KASSA — K-Governed Voice Architecture

A substrate-level architecture for voice AI: commit state once, hash-lock it, and let downstream agents consume K read-only.

KASSA explores the K-Governed Voice Architecture, a commitment-kernel approach to voice AI orchestration. The core thesis is simple:

Commit state before intelligence branches.

Instead of passing caller intent through a sequence of agents that reinterpret each other's lossy outputs, KASSA extracts a canonical state representation K, binds its projections with hashes, and validates downstream work against that fixed root.

Live Demo

Open the browser demo:

Commitment Kernel Demo v7

The demo compares:

• S1: standard sequential cascade
• S2: coordinated intermediate state
• S3: K-governed parallel fanout

The page runs without credentials using cached demo data. To execute live model calls, paste an Anthropic API key into the demo panel; it is stored in sessionStorage only. The earlier v6 demo remains archived.

Repo Map

Path	Purpose
demo/	Self-contained HTML/JS visual demos for latency and drift
harness/	Commitment Conservation Harness for recursive stability experiments
schema/	K-bound projection schema and state contract
docs/	Architecture notes, reviewer response, conservation paper, deep wiki
images/	Diagrams for S1/S2/S3 and architecture visuals
.planning/codebase/	Technical handoff notes and codebase map

Technical Principle

Current voice AI often behaves like this:

audio -> STT -> partial text -> agent A -> agent B -> agent C -> response

Each stage reinterprets a degraded signal. Drift compounds with the number of handoffs.

KASSA proposes:

audio -> extract K -> hash-lock K -> read-only fanout -> validate against K

That changes the failure mode:

• Drift becomes bounded against a shared root instead of compounding through a chain.
• Agents can run in parallel because they do not depend on each other's prose.
• Governance can validate outputs against a fixed commitment state.
• Multimodal channels can be preserved independently: text, prosody, temporal behavior, affect.

K-Bound Projections

The public schema models K as independently hashed channels bound by a parent hash:

K_text      lexical content
K_pros      prosodic structure
K_tmp       temporal behavior
K_affect    acoustic affect markers

See schema/K_bound_projections_schema.json and schema/README.md.

Run The Harness

Install:

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Run a quick single-signal comparison:

export OPENAI_API_KEY=...
python harness/quick_demo.py "The user must provide a valid email address before proceeding."

Run the convergence harness:

export OPENAI_API_KEY=...
python harness/run_convergence_v2.py

The harness is a public reproducibility scaffold. It does not disclose protected production-layer MO§ES™ internals. See harness/ip_boundary.md and harness/README.md.

What This Does Not Claim

KASSA is not a generic chatbot wrapper, and the public harness is not the full production enforcement substrate.

The public layer demonstrates:

• recursive transformation stress testing
• public proxy commitment extraction
• comparison of baseline, compression, and gated conditions
• measurable drift/stability behavior over loops

The protected layer remains outside this repo:

• internal production MO§ES™ implementation
• canonical enforcement substrate
• non-public commitment algebra and deployment machinery

Engineering Context

Latency numbers in single-turn model demos can be compressed by provider-side KV caching. The observed S1/S3 gap in the demo should be treated as a floor, not a production ceiling. In multi-turn production calls, sequential context bloat should widen the gap.

Related Work

• Conservation Law paper: docs/CONSERVATION_LAW_PAPER.md
• Deep wiki: docs/DEEP_WIKI.md
• Build doc: docs/K-governed-voice-architecture-build-doc.md
• Public IP boundary: harness/ip_boundary.md

License And IP Boundary

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)
• U.S. Serial No. 19/426,028 (Utility)

MO§ES™ is a trademark of Ello Cello LLC (USPTO IC 042).

See LICENSE for the full license, including the Lineage Custody Clause.

© 2026 Ello Cello LLC. All Rights Reserved.