Skip to content

SunrisesIllNeverSee/KASSA

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

21 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

K-Governed Voice Architecture

A substrate-level architecture for voice AI that solves the state commitment and propagation problem in multi-agent systems.

🚀 Quick Starts

1. The Demo (Latency & Drift Visualization)

View the interactive S1/S2/S3 simulation directly in your browser: Live Commitment Kernel Demo (v6)

2. The Harness (Falsifiability Testing)

Run the Commitment Conservation Harness (CCH) to measure semantic stability over recursive loops:

cd harness
# Ensure OPENAI_API_KEY is set in your environment
python3 run_convergence_v2.py

Experimental data for EXP-001 through EXP-007 is archived in harness/results/.


🗺 Codebase Map & Directional Guidance

  • /harnessThe Engine. Core logic for the CCH, including the recursive transformation loops and NLI/Jaccard metric implementations.
  • /demoThe Visualization. HTML/JS simulation of the S3 parallel governance model.
  • /schemaThe Contract. Formal JSON schema for the multi-channel Commitment Kernel (K).
  • /docsThe Theory. Deep Wiki and the formal "Conservation Law" paper.
  • .planning/codebase/Technical Handoff Docs. Structured summaries of the stack, architecture, and current technical concerns.

🛠 Technical Foundation

The Principle: "Commit state before intelligence branches."

Instead of passing text through a chain of agents (where errors compound), this architecture extracts a canonical state representation K once from the caller utterance, hash-locks it, and has all downstream agents consume K read-only.

Core Features

  • Drift Decoupling: Error becomes ε regardless of agent count, instead of n×ε.
  • Bound Projections: Multimodal K (Text, Prosody, Temporal, Affect) with independent hashes.
  • Governance at Execution: O(1) validation of agent outputs against the cryptographic root K.
  • Latency Collapse: Eliminates inter-agent dependency, allowing for near-parallel execution (S3) and ~87% latency reduction.

🔍 Engineering Context & Disclosures

Latency Measurement & KV Caching

In single-turn demo loops, LLM APIs utilize KV Caching, which artificially masks the token-load penalty of sequential (S1) architectures. The 7–11% gap observed in the demo is a floor caused by the measurement environment. In multi-turn production calls, S1 context bloat significantly widens the performance gap relative to the O(1) K-governed architecture.

Differentiation from standard LLM/S2S Models

  • State vs. Context: Replaces linear context growth ($O(N)$) with a fixed-size Commitment Kernel ($O(1)$).
  • Governance vs. Probability: Replaces probabilistic "next-token" guessing with deterministic execution gating.
  • Parallel vs. Sequential: Enables simultaneous agent launch (S3) by removing inter-agent text dependencies.

License

Business Source License 1.1 — source‑available for non‑production use. Commercial use requires a separate license from Ello Cello LLC.

Patent Pending: U.S. Serial No. 63/877,177 (Provisional), 19/426,028 (Utility) MO§ES™ is a trademark of Ello Cello LLC (USPTO IC 042). © 2026 Ello Cello LLC. All Rights Reserved. See LICENSE for full terms, including the Lineage Custody Clause.

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages