Temperature Tuning & the Curious Case of the Wandering Rising Sign

I ran a small experiment on the sonifyr code and it showed me something unexpected.

I set out to test how temperature affects:

• Structure

• Creativity

• Tone

• Schema reliability

• Astrological accuracy

I kept everything constant: birth data, prompt, schema, model, week range.

The only thing I changed was temperature.

Lower temperature produced more predictable phrasing.

Higher temperature produced more expressive variation.

Exactly as expected.

What I did not expect was identity drift.

What Worked Across All Temperatures

No matter the setting:

• The voice stayed warm and grounded

• The tone remained supportive

• The symbolism stayed rich

• The affirmations held

• The structure remained valid

• Creativity shifted slightly. Expression varied.

• But the creative layer was stable.

  
  

What Didn’t Work

The identity layer drifted.

• At some temperatures, the Sun, Moon, and Rising were correct.

• At others, the Rising sign shifted.

• At lower entropy, it even collapsed into the Sun sign.

This ruled out randomness as the cause. If temperature were responsible, higher settings would show more instability. Instead, identity inconsistencies appeared independently of entropy. That suggests something deeper than variation.

The Real Issue

When identity elements drift, it usually means one of three things:

• The model is being allowed to reinterpret context.

• Core data isn’t consistently injected or prioritized.

• The system is inviting inference instead of grounding.

In my case, all three were happening.

The system blurred the line between calculation and interpretation.

Sometimes the model was interpreting precomputed data.

Other times, it was reconstructing identity from raw inputs.

What This Experiment Actually Revealed

If a system mixes deterministic computation (Math) with generative interpretation (LLM) without clearly separating them, drift becomes inevitable.

The model doesn’t “get confused.” It fills in gaps. Especially in symbolic domains like astrology, where archetypes are statistically familiar and easy to complete. When precision is requested but not explicitly provided, the model supplies it probabilistically.

The Takeaway

This experiment started as a study in creativity tuning. It ended as a lesson in architecture.

Where creativity was stable, identity was not. Which means temperature was working correctly. The instability wasn’t randomness. It was ambiguity.

When math and meaning share the same layer, identity can wander.

Separate them and identity stays grounded.