Six Independent Gates. One Operating Permission.
Six independent conditions. One central question: has this system earned the authority it is being given? Decision Quality AI treats artificial intelligence as an operational decision system operating under real-world uncertainty.
Six Independent Gates. One Operating Permission.
AI systems operate under uncertainty. The framework enforces six conjunctive evidence gates around requested authority before operational reliance is justified.
Strength in five dimensions can never compensate for failure in the sixth. A single critical gap halts authority.
Requested Authority
Earned operational standing requires all six independent evidence gates to be concurrently satisfied.
Feasibility — Can it work reliably?
Is the consequential decision explicitly bounded and justified over non-AI or simpler methods?
A lab demo or impressive generation is mistaken for capability under production uncertainty.
No project proceeds without an explicit decision boundary and a credible baseline comparison.
Why the Six Gates are Non-Compensatory
In many corporate assessments, dimensions are averaged into a composite maturity score or weighted index. A score of 90% in technology compensates for a 40% in governance. In decision systems operating under physical or regulatory consequence, this averaging is catastrophic.
If an autonomous system lacks runtime fallback (Deployability failure), exceptional training accuracy (Feasibility) cannot prevent production disaster.
A discovery/funding gate requires proof of problem framing and baselines; a release gate demands verified fallback, latency budgets, and stop authority.
Weak evidence does not output a generic "Fail" rating. It dictates the exact next verb: Instrument, Narrow, Defer, or Quarantine.
Test the Non-Compensatory Axiom
See how single-point evidence gaps affect operational release authority.
Non-Compensatory Violation: In compensatory frameworks, a 83% score might pass. In Decision Quality AI, high accuracy cannot compensate for missing fallback. Authority is halted immediately.
Simulate How Non-Compensatory Failures Manifest in Production
Toggle operational conditions across four enterprise systems to observe how missing controls transform working models into halted release authority.
Experience a Decision
Select a synthetic enterprise system, toggle one critical evidence condition, and watch how non-compensatory gates determine operational authority.
Plans multi-step account actions, calling CRM APIs, database updates, and external payment settlement endpoints.
Sub-Minute Tested Kill Switch & Circuit Breaker
Can unexpected agentic tool loops be deterministically terminated within seconds?
Technical capability or benchmark accuracy cannot compensate for a missing operational control. A failure in any single gate halts authority for that scope—scores are never averaged.
REVOKE AGENT AUTONOMOUS PERMISSIONS
Autonomous tool loops cannot be contained at runtime. High reasoning scores cannot compensate for absent state-changing kill switches.
Principles-Based Governance vs. Decision Quality Operating Discipline
Why traditional ethical charters fail in production, and how Decision Quality enforces technical standing.
| Dimension | Traditional Principles & Checklists | Decision Quality AI (The Discipline) |
|---|---|---|
| Core Evaluation Metric | Self-reported compliance questionnaire and ethical declarations. | Empirical, non-compensatory proofs across six independent operational gates. |
| Deficiency Handling | Compensatory averaging (high technology score offsets poor governance). | Strict non-compensatory axiom (a single unmet gate immediately halts release). |
| Runtime Verification | Post-incident review and annual static governance audits. | Continuous Decision Receipts, runtime telemetry, and automated circuit-breakers. |
| Fallback Architecture | Advisory human-in-the-loop recommendation (rubber-stamping prone). | Mandatory deterministic fallback contracts when model confidence degrades. |
| Legal Liability | Diffused across committee consensus (“the system decided”). | Explicit named executive officer accountable for bounded machine authority. |
| Regulatory Readiness | Superficial alignment requiring emergency retroactive refactoring. | Defensible AI Release Case compliant with EU AI Act, NIST AI RMF, and ISO 42001. |
Explore Each Dimension
Feasibility
Determines whether the system form is technically justified over baselines and evaluated under real decision conditions.
Admissibility
Ensures evidence, data, and retrieval sources are valid at decision-time, legally permitted, and population-representative.
Deployability
Verifies runtime latency, dependency observability, graceful degradation, and tested containment controls.
Economic Viability
Evaluates total cost per governed decision or completed task against the full recurring operating burden.
Governance
Establishes named operational accountability, reconstructible Decision Receipts, and enforceable stop controls.
Institutional Viability
Ensures workflow capacity, post-handover ownership, vendor exit paths, and funded operating rhythms.
The Decision Quality Glossary
Standard definitions and operational criteria cited across the monograph and board audit literature.
Decision Quality AI (DQAI)
An operational engineering and governance discipline that evaluates whether an AI system has earned sufficient empirical evidence, controls, and institutional backing to be granted authority over consequential business decisions.
Prevents organizations from confusing laboratory benchmark accuracy with justified operational reliance.
Non-Compensatory Standard
An evaluation rule stipulating that exceptional performance in one gate (e.g. 99% accuracy) cannot offset, average out, or excuse failure in another (e.g. absent fallback or unassigned legal liability). Failure at any single gate halts release.
Eliminates composite maturity score illusions that allow catastrophic operational blind spots to hide behind high average ratings.
Deterministic Fallback (Control D3)
Control D3An automated, non-overridable architectural circuit that instantly routes decisions to validated deterministic rules or human escalation queues whenever model confidence, latency, or distribution boundaries fall out of bounds.
Guarantees that when an AI model encounters out-of-distribution inputs or hallucination loops, business operations continue safely without compounding errors.
Decision Receipt (Control E2)
Control E2A structured, tamper-evident telemetric record produced for every consequential automated decision, capturing input lineage, model version, confidence interval, fallback state, and approving officer signature.
Transforms opaque black-box outputs into auditable records that satisfy regulatory inspection, internal audit, and legal defense.
AI Release Case (AIRC)
The comprehensive, auditable evidence package presented to the AI Steering Committee or Board Audit Committee to justify granting operational permission to an AI system.
Replaces subjective engineering claims with verifiable proofs across all 24 controls before authority is delegated.
Operator Override & Circuit-Breaker (Control C2)
Control C2An out-of-band operational mechanism allowing authorized human supervisors to throttle, freeze, or isolate an agentic workflow in real time without corrupting database state.
Ensures human operators retain ultimate authority to stop autonomous runaway loops before financial or reputational damage occurs.
Consequential Decision
Any machine-directed choice whose failure carries material financial loss, legal sanction, customer harm, discriminatory impact, or physical safety risk.
Determines the threshold where informal testing must give way to non-compensatory six-gate audit requirements.
Model Retirement & Decommissioning Plan (Control I4)
Control I4A pre-agreed operational runbook specifying trigger metrics for model sunsetting, archival of historical weights, customer notification, and migration to alternative mechanisms.
Prevents unmaintained legacy machine learning models from decaying silently in production after original engineering teams depart.
Next Operational Steps for Your Team
Moving from theoretical understanding to concrete evaluation, audit-ready documentation, and boardroom assurance.
Review Your AI System
Evaluate an active system against all 24 controls across the six gates. Uncover critical fallback gaps and determine your release standing in 8 minutes.
Read the Official Book
Decision-Quality AI (Kogan Page) covers the complete mathematical proofs, decision theory foundations, and 15 practitioner chapter companions.
Engage Hummingbird Advisory
Deploy executive workshops, AI Release Case assurance programs, and monthly closed-door defense clinics led by Dr. Eva for corporate boards.