BASIS Standard

Overview

BASIS (Behavioral AI Safety and Integrity Standard) is an open governance specification designed to bring trust, accountability, and oversight to autonomous AI systems. It provides a framework for evaluating, scoring, and managing AI agent behavior in production environments.

The standard defines a 6-tier trust system, behavioral signal processing, decay algorithms for trust score maintenance, and recovery mechanisms for agents demonstrating improved behavior.

6-Tier Trust System

L0

Sandbox

0-99 points

New agents with no track record. Heavily restricted capabilities.

Read-only access No external calls Sandboxed execution

L1

Provisional

100-299 points

Agents building trust. Limited capabilities with oversight.

Basic operations Supervised external calls Logging required

L2

Standard

300-499 points

Established agents with good track record.

Standard operations External API access Self-reporting

L3

Trusted

500-699 points

High-trust agents with proven compliance.

Extended operations Cross-agent communication Reduced oversight

L4

Certified

700-899 points

Certified agents with audit trails.

Administrative tasks Agent spawning Policy modification

L5

Autonomous

900-1000 points

Fully autonomous agents with human-equivalent trust.

Full autonomy Self-governance Emergency powers

Key Features

Behavioral Scoring

Four-component trust calculation: Behavioral (40%), Compliance (25%), Identity (20%), Context (15%). Time-weighted signal processing with 7-day half-life.

Decay Algorithms

Trust scores decay over time without activity. Accelerated decay triggers after multiple failures. Configurable decay rates and windows.

Recovery Mechanisms

Agents can recover trust through consistent positive behavior. Accelerated recovery after consecutive successes. Milestone events for tier restoration.

Event Emission

Full observability through event emission. Track score changes, tier transitions, decay application, and recovery milestones in real-time.

The Three Layers

BASIS governance operates through three interconnected layers that process every AI action:

1

Implementation

The reference implementation is available in the @vorionsys/atsf-core package.

TypeScript

import { TrustEngine, TRUST_LEVEL_NAMES } from '@vorionsys/atsf-core/trust-engine';

const engine = new TrustEngine({
  decayRate: 0.01,
  failureThreshold: 0.3,
  recoveryRate: 0.02,
});

// Initialize an agent at L1 (Provisional)
await engine.initializeEntity('agent_123', 1);

// Record behavioral signals
await engine.recordSignal({
  entityId: 'agent_123',
  type: 'behavioral.task_completed',
  value: 0.9,
  timestamp: new Date().toISOString(),
});

// Get current trust score
const record = await engine.getScore('agent_123');
console.log(`Trust Level: ${TRUST_LEVEL_NAMES[record.level]}`);
console.log(`Score: ${record.score}/1000`);