vinfty 0.1.0

V∞ — cognitive observability layer for any LLM orchestration framework

V∞ (vinfty)

Cognitive observability layer for any LLM orchestration framework.

Vinfty is a featherweight (~300 lines) Python library that adds cognitive observability to your LLM pipelines — without replacing your existing tools. It measures ont_self (self-consistency), tracks C_ij coupling (cross-session memory coherence), and detects HMM s₀/s₁ state transitions — so you know whether your system is healthy, not just whether it returned a valid JSON.

pip install vinfty

---

Why?

LangChain, CrewAI, AutoGen — they all solve "how to call LLMs in sequence." None of them answer:

• Is my agent system getting more coherent or more chaotic over time?
• Are my tool calls producing consistent results or contradicting each other?
• Is the system in a stable (s₀) or active-search (s₁) mode?

Vinfty adds that layer. It wraps any orchestration framework and gives you a cognitive dashboard: ont_self trajectory, coupling matrix, HMM state, palace routing.

Quick Start

from vinfty import V9Orchestrator

engine = V9Orchestrator(active_k=100)

# Register any function as a tool → maps to a Palace domain
@engine.register(palace="P_search")
def search_web(q: str) -> str:
    return f"search results for {q}"

@engine.register(palace="P_analysis")
def calculate(expr: str) -> float:
    return eval(expr)

# Each call updates the cognitive state
engine.step("search latest LLM papers", tool=search_web)
engine.step("how many published this month?", tool=calculate)

# Check system health
report = engine.report()
# → {
#   "ont_self": 0.72,
#   "hmm_state": "s0"|"s1",
#   "palace_flow": ["P_query", "P_search", "P_analysis"],
#   "c_ij_density": 0.34,
#   "memory_count": 12,
# }

Core Concepts

Concept	What it measures	Analogy
ont_self	System self-consistency	"Is my agent making consistent decisions?"
C_ij	Cross-session memory coupling	"Do past decisions influence present ones coherently?"
HMM s₀/s₁	Cognitive mode	"Is the system in stable execution (s₀) or active exploration (s₁)?"
Palace	Functional domain	"Which cognitive domain is each tool call serving?"

Adapters

Vinfty ships with adapters for common frameworks:

# LangChain adapter
from vinfty.adapters.langchain import LangChainAdapter

adapter = LangChainAdapter(agent)
report = adapter.run_with_observability(user_query)
# Same result as agent.run(), plus cognitive report

No LLM Required

Vinfty is pure symbolic computation — no API keys, no GPU, no model weights. The cognitive metrics are derived from the structure of your tool calls and memory traces, not from the content of LLM responses.

License

MIT