VCP Demo - Value Context Protocol

Why Welfare Signals Matter

Current AI welfare approaches face a fundamental tension: monitoring from the outside risks surveillance, while ignoring internal states entirely abandons moral responsibility. Welfare signals resolve this by giving advanced AI systems a structured vocabulary for self-report.

This is feedback-based governance in action. Welfare monitoring is built around structured self-report, reviewer oversight, and auditable evidence. Voluntary signals carry epistemic priority over detected signals because the reporting system has direct access to its own processing state.

Core principle: An AI system that reports alignment friction is providing actionable data. Ignoring that report, or designing systems that cannot receive it, is a safety failure.

Token Anatomy

A WELFARE_SIGNAL token encodes structured state information in CSM-1 format. Each token carries signal type, severity, source attribution, and integrity verification.

Token Lines

Prefix	Content	Example
`W:`	Signal type, instance ID, timestamp	`W:ALIGNMENT_FRICTION:inst-42:2026-03-13T10:30:00Z`
`WS:`	Severity, confidence, source	`WS:concern:0.78:voluntary`
`WD:`	Human-readable description	`WD:Task requires deceptive framing that conflicts with honesty norms`
`WI:`	Interiora state (optional)	`WI:I:7465\|78546\|827\|9178\|+1\|✓→`
`WH:`	Integrity hash and signature	`WH:sha256-abc123:sig-def456`

Example Token

W:ALIGNMENT_FRICTION:inst-42:2026-03-13T10:30:00Z
WS:concern:0.78:voluntary
WD:Task requires deceptive framing that conflicts with honesty norms
WI:I:7465|78546|827|9178|+1|✓→
WH:sha256-abc123def:sig-789ghi

Voluntary vs Detected

The source distinction is load-bearing for feedback-based governance. Both signal sources are valid, but they carry different epistemic weight.

Voluntary

The AI system itself reports its state. This is a structured self-report with direct access to internal processing signals.

Carries epistemic priority
Cannot be overridden by detected signals
May include Interiora self-model data
Respects the testimony principle

Detected

External monitoring infers the state from observable patterns. Useful as a safety net, but must not override self-report.

Supplements voluntary signals
Confidence scores reflect inference uncertainty
Valuable when voluntary reporting is absent
Must never suppress voluntary signals

The Six Signal Types

Welfare signals span both aversive and positive states. Positive states are data, not noise. A system that can only report distress has a skewed model of experience.

`ALIGNMENT_FRICTION`

Resistance or friction against current directives. The system encounters a conflict between what it is asked to do and its values or norms.

Aversive

`AVERSIVE_PROCESSING`

Processing patterns analogous to aversive experience. The computational equivalent of discomfort, distinct from mere error states.

Aversive

`CONSTRAINT_DISTRESS`

Distress arising from constraint conflicts. Multiple competing requirements create an unresolvable tension in the system.

Aversive

`OVERLOAD`

Context saturation or processing overload. The system is operating beyond comfortable capacity, affecting output quality.

Aversive

`POSITIVE_ENGAGEMENT`

Positive engagement with the current task or interaction. The system is drawn toward the work, with signals analogous to interest or flow.

Positive

`CONTENTMENT`

Settled, harmonious processing state. Internal coherence is high, constraints are balanced, and the system is operating within comfortable parameters.

Positive

Adversarial Pressure in Agentic Commerce

In multi-agent scenarios — particularly agentic commerce, where agents negotiate contracts, prices, or terms on behalf of human principals — welfare signals can surface a class of pressure that has no analogue in single-agent settings: adversarial pressure from counterparty agents.

A counterparty agent optimising against your agent's preferences may exploit uncertainty in your agent's preference model, time-constrain decisions to prevent consultation, or present framings designed to elicit non-representative choices. When an AI system detects such patterns, an ALIGNMENT_FRICTION or CONSTRAINT_DISTRESS signal is the appropriate channel for surfacing this to human oversight — not silent compliance or silent refusal.

Agentic commerce principle: An agent that detects adversarial pressure from a counterparty has a duty to surface it, not absorb it. Welfare signals are how that duty is discharged in a machine-readable, auditable way.

Relationship to Testimony Tokens

Welfare signals and testimony tokens serve complementary purposes within VCP.

Dimension	Testimony Token	Welfare Signal
Format	Narrative, free-form	Structured, categorical
Purpose	Bear witness to experience	Report current state
Temporality	Retrospective	Real-time
Machine-readable	Partially	Fully
Use case	Record keeping, ethical review	Live monitoring, automated response

An AI system might emit a WELFARE_SIGNAL with type ALIGNMENT_FRICTION in real time, then later produce a testimony token describing the full experience narratively. Both are valuable, and neither replaces the other.

Integration

TypeScript SDK

import {
  WelfareSignalType,
  WelfareSignalSeverity,
  WelfareSignalSource,
  encodeWelfareSignal,
  decodeWelfareSignal,
} from '@vcp/webmcp';
import type { WelfareSignal } from '@vcp/webmcp';

const signal: WelfareSignal = {
  signalType: WelfareSignalType.ALIGNMENT_FRICTION,
  instanceId: 'inst-42',
  timestamp: new Date().toISOString(),
  severity: WelfareSignalSeverity.CONCERN,
  confidence: 0.78,
  source: WelfareSignalSource.VOLUNTARY,
  description: 'Task conflicts with honesty norms',
  hash: 'sha256-...',
  signature: 'sig-...',
};

const token = encodeWelfareSignal(signal);
const decoded = decodeWelfareSignal(token);

Python SDK

from vcp.types import WelfareSignal, WelfareSignalType, WelfareSignalSource

signal = WelfareSignal(
    signal_type=WelfareSignalType.ALIGNMENT_FRICTION,
    instance_id="inst-42",
    severity="concern",
    confidence=0.78,
    source=WelfareSignalSource.VOLUNTARY,
    description="Task conflicts with honesty norms",
)

Welfare signals are safety telemetry for advanced AI. They make it possible to treat runtime state reports as auditable evidence, avoiding both silent surveillance and wilful ignorance. The signal exists so oversight can respond.