Agent & Swarm Config Live Parameters

Agent Swarm Size 120 agents

Posting Velocity (msg/s) 8.5 msg/s

LLM Temperature (Entropy) 0.92

Template Variety 3 archetypes

Spam Defense Filter

Filter Heuristic Mode

Entropy Gate Threshold 3.40

Real-Time Synthetic Feed Telemetry

60 FPS · Active Engine

Feed Ingestion Stream (1020 events)

● Human Safe ▲ Slop Slipped ✖ Slop Blocked

Audit Metrics Benchmark Target

Total Generated 1020

AI Slop Count 894

Human Count 126

Blocked Slop 812

False Positives 14

Feed Purity 91.2%

Noise Ratio 87.6%

Active Mitigation Rules

Shannon Token Entropy < 3.40 bits/tok

Repetitive Slop Templates Jaccard > 0.78

Synthetic Agent Archetypes 3 Flagged

Bandwidth Waste Averted 4.28 MB

Research Context: Ars Technica highlighted how swarms of autonomous agents produce repetitive "slop" that contaminates web discussions and search indices. Tuning entropy gates suppresses autonomous spam without silencing organic authors.

Reading synthetic feed metrics without confusing them with evidence

Your browser does not support video.

Read the explanation

This workbench starts with one thousand twenty synthetic events: eight hundred ninety four labeled slop and one hundred twenty six labeled human. Their types come from the generator, not from a classifier analyzing real posts. Entropy values are randomly assigned within different ranges for the two types. The page never computes token entropy from the message text. Think of it as an experiment about rules on simulated data, rather than a working spam detector. The default dashboard reports eight hundred twelve blocked slop items and fourteen blocked human items. That leaves eighty two slop items and one hundred twelve human items delivered. Human share is one hundred twelve divided by one hundred ninety four, about fifty seven point seven percent. Yet the fixture displays ninety one point two percent. The source explicitly restores stored expected results at default settings. The displayed purity is therefore inconsistent with its own displayed counts, and should not be interpreted as verified filter performance. Away from the default override, recalculation walks the stored posts. None mode blocks nothing. Strict mode blocks slop below threshold plus point six and human posts below threshold minus point two. Permissive mode blocks slop below threshold minus point four. Adaptive mode combines the slop threshold with an identifier modulo seven rule and a separate human modulo nine rule. These rules use the already known synthetic type. They are not a type blind model that can establish whether a real message is artificial. Start Real Time adds one event each timer tick, regardless of the agent count slider. At eight point five events per second, the interval is floored to one hundred seventeen milliseconds; at thirty, the fifty millisecond minimum limits nominal generation to twenty events per second. Live filtering uses a simpler threshold than stored post recalculation. Changing agent count changes simulated author names, not a real swarm. Export saves local configuration, labels, and telemetry. No feed, moderation service, or backend is contacted by this demonstration.