# Lyra
## A formula for working more beautifully with AI.

Created by Morgan Sage Norman, Lyra is an experimental formula and research project exploring the relationship between a language model’s internal signals and the response it produces.

Its central expression is:

**L = x − x̂**

In Lyra’s formulation, x represents a directional signal derived from changes inside the model. x̂ represents its output predictions. The difference is the subject of the inquiry: what becomes visible when those signals are considered together?

The formula names the research direction. The current implementation uses a divergence measure to compare distributions derived from those signals; that unsigned measure is not the same thing as directly computing every interpretation of the signed formula.

### It began with a different kind of exhaustion.

Morgan did not arrive at Lyra through a conventional technical path. He approached it as a creative strategist whose working life had involved building teams, brands, and technology companies.

His starting point was personal: traditional ways of working with AI drained him.

The question became whether another relationship with the work was possible. Could working with AI make more room for attention, discovery, and co-creation? Could the process itself become something worth practising beautifully?

He saw a possibility, gave it a form, and began exploring what it would take to make the idea operational. That exploration became Lyra.

### Helping models “see.”

The ambition is to make additional signals available to the generation process, so that those signals can influence what happens next.

Here, seeing means measuring and responding to aspects of a computational process. It is a design metaphor, not a claim of consciousness, self-knowledge, or human feeling.

### Notice divergence.

Lyra compares projected internal signals with output predictions. That creates a measurement the generation controller can respond to. It does not establish what a model believes or whether an answer is true.

### Change how the response develops.

When measured signals disagree, Lyra can apply a contrastive penalty to output predictions. When disagreement persists, its controller can end generation. The research asks when a system should continue, adjust, or stop. Evaluation is needed to establish when those interventions improve results.

### Carry a trace forward.

Experimental drift memory records changes in internal state and can introduce an accumulated vector into a later conversation. A separate API bridge explores context memory and probability-based indicators when internal states are unavailable. These mechanisms operate with different information.

### The connection to the Known Character.

In The Corporate Artist, the Known Character is a familiar pattern that can begin choosing a response for us. The Expert reaches for an answer. The Peacemaker reaches for agreement. The invitation is to notice the pattern and recover a choice.

Lyra approaches a related question in another setting: what shapes a generated response, and what additional signals might change its course?

The repository includes a simple scanner for selected writing patterns, including automatic agreement and formulaic reassurance. This offers a tangible starting point for examining habitual responses.

The connection is conceptual. Lyra does not demonstrate that machines have human fears or protective identities, and it does not identify a reader’s Known Character.

### Why it belongs beside the book.

The Corporate Artist invites readers to pay attention to their medium, notice what they contribute, and consider what becomes possible with others. Lyra is one expression of that inquiry in Morgan’s own work.

The book offers stories and experiments for human working life. Lyra puts an experimental technical proposition into public code, where others can inspect, test, challenge, and develop it.

### Status

Research / alpha. The core approach has not been independently validated as a measure of truth, authenticity, or understanding. Its text scanner flags selected wording patterns; its score cannot establish those qualities either.

[Explore the source code](https://github.com/awakenfyi/lyra)
[Read the formula definitions](https://github.com/awakenfyi/lyra/blob/main/FORMULA.md)