YANA RUNTIME

The model proposes.
The human holds authority.

A bounded TurnEngine, formal capability, explicit approval, HALT and audit receipts — authority never lives in a prompt.

01 · Modelproposes a capability
02 · Authority chainallow · ask · deny
03 · Capabilityexecutes within scope
04 · Evidencerecords the result

AI CAN ACT

A model can write files, run commands, call tools — unbounded, with no one watching.

Put authority outside the model. That's Yana Runtime.

Anthropic
OpenAI
Gemini
Groq
Ollama (local)
+ DeepSeek, OpenRouter
Yana RuntimeMission · Context · Turn Engine · Governance · Capability · Evidence
Files
Git
Shell / Process
System

ONE RUNTIME

Several different interaction surfaces.

All going through one Yana Runtime — not a separate copy per surface.

Yana Studio
Runtime CLI
MCP (opt-in, experimental)
Yana RuntimeStudio spawns the exact same yana-rt binary the CLI uses — same TurnEngine, not two implementations

ANY AI

Change the model.

Keep the entire system.

MODEL IN USE
AnthropicOpenAIGeminiGroqDeepSeekOpenRouterOllama (local)

Pick a provider — Yana automatically uses that provider's default model, no catalog lookup needed.

DOESN'T CHANGE
✓ Mission
✓ Tasks
✓ Project memory
✓ Authority chain
✓ Evidence

HUMAN-GOVERNED

The model proposes an action.

A human decides whether it happens.

AI proposal"Modify 43 files"
Yana evaluatesYanaAuthorityChain
Blast radius policysrc/guard/blast_radius.rs
Approval requiredwaiting, not executed
Human decisionallow · limit · deny

DETERMINISTIC CONTROL

AI doesn't decide for itself how far it reaches.

Three capability groups, each with its own boundary.

Access

file readgit statusgit log

Change

file_mutationconfig_writegit commit

Execute

command (shell)processsystem

GOVERNANCE

Authority is architecture,
not a badge.

Policy lives outside the model. Capability leases are scoped and revocable. Giám Thị HALT holds the highest stop authority.

YanaAuthorityChain

Deterministic preflight before any real action.

filesystem workspace only shell ask network ask destructive deny

Capability leases

Time-boxed, scoped, subject-specific permission.

subject current session scope /workspace ttl 27 min revoke available

Fail loudly

Missing governance must surface, never silently skip a permission check.

guard missing → stop denied → visible HALT → highest

CONTINUITY

Work outlives a single chat.

Typed task graph, mission, L1/L2 memory and project memory keep dependency, decisions and provenance across sessions.

Missionactive
Task graph6 nodes · 8 edges
Project facts34 indexed
Conversationreplaceable surface

EVIDENCE ENGINE

No evidence, not called done.

File diff, test result, build result, exit code and approval form an evidence package a human reviews, accepts or rolls back.

Implementation
Tests38 / 38
BuildExit 0
Human gatewaiting