# AI MusiMuse # 010_MUSICAL_DECISION_MODEL.md Version 1.0 --- # Purpose The goal of AI MusiMuse is not to predict audio. The goal is not to predict notes. The goal is to predict musical decisions. Every composition is viewed as a sequence of creative decisions made by a composer. The AI should eventually learn these decisions and generate new ones. --- # Philosophy Traditional generators learn Audio ↓ Next Audio or Token ↓ Next Token AI MusiMuse learns Situation ↓ Decision ↓ New Situation Every generated track is therefore a chain of musical decisions. --- # Decision Pipeline ``` Current State ↓ Decision ↓ State Update ↓ Decision ↓ State Update ↓ Decision ↓ ... ``` Generation is iterative. --- # Musical State At every moment the Composer has an internal state. Example ``` Current section Current energy Current density Current harmony Current rhythm Current texture Current tension Elapsed time Remaining duration ``` The state continuously evolves. --- # Decision Types Every decision belongs to a category. Examples ``` Structure Harmony Rhythm Texture Dynamics Energy Effects Timing Variation Ending ``` --- # Example Decisions ``` Create Intro Increase Energy Reduce Density Introduce Bass Introduce Melody Open Filter Repeat Theme Change Harmony Remove Percussion Create Break Return Theme Finish Track ``` --- # Decision Object Every decision contains ``` id category time confidence parameters ``` Example ``` Decision IncreaseEnergy time = 62 s confidence = 0.93 target = 0.72 ``` --- # Decision Sequence Example ``` Start Atmosphere ↓ Add Texture ↓ Introduce Bass ↓ Increase Density ↓ Introduce Theme ↓ Break ↓ Return Theme ↓ Ending ``` The sequence becomes the true representation of composition. --- # Decision Dependencies A decision depends on Current State + Style + History + Constraints The same decision should not always occur. The Composer should remain creative. --- # Creativity Every decision has a probability. Example ``` Return Theme 0.84 Create New Theme 0.16 ``` Random Seed determines which path is chosen. --- # Composer Memory The Composer remembers Current composition only. It never memorizes generated tracks. Training data is immutable. --- # Decision Graph Instead of a linear sequence, the Composer internally builds a graph. Example ``` Atmosphere ↓ Theme ↓ Development ├──── Break └──── Variation ↓ Return ↓ Ending ``` Generation is graph traversal. --- # Constraints The user may constrain decisions. Examples ``` No percussion Dark atmosphere Long intro Slow evolution Minimal harmony Ambient only No climax ``` Constraints influence probabilities. They never directly dictate music. --- # Style Guidance The Style Model modifies probabilities. Example ``` Author usually adds bass after 45 seconds ↓ Probability increases. ``` --- # Decision Evaluation Every decision is evaluated. Questions ``` Does it fit the style? Does it increase coherence? Does it violate constraints? Does it create repetition? Does it improve variation? ``` Only accepted decisions become part of the composition. --- # Future Learning The AI will eventually learn ``` State ↓ Decision ``` instead of ``` Audio ↓ Audio ``` This dramatically reduces training complexity. --- # Renderer Independence The Composer never generates sound. The Composer generates decisions. Another subsystem converts Decisions ↓ Timeline ↓ Events ↓ Audio The rendering engine may change without affecting the Composer. --- # Long-Term Goal The Composer should eventually make decisions similarly to the human author. Not because it memorized songs, but because it learned the author's compositional habits. Every generated piece should therefore feel authentic, while remaining completely new.