Stop telling your LLM what not to do. You’re accidentally expanding its universe of failure. 🌌
I ran an experiment comparing negative vs. affirmative constraints across 100 independent generations. The result? Negative prompts lead to a “leak of certainty” in the softmax distribution that compounds across agentic loops.
In this piece, I look at:
✅ Attention Sinks &amp; Semantic Failure: Seeing the raw attention weights that cause models to ignore “not.”
✅ Admissible Space: Using the RGB color cube to visualize how entropy control shrinks the space where a model is allowed to be wrong.
✅ Compounding Risk: The math behind why “95% reliable” is a project killer for multi-step agents.
If you’re building agentic workflows, reliability is a probability management problem.
<a href="https://open.substack.com/pub/tareksanger/p/llm-reliability-entropy-control?yr=5b1qx8&amp;utm_medium=ios" target="_blank" rel="noopener nofollow">Full Article Here</a>

Squad for devs building apps with GenAI and LLMs! Connect, share, exchange ideas, and get feedback. 

Build With GenAI

Negative constraints in LLM prompts (telling the model what NOT to do) expand the failure space by causing 'certainty leaks' in softmax distributions. An experiment across 100 generations compares negative vs. affirmative prompting, examining attention sinks that cause models to ignore 'not', using an RGB color cube to visualize entropy control and admissible output space, and the compounding math showing why even 95% per-step reliability becomes a project killer in multi-step agentic workflows. The core argument: LLM reliability in agentic systems is fundamentally a probability management problem best addressed through affirmative constraints.

LLL Reliability is Entropy Control