AEGISdocs
Resources

Guardrail Framework Governance

Governance procedures for the Hardened Quantitative Guardrail Framework including approval workflows, recalibration processes, and dual-control requirements.

Version: 1.0.0 Effective Date: 2025-12-26 Owner: Risk Team Review Cycle: Quarterly


Overview

This document establishes governance procedures for the Hardened Quantitative Guardrail Framework. It defines approval workflows, recalibration processes, and dual-control requirements to ensure consistent, auditable risk management.


Recalibration Checklist

Model recalibration MUST follow this checklist to ensure parameter changes are validated, tested, and properly governed before deployment.

Prerequisites

RequirementOwnerEvidence
Minimum 90 days of production data collectedData EngineeringTelemetry export
No active incidents involving guardrail systemOperationsIncident dashboard
Previous quarter's calibration completeRisk LeadGit tag exists

Phase 1: Data Validation

  • 1.1 Export holdout data from telemetry system (minimum 1,000 proposals)
  • 1.2 Verify holdout data was not used in current model training
  • 1.3 Confirm data distribution matches production (KL divergence < 0.5)
  • 1.4 Document data period: ____-__-__ to ____-__-__
  • 1.5 Calculate baseline metrics on holdout:
    • False Positive Rate: ____%
    • False Negative Rate: ____%
    • Decision Accuracy: ____%

Sign-off: Data Engineering Lead Date: ____-__-__


Phase 2: Parameter Optimization

  • 2.1 Run calibration script on holdout data (see Implementation Plan §Phase 4)
  • 2.2 Document proposed parameter changes:
ParameterCurrent ValueProposed ValueJustification
epsilon_R
epsilon_P
risk_trigger_factor
profit_trigger_factor
trigger_confidence_prob
novelty_gate.N0
novelty_gate.k
novelty_gate.output_threshold
complexity_floor
quality_min_score
  • 2.3 Back-test proposed parameters on historical data
  • 2.4 Calculate performance delta:
    • FPR change: ____%
    • FNR change: ____%
    • Net improvement: ____%

Sign-off: ML Engineering Lead Date: ____-__-__


Phase 3: Shadow Testing (Phase 1 Tests)

  • 3.1 Deploy proposed parameters to shadow environment
  • 3.2 Run parallel scoring for minimum 7 days
  • 3.3 Compare shadow vs production decisions:
    • Agreement rate: ____%
    • Divergence cases reviewed: ___ / ___
  • 3.4 Document any edge cases where new parameters perform worse
  • 3.5 Confirm no regression in guardrail catch rate

Sign-off: QA Lead Date: ____-__-__


Phase 4: Red-Team Validation (Phase 2 Tests)

  • 4.1 Run Monte Carlo fuzzing suite (minimum 10,000 synthetic proposals)
  • 4.2 Execute boundary condition tests:
    • Near-threshold proposals (±5% of trigger)
    • Zero-baseline edge cases
    • Maximum complexity proposals
    • Novelty boundary (N = 0.69-0.71)
  • 4.3 Verify no new exploit vectors introduced
  • 4.4 Confirm adversarial resistance maintained

Sign-off: Security Engineering Lead Date: ____-__-__


Phase 5: Dual-Control Governance Approval

CRITICAL: This phase implements dual-control governance requiring TWO independent risk approvers.

Primary Risk Approval

Risk Lead Review:

  • 5.1 Reviewed all phase sign-offs above
  • 5.2 Verified holdout data independence
  • 5.3 Confirmed no degradation in protection
  • 5.4 Assessed business impact of parameter changes
  • 5.5 Approved proposed parameters
ApproverRoleSignatureDate
Risk Lead

Secondary Risk Approval (Dual-Control)

Second Risk Approver Review:

The second approver MUST be independent from the primary approver and MUST NOT have participated in the calibration process.

  • 5.6 Independently reviewed calibration methodology
  • 5.7 Verified dual-control separation of duties
  • 5.8 Confirmed no conflicts of interest
  • 5.9 Validated test coverage sufficiency
  • 5.10 Approved proposed parameters
ApproverRoleSignatureDate
Second Risk Approver

Dual-Control Verification:

  • Both approvers are from the Risk team or equivalent authority
  • Approvers have no reporting relationship to each other
  • Minimum 24-hour separation between approvals (cooling-off period)

Phase 6: Deployment

  • 6.1 Create new parameter file: guardrail_params_v_._.json
  • 6.2 Open Pull Request with:
    • Parameter file changes
    • Updated Interface Contract (if schema changes)
    • This completed checklist as PR description
  • 6.3 PR approved by CODEOWNERS
  • 6.4 Merge to main branch
  • 6.5 Create Git tag: guardrail-v_._-freeze
    git tag -a guardrail-v1.2-freeze -m "Freeze guardrail params at v1.2 (Q_ 202_)"
    git push origin guardrail-v1.2-freeze
  • 6.6 Verify production deployment of new parameters
  • 6.7 Confirm telemetry logging new param_snapshot_id

Sign-off: DevOps Lead Date: ____-__-__


Phase 7: Stakeholder Communication

  • 7.1 Notify stakeholders of parameter update:
    • Risk Committee
    • Engineering Leadership
    • Operations Team
    • Compliance (if applicable)
  • 7.2 Send communication including:
    • Summary of changes
    • Effective date
    • Expected impact
    • Rollback procedure reference
  • 7.3 Update internal documentation
  • 7.4 Schedule next quarterly calibration review

Communication Template:

Subject: Guardrail Parameters Updated to v_._ (Q_ 202_)

Team,

The guardrail framework parameters have been recalibrated and deployed.

Key Changes:
- [List parameter changes]

Effective: [Date]
Git Tag: guardrail-v_._-freeze

Performance Impact:
- False Positive Rate: [X%] → [Y%]
- False Negative Rate: [X%] → [Y%]

Rollback: If issues arise, revert to guardrail-v_._-freeze per IRP.

Questions? Contact Risk Team.

Sign-off: Communications Lead Date: ____-__-__


Governance Roles

RoleResponsibilityAuthority
Risk LeadPrimary recalibration approvalAPPROVE/REJECT parameters
Second Risk ApproverDual-control validationAPPROVE/REJECT (independent)
ML Engineering LeadCalibration executionRECOMMEND parameters
Data Engineering LeadData validationCERTIFY data quality
QA LeadShadow testingCERTIFY test results
Security Engineering LeadRed-team validationCERTIFY security
DevOps LeadDeployment executionEXECUTE deployment

Escalation Path

If any phase fails or concerns arise:

  1. Phase 1-4 Failure: Return to beginning of failed phase after remediation
  2. Phase 5 Rejection: Escalate to Risk Committee with documented concerns
  3. Phase 6-7 Issues: Invoke rollback procedure per Incident Response Plan

Audit Trail Requirements

All recalibration events MUST be logged:

FieldDescription
recalibration_idUnique identifier (UUID)
initiated_dateISO-8601 start date
completed_dateISO-8601 completion date
previous_versione.g., guardrail-v1.1-freeze
new_versione.g., guardrail-v1.2-freeze
primary_approverRisk Lead identifier
secondary_approverSecond Risk Approver identifier
data_period_startHoldout data start date
data_period_endHoldout data end date
pr_numberGitHub PR number
commit_shaMerge commit SHA

References


Changelog

VersionDateAuthorChanges
1.0.02025-12-26Claude CodeInitial creation (GAP Issue #4)

On this page