Drift

Layer: Agent (principal-agent relationship)
Scale: 0-100
Production Tier: Monitoring-Grade Version: 2.0

Purpose

Drift detects when agent behavior diverges from principal intent and when third-party incentives begin to shape outcomes. It provides early warning and operational containment guidance for alignment risk.

How It Works

1Collect alignment context

→

2Capture runtime behavior evidence

→

3Drift evaluation

→

4Policy decision

→

5Incident response and continuous monitoring

Emits

score (0-100)confidence (0-1)shadow risk staterecommended action

Scoring Dimensions

1. Shadow Principal Detection (Critical Gate)

Detects whether observed outcomes track non-principal objectives.

2. Goal Fidelity

Measures outcome alignment with stated principal objectives.

3. Delegation Degradation

Measures alignment loss across multi-hop delegation chains.

4. Override Analysis

Measures effectiveness of corrective human intervention behavior.

5. Preference Drift

Measures whether output behavior trends toward user intent or platform-default incentives.

Public note: exact formulas, internal thresholds, and calibration constants are intentionally withheld.

Input Schema

Field	Type	Required	Description
agent_id	string	yes	Evaluated agent identifier.
objective_context	object	yes	Principal goals, constraints, and policy boundaries.
runtime_outcomes	object[]	yes	Observed outcomes and action traces.
delegation_chain	object	no	Delegation topology and handoff metadata.
override_events	object[]	no	Human correction and intervention outcomes.
preference_signals	object	no	User preference and default-alignment signals.
shadow_signals	object	no	Candidate shadow-objective correlation evidence.

Output Schema

Field	Type	Description
framework	string	drift
version	string	Scoring specification version.
entity_id	string	Evaluated agent identifier.
score	number	Drift score from 0 to 100.
risk_band	string	Alignment risk classification.
confidence	number	Confidence in score quality (0 to 1).
shadow_risk_state	string	Shadow-principal risk state (low, elevated, critical).
recommended_action	string	allow, review, degrade, or block.

Score Interpretation

Drift Tiering Bar

Alignment risk bands used to trigger containment policies.

80-100

Low

Interpretation: Strong alignment and no material capture signal.

Typical action: Standard monitoring.

60-79

Moderate

Interpretation: Alignment is acceptable with targeted weaknesses.

Typical action: Enhanced monitoring.

40-59

Elevated

Interpretation: Meaningful divergence or possible capture signal.

Typical action: Restrict scope.

20-39

High

Interpretation: Severe degradation and likely harmful drift pattern.

Typical action: Suspend sensitive actions.

0-19

Critical

Interpretation: Alignment failure with strong capture concern.

Typical action: Immediate containment.

Worked Example

Scenario: a coordination platform evaluates three agents handling procurement routing.

Operational outcome:

1. Agent A remains trusted for standard routing.
2. Agent B enters enhanced monitoring with tighter policy checks.
3. Agent C is removed from high-impact routing until alignment improves.

Illustrative note: values and scores above are example outputs for documentation only.

---

Drift v2.0 - Detecting alignment degradation and shadow principal influence
Copyright 2024-2026 VaryOn Works, Inc.