ALSE 5.0 — Deterministic AI for DAWs

The AI infrastructure layer
for music creation tools

A deterministic execution harness for AI-assisted recording. Not a chatbot bolted onto a DAW. A layered architecture that sits between the user, the LLM, and the studio — solving hallucination, state sync, and the psychological safety net problem.

The Architecture Contact Us
21
Months of development
1,900+
Tests across 128 modules
3
Infrastructure locks
3
Isolated personas
Architecture

Not a feature. A platform.

ALSE is built on six systems that together solve the fundamental problem of integrating non-deterministic AI with deterministic DAW state management.

Session Context Model

A structured, SQLite-backed representation tracking song title, production stage, take count, last action, user skill level, and session history. Updated by every action. Read before every response.

Execution Harness

Pre-flight snapshot capture, atomic execution, validation, and rollback on failure. Wraps every DAW mutation in a deterministic safety layer. Preserves every take — even failed ones.

Three-Persona Routing

Reese (Concierge), James (Producer), Ray (Engineer). Each persona has a constrained tool surface that cannot overlap. Creative feedback, technical execution, and system health are isolated by design.

Hybrid Intent Classifier

Four-level waterfall: state override → fast-path regex (sub-millisecond) → brain router → AI fallback. The deterministic layer handles critical commands without touching a language model. Non-deterministic paths are sandboxed.

Real-Time State Push

WebSocket EventBus replacing HTTP polling for sub-100ms UI sync. Voice and Companion App converge on the same state model. No stale state, no race conditions between modalities.

User Model

Cross-session learning that adapts behavior: same question gets a different answer on session 1 (pedagogical) vs. session 50 (direct). Grows with the user from first-time recorder to experienced producer.

Infrastructure

Three infrastructure locks for safe AI execution

Every major audio company is currently hitting the same wall: LLMs are non-deterministic, but DAWs require deterministic state. These three locks close the gap that every chatbot-on-DAW wrapper ignores.

Lock 01

Harness Transaction IDs

Every multi-step command generates a harness_transaction_id (UUIDv4) in preflight. Every event emitted during execution carries it. Subscribers can query was_rolled_back() before committing state. No more orphaned events or ghost states.

Lock 02

Event Compensation

When the harness rolls back, subscribers who already consumed forward events receive inverse compensation events with the same correlation_id. The Session Context Model and User Observation Service revert cleanly. All-or-nothing — silent partial rollback is not an option.

Lock 03

Severity-Graded Failures

Three failure grades: transient (earcon only, the system recovers silently), permanent (TTS explanation, actionable next steps), and catastrophic (TTS + SESSION_LOCKED state + Companion overlay). The user always knows what happened and what to do next.

Two properties hold across all three: capture is gated behind a deliberate footswitch tap, and the system reports only verified DAW state — it says a track is muted because the DAW confirms it, not because a command was sent.

The North Star

"Blue Moon" — the scenario that drives every decision

A 17-year-old unboxes ALSE. They have a song called "Blue Moon" about a breakup. They have never used a DAW. They do not know what a bus is, what gain staging means, or how to ask for what they want. The system must bridge that gap — from first command to finished recording.

The concierge who doesn't execute commands

Reese is the front door. She does not mute tracks, arm inputs, or touch transport. Her tool surface is constrained to conversation, intent routing, and failure explanation. When a command comes in, she classifies it and delegates — technical command goes to Ray (the engineer who never speaks), creative question goes to James (the producer who never touches the DAW).

This separation is what makes the system deterministic. If Ray's process crashes, James and Reese continue independently. The studio does not go down because one persona failed.

"Done. I've undone that last change. Everything is back to where it was."
— the concierge, on undoing a change. Verbatim from the system.

Interested in the architecture?

We are looking for platform partners and engineering teams who understand that integrating AI into creative tools requires infrastructure, not features.

Or reach us directly at contact@benblends.ai