Skip to content

arkavo-org/specifications

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

14 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Arkavo Specifications

Technical specifications for the Arkavo ecosystem.

Specifications

game-rl/

Game-RL Protocol: Multi-Agent AI Interface for Game Environments

The Game-RL Protocol defines a standard interface for multi-agent AI systems to observe and interact with game environments, functioning either as embodied characters (APCs/NPCs) or systemic controllers (Game Masters). Built on MCP (Model Context Protocol), it provides:

  • Dual Agency Scopes: Embodied agents (physics-bound) vs Systemic agents (god-mode)
  • Clock Modes: Training (agent-owned, deterministic) vs Live (engine-owned, real-time)
  • Session Topologies: Exclusive (stdio) vs Shared (IPC for Claude/ChatGPT Desktop)
  • Role-Based Permissions: GameMaster can spawn/teleport; EntityBehavior can only move/interact
  • Vision Streams: High-performance shared memory for pixel observations

Wire format: JSON-RPC 2.0 over MCP (stdio/IPC)

Latest Draft: draft-arkavo-game-rl-00

JSON Schemas: schemas/game-rl/draft-00/

Reference Implementation: GameOfMods


ntdf-rtmp/

NTDF-RTMP: NanoTDF Manifest Transport over RTMP

NTDF-RTMP specifies methods for transporting NanoTDF policy manifests over RTMP streaming connections, enabling real-time encrypted media with dynamic policy updates. It provides:

  • In-Band Transport: Manifests via AMF onMetaData or RTMP Data Messages
  • Session State Machine: AWAITING_MANIFEST → ACTIVE → REKEYING states
  • Receiver Protocol: Buffering, manifest parsing, and key derivation
  • Compatibility: Works with FFmpeg, OBS, Wirecast via standard onMetaData

Wire format: AMF-encoded manifest in RTMP metadata

Latest Draft: draft-arkavo-ntdf-rtmp-02

Reference Implementation: arkavo-rs


ntdf-token/

NTDF Tokens: NanoTDF-based Authentication Tokens

NTDF tokens are cryptographically-bound authentication tokens that replace traditional JWT Bearer tokens. Built on OpenTDF's NanoTDF specification, they provide:

  • Proof-of-Possession: DPoP (RFC 9449) integration prevents token theft
  • Policy Binding: Cryptographic GMAC binding enforces access control at the KAS
  • Confidentiality: AES-256-GCM encrypted payload protects claim data
  • Provenance: Ed25519 signatures ensure token integrity and origin

Wire format: Authorization: NTDF <Z85-encoded-nanotdf>

Latest Draft: draft-arkavo-ntdf-token-00

Reference Implementation: arkavo-rs


torg-decision/

TØR-G: Token-Only Reasoner (Graph) Intermediate Representation

TØR-G is a zero-parser, boolean-circuit Intermediate Representation designed for AI agent policy synthesis. It treats LLM output as a direct stream of graph construction instructions, providing:

  • No Text: The language consists solely of atomic tokens
  • No Parsing: The "compiler" is a deterministic state machine
  • Strict Booleanism: All logic reduced to boolean decision combinatorics
  • Verifiability: Every generated program is a finite DAG amenable to SAT/BDD verification

Wire format: Direct token stream with logit masking enforcement

Latest Draft: draft-arkavo-torg-decision-00


tdf-json/

TDF-JSON: Trusted Data Format — JSON Serialization

TDF-JSON defines a JSON-based serialization for the Trusted Data Format, providing encrypted payloads with policy-bound key access.

Latest Draft: draft-00


tdf-cbor/

TDF-CBOR: Trusted Data Format — CBOR Serialization

TDF-CBOR defines a compact CBOR-based serialization for the Trusted Data Format, optimized for constrained environments and reduced wire size.

Latest Draft: draft-00


agent-runtime-policy/

Agent Runtime Policy (ARP): Runtime Adaptation for AI Agents

ARP provides a standard format for describing how AI agents adapt their behavior at runtime. A companion to the Agent Definition Language (ADL), ARP defines the adaptation machinery that governs an agent's operational evolution within its declared boundaries:

  • Adaptation Engine: Thompson Sampling, UCB1, epsilon-greedy with Bayesian priors
  • Multi-Timescale Feedback: Per-event quality gates, PolicyCache with temporal decay, peer gossip, offline consolidation
  • One-Way Ratchet: Automated adaptation may only tighten policy; relaxation requires human approval
  • Cross-Layer Escalation: Events in one layer (cognitive, execution, data, network) trigger policy tightening in others
  • Budget & Velocity Constraints: Per-task ceilings, per-minute spend limits, degradation chains
  • Cryptographic Integrity: Document signing, ADL binding, tamper-evident state storage

Wire format: application/arp+json (JSON canonical), .arp.toml (authoring)

Latest Draft: arp-spec-draft-00

JSON Schemas: schemas/agent-runtime-policy/draft-00/


JSON Schemas

Machine-readable schema definitions for protocol validation:

Specification Schema Directory
Game-RL schemas/game-rl/draft-00/
Agent Runtime Policy schemas/agent-runtime-policy/draft-00/

License

Apache 2.0 - See LICENSE

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors