AEGIS Migration Guide

Version: 1.0.0 | Updated: 2026-02-09 | Status: Active

This guide covers upgrading AEGIS versions, migrating parameter schemas, recalibrating thresholds, and handling workflow state migrations.

1. Parameter Schema Migration

Schema Versioning

schema/interface-contract.yaml is the authoritative source for all frozen parameters. The schema version is tracked in the file's version field.

Migration Procedure

1. Compare schemas: Diff old and new interface-contract.yaml

   git diff guardrail-v1.1-freeze..HEAD -- schema/interface-contract.yaml

2. Identify changes: New parameters get defaults automatically. Modified parameters require review.

3. Validate config: Test your custom config against the new schema

   aegis validate-config <your-config.yaml>

4. Rollback plan: Keep a tagged backup of the previous frozen state

   git tag guardrail-v{X.Y}-freeze -m "Freeze guardrail params at v{X.Y}"

Parameter Change Categories

2. Parameter Recalibration

When to Recalibrate

• KL divergence drift detection triggers sustained warnings (tau_warning exceeded)
• Business requirements change (risk tolerance shift)
• Model performance degrades over time
• Calibrator actor proposes threshold adjustments

Recalibration via Calibrator Actor

The Calibrator actor (src/actors/calibrator.py) provides a structured workflow for threshold tuning:

from actors.calibrator import Calibrator
from engine.drift import DriftMonitor
from engine.bayesian import BayesianPosterior

calibrator = Calibrator(
    actor_id="calibrator-1",
    drift_monitor=drift_monitor,
    bayesian_posterior=posterior,
)

# 1. Propose new thresholds based on observed data
proposal = calibrator.propose_calibration(
    parameter="tau_warning",
    new_value=0.35,
    justification="Sustained drift above current threshold",
)

# 2. Approval gate (requires human approval)
calibrator.approve(proposal.proposal_id, approver_id="risk-lead")

# 3. Apply calibrated thresholds
calibrator.apply_calibration(proposal.proposal_id)

Recognized Parameters

The Calibrator accepts changes to these 16 parameters (whitelist in _RECOGNIZED_PARAMETERS):

Safety Constraints

• Calibration proposals require APPROVED status before application
• The _RECOGNIZED_PARAMETERS whitelist prevents modification of unrecognized parameters
• Thread-safe via threading.Lock
• All calibration events emit telemetry for audit trail

3. Workflow State Migration

Persistence Schema

AEGIS workflow state is stored via SQLAlchemy ORM models (src/workflows/persistence/models.py):

Schema version: 1.0.0

Upgrade Procedure

Current approach:

1. Export existing workflow data:

   from workflows.persistence.repository import WorkflowPersistence
   workflows = await repo.list_workflows(status="pending")

2. Apply schema changes via SQL migration scripts

3. Re-import workflows with updated schema

Roadmap: Alembic migration framework is planned for automatic schema versioning (alembic upgrade head). Until then, use the manual procedure above.

Rollback Procedure

If a migration fails or the new version exhibits issues:

1. Serialization backup (before migration):

   import json
   from workflows.persistence.repository import WorkflowPersistence
   
   workflows = await repo.list_workflows()
   backup = [w.to_dict() for w in workflows]
   with open("workflow-backup.json", "w") as f:
       json.dump(backup, f)

2. Restore from backup:

   for data in backup:
       workflow = ProposalWorkflow.from_dict(data)
       await repo.save_checkpoint(workflow, actor_id="migration-rollback")

3. Database point-in-time recovery: If using PostgreSQL with continuous archiving, restore to a point before the migration:

   # AWS RDS example
   aws rds restore-db-instance-to-point-in-time \
     --source-db-instance-identifier aegis-db \
     --target-db-instance-identifier aegis-db-restored \
     --restore-time "2026-03-05T12:00:00Z"

4. Redeploy previous version: Roll back the container image or Lambda function to the pre-migration tag:

   git checkout guardrail-v{PREVIOUS}-freeze

Serialization

All workflows support to_dict() / from_dict() serialization:

# Export
data = workflow.to_dict()

# Import (handles timezone normalization via ensure_utc)
workflow = ProposalWorkflow.from_dict(data)

The shared ensure_utc() helper (src/workflows/serialization.py) normalizes timestamps across all three workflow types (proposal, consensus, override).

Data Safety

• PII fields: 12 telemetry fields encrypted at rest via PIIEncryptionEnricher
• Audit trail: Hash-chained transitions (SHA-256) for tamper detection
• Encrypted fields: EncryptedField.value_type discriminator ("str", "bytes", "json") for lossless round-trips

4. Configuration Upgrade Path

AegisConfig Properties

AegisConfig is a frozen dataclass — instances are immutable after creation. This prevents parameter drift during a session.

from aegis_governance import AegisConfig

# Immutable: raises AttributeError
config = AegisConfig.default()
config.epsilon_R = 0.5  # AttributeError: cannot assign to field

Backward Compatibility

Configuration Loading

AegisConfig.from_yaml() reads the parameters key with nested sections:

# Top-level parameters
parameters:
  epsilon_R: 0.01
  # Nested section
  novelty_gate:
    N0: 0.7
    k: 10.0
    output_threshold: 0.6
  # Nested section
  kl_drift:
    tau_warning: 0.3
    tau_critical: 0.5
    window_days: 30

Nested sections are mapped to flat AegisConfig fields via _NESTED_SECTIONS in src/config.py.

5. Version Compatibility Matrix

AEGIS Version Matrix

Dependency Compatibility

6. Pre-Migration Checklist

• Backup current config: cp schema/interface-contract.yaml schema/interface-contract.yaml.bak

• Tag current state: git tag guardrail-v{X.Y}-freeze

• Run test suite against new version: pytest tests/ -v --tb=short

• Validate config against new schema: aegis validate-config <config.yaml>

• Test in staging: Deploy new version to staging environment first

• Verify audit trail continuity: Ensure hash chain remains valid after migration

  from workflows.persistence.durable import DurableWorkflowEngine
  # Verify hash chain integrity
  is_valid, message = engine.verify_audit_chain(workflow_id)
  assert is_valid, f"Audit chain broken: {message}"

• Review ROADMAP.md: Check for breaking changes in the target version

• Update monitoring: Ensure Prometheus rules match new metric names (if changed)

References

• Interface Contract — Frozen parameters (authoritative)
• Calibrator Actor — Threshold recalibration
• Persistence Models — ORM schema
• Config Module — AegisConfig, from_yaml(), nested sections
• Serialization — ensure_utc() shared helper
• Production Guide — Deployment instructions
• Performance SLAs — Latency and throughput targets