FBN
Exhibit VI / 16AI/ML — 2026Research engineer, sole build← Index

Computational Entropy Lab

Does a muddled prompt cost more electricity? A research laboratory that measures the energy price of ambiguity, token by token.

Computational Entropy Lab cover

FBN.2026.06

Computational Entropy Lab

Research engineer, sole build

Python, FastAPI, React, Ollama, 2026

9

controlled mutation conditions

mJ

per token, measured on the GPU

6

statistical tests, with effect sizes

1

click to export a reproducibility package

The problem

Language models are judged on what they answer and almost never on what answering costs. No one measures whether a vague, contradictory or bloated prompt burns more energy than a clear one.

The approach

A mutation engine rewrites prompts under nine controlled conditions (typo noise, verbosity, ambiguity, contradiction, negation, reordering, formality, code-switching and a clean baseline) at tunable intensity, with length-matched controls that separate how long a prompt is from what it means. Each run records energy per token, using real GPU wattage through NVIDIA NVML and labelling every estimated figure as an estimate. A statistics suite of Pearson, Spearman, Kendall, ANOVA, Kruskal-Wallis and effect sizes then correlates the Semantic Instability Index against energy to yield the Prompt Entropy Coefficient.

The outcome

A reproducible laboratory, from a live WebSocket dashboard to LaTeX tables and a downloadable reproducibility package, built on one principle: real and estimated measurements are never silently blended.

What it does

4 capabilities

A new coefficient

The Prompt Entropy Coefficient relates how unstable a prompt is to how much energy each token costs.

Honest instruments

Every measurement is tagged as real NVML wattage, a TDP proxy or a simulation, so no estimate masquerades as data.

Controls that matter

Length-matched variants separate a longer prompt from a more confusing one.

Ready for a paper

LaTeX tables, CSV and JSON export, and a zipped package of raw data, protocol and environment.

The strata

How it stands
01 · DASHBOARDReact with Recharts, live over WebSocketsfig. I02 · APIFastAPI orchestrating experiments and persistencetav. II03 · MUTATION ENGINENine seeded, reproducible conditions with readability scoringc. III04 · INFERENCEOllama (phi3.5 by default), optional hosted providersnota IV05 · ENERGYNVIDIA NVML for the GPU, labelled estimates for CPU and memory…fol. V06 · STATISTICSCorrelation with confidence intervals, ANOVA, Kruskal-Wallis, …fig. VIComputational Entropy Lab, delineavit F.K.R.

  1. Dashboard

    React with Recharts, live over WebSockets

  2. API

    FastAPI orchestrating experiments and persistence

  3. Mutation engine

    Nine seeded, reproducible conditions with readability scoring

  4. Inference

    Ollama (phi3.5 by default), optional hosted providers

  5. Energy

    NVIDIA NVML for the GPU, labelled estimates for CPU and memory, CodeCarbon cross-check

  6. Statistics

    Correlation with confidence intervals, ANOVA, Kruskal-Wallis, effect sizes

Next exhibit — VII

BlogHaven