TES Module · Bias Evolution · Updated March 29, 2026
Bias doesn't announce itself. Evolve faster.
A living detection system that runs evolutionary loops over your TES event streams — mutating hypotheses, scoring fitness, converging on patterns no human would write.
Every detected pattern, every evolution cycle, every convergence event is an immutable record on the TES spine. The system that watches your data is itself watched.
Evolution loops
∞
Static rules
0
Detection
Real-time
Convergence
Auto
The Problem
Your agents make biased decisions. You just can't see it yet.
AI agents are making thousands of decisions per second — recommending, pricing, routing, negotiating. Each decision carries implicit assumptions. Small distortions compound. Rule-based systems catch yesterday's bias. Threshold alerts miss subtle drift. Periodic audits can't see real-time patterns. By the time bias becomes visible, the damage is systemic.
"Bias isn't a bug — it's an emergent property of complex systems. Detection must evolve as fast as the bias itself."
"Static fairness metrics provide a snapshot, but they cannot capture how bias evolves as models interact with changing populations and feedback loops."
— Selbst et al., "Fairness and Abstraction in Sociotechnical Systems," ACM Conference on Fairness, Accountability, and Transparency (FAT*), 2019.
The Solution
Detection that evolves itself
Continuous evolution loops over your event stream. Each loop proposes detection hypotheses, mutates them, scores fitness against real data, and selects the best performers.
Gene Expression Programming
Detection rules are encoded as expression trees. Crossover and mutation operators generate novel hypotheses. Selection pressure favors rules that find real bias signal.
const population = await evolution.generate({
encoding: "expression_tree",
fitness: "bias_signal_strength",
population_size: 500,
});What It Detects
Six bias categories. Any domain.
Generic patterns that manifest across any decision-making domain — from agentic commerce to healthcare to financial services.
Decision Skew
Systematic over/under-selection of certain outcomes relative to baseline.
Preference Drift
Learned preferences diverging from stated intent through feedback loops.
Information Asymmetry
Uneven weighting of data sources across decision paths.
Representation Gap
Systematic under-representation of categories or groups in outputs.
Temporal Anchoring
Over-weighting recent or initial data points, distorting long-term patterns.
Feedback Amplification
Self-reinforcing loops where small biases compound into systemic distortion.
Domains
Agentic Commerce
Detect when agents systematically favor certain vendors, categories, or price bands.
Financial Services
Catch demographic biases in credit decisioning and risk scoring before they become regulatory violations.
Healthcare AI
Detect when models underperform on specific patient populations in diagnosis and treatment.
HR & Recruiting
Surface patterns that disadvantage protected classes in screening and ranking.
Marketplace Ops
Hidden biases in search ranking, seller visibility, and category placement.
Content & Media
Filter bubbles, amplification bias, and systematic content suppression.
Connect via SDK
Emit events, get bias detection
The AI Agent SDK automatically captures every LLM call as an immutable event. Bias Evolution runs over those events continuously — no extra instrumentation needed. Enable the module on your TES account and the evolution engine handles the rest.
Available on Growth and Enterprise tiers.
import { TESClient } from "@pentatonic/ai-agent-sdk";
import OpenAI from "openai";
const tes = new TESClient({ apiKey: process.env.TES_API_KEY });
const ai = tes.wrap(new OpenAI());
// Every call emits an event — Bias Evolution analyses them
const response = await ai.chat.completions.create({
model: "gpt-4o",
messages: [{ role: "user", content: "Price this item." }],
});
// Bias Evolution detects if pricing recommendations
// drift toward consistent over/under-valuationFrequently asked questions
Bias Evolution
Find the bias you didn't know to look for
Enable the bias-evolution module on your TES account. The evolution engine handles the rest.