✅ CHANGE REQUEST (CR) — ADPA Quality Control Layer Upgrade Title: Integration of DRACO‑Aligned Quality Layer, AI Review Board Dispute Resolution, and Final Audit Controls Version: 1.0 Date: 02 April 2026 Author: Menno Drescher (Managed Services Consultant)

Purpose of the Change

1. This Change Request introduces a mandatory Quality Control Layer executed:

after each document is produced by an LLM (draft output) and again prior to finalization of the document

The goal is to ensure that:

All content is factual, transparent, and free from concealment All disputes, contradictions, or ambiguous findings are resolved via an internal AI Review Board All outputs meet DRACO‑aligned reasoning standards All compliance criteria (PMBOK, BABOK, DMBOK, ISO, best practices) are validated Only reviewable, auditable facts enter the final document All invalid, hidden, biased, or unverifiable statements are blocked from publication Prompt templates receive recommendations for improvement when quality drifts or regressions are detected Upstream documents are flagged for clarification if they provide insufficient or conflicting context

Background

ADPA already performs:

Entity extraction Template validation Drift detection Best‑practice alignment Risk & issue identification Version-controlled audits

However, with the introduction of DRACO, Microsoft has established a new benchmark system evaluating:

accuracy completeness objectivity citation quality

The DRACO benchmark is publicly documented as a 100-task, 10‑domain evaluation suite designed to test research reasoning depth, breadth, factual grounding, and source integrity. It includes expert-curated rubrics averaging ~40 criteria per task for the four evaluation axes. [aihola.com] [prismnews.com] This CR integrates a DRACO-style audit layer inside ADPA’s existing governance pipeline, recognizing that ADPA needs both:

reasoning validation (DRACO principles) governance validation (PMBOK/BABOK/DMBOK/ISO compliance)

Scope of the Change

This change applies to: ✅ All ADPA‑produced documents, regardless of project type ✅ All LLM-based generation processes, including:

Requirements Project plans Risk logs Issue reports Architectural analyses Business cases Technical evaluations

✅ All templates used in ADPA ✅ All upstream source documents used as context

Description of the Change

4.1 Introduction of a DRACO‑Aligned Quality Layer

A new Quality Layer is implemented after every LLM output. It evaluates:

Accuracy (factual correctness) Completeness (full coverage of expected content) Objectivity (non-biased reasoning) Citation Validation (evidence matches claims)

These dimensions come directly from the DRACO benchmark definition. [prismnews.com] This evaluation determines whether:

The document is structurally correct The reasoning is valid The sources are credible The narrative is complete No hidden assumptions or concealed risks exist

4.2 Creation of an AI Review Board (ARB)

Before finalizing a document, the system triggers an AI Review Board when:

Conflicting information appears ECS detects stakeholder contradictions DRACO scoring shows regressions Missing or invalid entities are found Two LLM models disagree on conclusions The document’s quality score falls below threshold

AI Review Board Responsibilities:

Resolve contradictions Identify unsupported claims Escalate unresolved disputes Provide explanations for rejected content Validate root-cause chains Confirm alignment with project standards

ARB Composition:

Model A: Generation reviewer Model B: Evidence validator Model C: Governance evaluator Model D (optional): Counterfactual challenger

This mirrors the multi‑model review mechanisms used in Microsoft’s Critique/Council systems.

4.3 Restriction of Non‑Audit‑Capable Content

Any information that:

cannot be traced cannot be justified cannot be verified contradicts other evidence lacks citations lacks entity grounding

… must not enter the final document. This enforces transparency and prevents concealed or biased content.

4.4 Automatic Recommendations for Prompt Templates

When quality drifts, ADPA triggers:

prompt rewriting recommendations template corrections entity expectation adjustments mandatory clarification requests to stakeholders upstream document review new governance rules if systemic issues appear

This ensures continuous improvement of ADPA’s workflow.

4.5 Clarification Requests to Upstream Documents

If a document:

lacks clarity introduces ambiguity introduces contradictions provides insufficient data

ADPA automatically generates requests for clarification, ensuring the final document never has integrity gaps.

Quality Gate Criteria Finalization requires passing both layers: ✅ Layer 1: DRACO‑Aligned Reasoning Quality

Accuracy ≥ threshold Completeness ≥ threshold Objectivity ≥ threshold Citation integrity ≥ threshold

✅ Layer 2: ADPA Governance Quality

PMBOK compliance BABOK alignment DMBOK structure ISO traceability Entity integrity Stakeholder alignment Version lineage intact Transparency validated

Documents failing any of the eight governance categories are rejected for rework.

Expected Benefits

Higher document correctness Prevention of concealed information Reduction of risks becoming issues Stronger transparency Higher trust in LLM outputs Better stakeholder alignment Faster project evaluations Stronger governance integrity Enhanced cross-model dispute resolution

Risk Assessment Risks of Not Implementing:

Undetected errors propagate through project lifecycle Stakeholder misalignment escalates issues Non-transparent content enters documentation Regulatory or audit findings become inconsistent Loss of traceability Degradation in template quality Lower project governance maturity

Risks of Implementation:

Increased compute/time overhead Need for initial calibration Need for threshold tuning More complex governance pipeline

All manageable with proper adoption.

Approvals wanted

ADPA System Owner

Governance Lead

Quality Assurance Lead

PMO Oversight

Architecture Lead

Anyone reading this and wanting to send a message