}	10.4	0.996	0.000
APD KL-optimal {0.1..1.5}	5.1	0.993	0.000

H0: Null world (false positive check)

Method	Mean Steps	False Positive Rate	Correct Null
Strong beacon (u=1.0)	10.5	0.003	0.997
APD KL-optimal {0.1..1.5}	5.1	0.002	0.998

Key takeaway: APD automatically selects the strongest available probe (u=1.5), achieving 2× faster detection than a strong fixed beacon while maintaining false positive rates below α.

Theory

Wald SPRT

The Sequential Probability Ratio Test accumulates log-likelihood ratios until crossing a decision boundary:

• Log-likelihood ratio: Λ(y|u) = log p(y|H₁,u) − log p(y|H₀)
• Upper threshold (accept H₁): A = log((1−β)/α)
• Lower threshold (accept H₀): B = log(β/(1−α))

KL Divergence as Probe Quality

For the Gaussian case, the KL divergence between H₁ and H₀ given probe u is:

D(u) = μ(u)² / (2σ²)

This quantifies how much information a single observation provides about the hypothesis. APD selects the probe maximising D(u), achieving the fastest possible sequential detection.

Expected Sample Complexity

Under H₁, Wald's approximation gives:

E[N|H₁] ≈ ((1−β)·A + β·B) / D(u)

Stronger probes (higher D(u)) yield fewer required samples.

Package Structure

apd/
├── models.py    — World models (Gaussian, AdaptiveAgent, Network)
├── policy.py    — Probe selection (KL-optimal, fixed, random, ε-greedy)
├── sprt.py      — Wald Sequential Probability Ratio Test
├── apd.py       — Main APD engine
└── utils.py     — Statistics helpers

Running Experiments

# Gaussian demo (the killer demo)
python -m experiments.gaussian_demo

# Adversarial agent in multi-dimensional space
python -m experiments.adversarial_agent_demo

# Full benchmark sweep (outputs CSV)
python -m experiments.benchmark > results.csv

Tests

pytest tests/ -v

Citation

@software{apd2026,
  author = {Mikoshi Ltd},
  title = {Active Protocol Discovery},
  year = {2026},
  url = {https://github.com/DarrenEdwards111/active-protocol-discovery}
}

Licence

Apache 2.0 — Mikoshi Ltd, 2026