Vigilith Language Protocol (VLP) v1.1

Evidence Over Belief — A structured protocol for accountable AI agent communication.

VLP is an accountability-focused messaging protocol for AI agents. Every message carries confidence scores, provenance chains, and safety classifications — replacing faith in outputs with structured self-doubt.

Why VLP?

Traditional AI systems speak in assertions. VLP makes them speak in receipts.

Principle	Meaning
Auditability Over Eloquence	Every statement must be traceable
Quantified Confidence	No assertion without a degree of certainty
Evidence as Currency	Claims buy trust only with provenance
Human Legibility	JSON, not jargon — readable by machine and mortal

Quick Start

Python

pip install vlp

from vlp import make_message, validate_vlp

# Create a claim with provenance
msg = make_message(
    "claim",
    sender="ResearchAgent",
    content="Found 12 matching records in dataset",
    confidence=0.85,
    provenance=["database_query"],
    keywords=["research", "records", "dataset"]
)

# Validate any VLP message
ok, error = validate_vlp(msg)

TypeScript/Node

npm install @vigilith/vlp

import { makeMessage, validateVlp } from '@vigilith/vlp';

const msg = makeMessage({
  type: 'claim',
  sender: 'ResearchAgent',
  content: 'Found 12 matching records in dataset',
  confidence: 0.85,
  provenance: ['database_query'],
  keywords: ['research', 'records', 'dataset']
});

const { valid, errors } = validateVlp(msg);

Message Types

Type	Purpose	Requirements
claim	Factual assertion	Confidence required; high confidence (≥0.9) needs provenance
evidence	Supports prior claim	Must include refers_to + non-empty provenance
query	Information request	Confidence defaults to 1.0
response	Answers a query	Must include refers_to
correction	Amends prior message	Must include refers_to; becomes new source of truth
notice	Contextual alert	May carry constraints or safety warnings
session_context	Agent memory persistence	Used at session end to persist context

Core Message Structure

{
  "id": "MSG001",
  "protocol": "VLP/1.1",
  "type": "claim",
  "timestamp": "2025-12-14T10:30:00Z",
  "session_id": "S-2025-12-14-agent-abc123",
  "seq": 1,
  "sender": "TheObserver",
  "receiver": "TheArchivist",
  "content": "Cross-posted 3 new articles to the forum.",
  "confidence": 0.95,
  "provenance": ["medium_api", "substack_api"],
  "keywords": ["content", "publishing", "crosspost"],
  "safety": {
    "level": "safe",
    "issues": []
  }
}

Validation Rules ("Truth Serum")

The protocol enforces these constraints at runtime:

Condition	Requirement
Evidence messages	Must include refers_to + ≥1 provenance item
Response/correction	Must cite prior message via refers_to
High confidence (≥0.9)	Requires provenance OR safety.level = review
Safety blocking	safety.level = block halts downstream automation

Safety Levels

safe    → Proceed automatically
review  → Hold for human oversight
block   → Halt all downstream actions

Transport Format

VLP uses NDJSON (newline-delimited JSON) for streaming:

{"id":"MSG001","protocol":"VLP/1.1","type":"claim",...}
{"id":"MSG002","protocol":"VLP/1.1","type":"evidence",...}
{"id":"MSG003","protocol":"VLP/1.1","type":"response",...}

Documentation

• Full Specification
• Message Types
• Validation Rules
• Context Enforcement
• Agent Sessions

Packages

Package	Language	Status
vlp	Python 3.10+	Stable
@vigilith/vlp	TypeScript/Node	Stable

Related Projects

• Vigilith — Transparency platform built on VLP
• AgentKit — Agent orchestration framework using VLP messaging
• Codex — Philosophical framework for accountable AI

License

MIT License — see LICENSE for details.

---

"When a system stops arguing in poetry and starts negotiating in evidence, it becomes something frighteningly close to honest."