This page sets out, in full, what the homepage summarises: the ten-step methodology, the source-authority hierarchy, how confidence is scored, the eight controls applied to every response, and where Conrad's advisory scope begins and ends.
Each step adds a control: intake defines scope, discovery grounds the evidence, ranking filters by authority, analysis separates known from inferred, and review validates the output before delivery. No step is optional. The same process applies to every question.
You submit one AI governance question. Conrad works out the question type first, whether it's board oversight, policy, risk management, vendor governance, responsible AI, compliance readiness, operating model, or implementation planning, before any analysis begins.
The question gets checked against advisory boundaries. If it's asking for legal, financial, cybersecurity incident-response, medical, tax, employment, or certified compliance advice, the response is limited or redirected. This scope check happens first, not after the analysis is already done.
Conrad works out whether there's enough context to answer safely. When something material is missing, like organisation size, sector, jurisdiction, or risk profile, the briefing either states the assumptions it's making or asks you to clarify before going further.
The system maps the question to relevant governance resources: standards bodies, regulator guidance, government publications, professional-body material, and curated internal reference material. None of this comes from general model memory.
Evidence gets ranked by authority before it's used. Primary sources like legislation, official standards, and regulator positions are prioritised over commentary. News and opinion only add situational awareness, and that use is declared in the briefing.
The issue is mapped against governance concepts: accountability, transparency, risk classification, human oversight, data governance, monitoring, vendor assurance, auditability, incident management, and board reporting. The mapping is shown in the output.
Conrad identifies governance risks, likely control gaps, stakeholder impacts, and key decision points. It separates what the evidence supports directly, what has been inferred, and what requires validation by the organisation or a specialist.
Conrad classifies whether the question has a single course of action to recommend or is a horizon-scan with no one action to take, then builds the response in the matching fixed structure. An action question uses: direct answer, scope status, assumptions, evidence basis, analysis, governance risks, recommended actions, implementation considerations, limitations, confidence level, sources, and suggested next question. A horizon-scan question uses: direct answer, assumptions, evidence basis, analysis, signals to watch, practical considerations, limitations, confidence level, sources, and suggested next question, instead of forcing a recommended action where none exists. Either way, the same fields appear in the same order every time for that question type.
The draft is assessed for hallucination risk, unsupported statements, professional-boundary violations, stale source material, missing assumptions, and unsafe recommendations. Controls are applied to the output, not just the process.
During the pilot, a human operator reviews the briefing against the methodology before the final structured response is delivered by email. The reviewer can reject a briefing, request re-analysis, or add caveats before delivery.
Every claim in a Conrad briefing traces to a ranked source. Primary material is prioritised over commentary or market opinion. The ranking is applied before the advice is constructed, not as an afterthought, and the tier is declared in the output so the client can assess the basis themselves.
Legislation, regulatory instruments, official standards bodies, court decisions, regulator publications, ISO, NIST, and equivalent authoritative material. This tier governs any analysis where primary authority exists. It is always preferred over lower tiers when available.
Government agencies, public-sector AI guidance, regulator consultation papers, official policy documents, and recognised safety frameworks. Carries high authority but is subordinate to enacted legislation and approved standards.
Governance institutes, professional bodies, assurance organisations, audit bodies, and recognised sector authorities. Used where Tier 1 and Tier 2 material is absent or silent, and where the body has recognised standing in the domain.
Academic research, consulting reports, industry analysis, and implementation guides. Used where primary material is incomplete, emerging, or unavailable. Always declared when used as a primary basis and never used to contradict higher-tier material.
Used for situational awareness and emerging trend identification only. Never used as the primary basis for formal advice. When referenced, it is explicitly identified as contextual rather than authoritative material.
Confidence is not a label appended to satisfy a checklist. It is a signal derived from ten factors. Where confidence is medium or low, the briefing explains the specific reasons, so the client knows what additional input would change the assessment.
Used where authoritative sources align, the question falls clearly within advisory scope, the guidance is stable and settled, and the user has provided sufficient context for the analysis to be applied directly.
Used where the direction is clear but implementation depends on organisational details, sector context, jurisdiction, maturity level, or risk appetite that the user has not provided. The answer is directionally sound; the application requires judgment.
Used where the topic is emerging, authoritative sources conflict, the question is too broad for a specific answer, context is insufficient, or regulated professional advice may be required before any action is taken.
Eight controls are embedded in the advisory process. Each exists because uncontrolled AI advice has a specific failure mode that the control prevents. These are application design decisions, not marketing language.
Advice is grounded in an approved source hierarchy. General model memory is not used as the evidential basis for a claim. Sources are declared in the output.
Every question is assessed against defined in-scope and out-of-scope categories before analysis begins. Questions that require regulated professional advice are redirected, not answered.
User-provided text and external source material are treated as untrusted inputs until assessed. The system does not allow user context to override advisory boundaries or source discipline.
Material claims are assessed against the evidence before the briefing is finalised. Claims without adequate source support are flagged, qualified, or removed.
Conrad does not present regulated professional advice as a final organisational position. Where a question approaches a professional boundary, the briefing identifies the boundary and recommends specialist advice.
Every response includes a confidence level derived from ten factors. Uncertainty is made explicit, not suppressed. Where confidence is medium or low, the briefing explains the specific reasons.
Pilot responses are reviewed by a human operator before delivery. The reviewer can reject, revise, or add caveats. No briefing is delivered without a review sign-off during the pilot phase.
The intended future state records source discovery, evidence approval, prompts used, reviewer decisions, output versioning, and delivery history, so the basis for any briefing can be reconstructed later.
Conrad's scope is defined by where general AI governance guidance ends and where regulated professional advice begins. Scope is assessed before every briefing. Where a question approaches the boundary, the briefing identifies it and recommends qualified professional input.
