A newer version of the Gradio SDK is available: 6.26.0
title: Instruction Drift Simulator
emoji: 😻
colorFrom: gray
colorTo: green
sdk: gradio
sdk_version: 6.3.0
app_file: app.py
pinned: false
license: mit
short_description: Instruction Drift Simulator
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
👩💻 Author
Aditi Khare — AI Product & Research Leader
Writing on AI research, product thinking, and system architecture
🌐 Website: aditikhare.com
🔗 GitHub Repository: Instruction Drift Simulator
🤗 Live Demo: View on Hugging Face
💼 LinkedIn: Aditi Khare
🧠 Instruction Drift Simulator
The Instruction Drift Simulator demonstrates how model behavior can systematically drift away from original instructions as prompts evolve, context accumulates, or constraints weaken over time.
This Space is a controlled evaluation environment designed to make instruction drift visible, measurable, and comparable—not to correct or optimize prompts.
What This Demo Shows
- How initial instructions degrade across multi-step interactions
- How ambiguity, overload, or soft constraints introduce drift
- How behavior changes without any change in model weights
- How instruction adherence weakens over time
Evaluation Focus
The simulator evaluates behavior step-by-step across:
- Instruction adherence
- Constraint leakage
- Semantic deviation
- Accumulated ambiguity
- Termination correctness
Drift is surfaced explicitly, not inferred.
What This Is
- A behavioral evaluation demo
- A controlled instruction-drift simulation environment
- A non-autonomous, non-agentic system
- A diagnostic tool for GenAI reliability
What This Is Not
- A chatbot
- A prompt optimization tool
- A RAG system
- An autonomous agent
- A mitigation or correction engine
This simulator does not fix drift—it exposes it.
Why This Matters
Instruction drift is a major source of:
- Silent failures in agentic systems
- Policy non-compliance
- Evaluation blind spots
- Misleading benchmark performance
This demo treats instruction fidelity as a first-class evaluation signal.
License
© 2026 Aditi Khare. All rights reserved.