Research & Publications

CAP and PACELC never addressed AI systems that decide outcomes for people. CAAP argues Accountability belongs in the same architectural class as Consistency, Availability, and Partition Tolerance — with four trade-off configurations, testable propositions, implementation patterns, and alignment to the EU AI Act, NIST AI RMF, and India's DPDP Act.

• •Accountability is a first-class architectural constraint — not an ethics overlay that can be bolted on after launch.
• •CAAP formalises four trade-off configurations and two testable architectural propositions for high-stakes AI.

As AI takes on screening, performance scoring, and workforce planning, enterprises lack a clear rule for what machines may decide alone. This working paper introduces AI-MDAF — four authority tiers, seven weighted criteria, a scoring model, and a maturity assessment — plus the Named Human Principle for accountable managerial AI.

• •AI-MDAF defines four authority tiers: automate, advise, co-decide, and human-only — mapped across the employee lifecycle.
• •Seven weighted evaluation criteria and a quantitative scoring model help teams decide which managerial decisions can be automated.

Senior-living operators face staffing pressure and rising resident expectations. Agentic workflows can handle routine requests — dining, maintenance, activities — but vulnerable populations demand stricter trust boundaries than typical B2B SaaS. This brief covers voice-agent architecture, mandatory human escalation for health intents, and audit patterns from OEAT pilot deployments.

• •Routine vs. sensitive intent routing must be explicit in architecture, not prompt-only.
• •Voice adds latency and accessibility benefits but increases mis-hear risk — always offer human fallback.

Production AI agents fail in predictable ways — not because models are weak, but because orchestration, tools, and guardrails are under-designed. This white paper documents six failure modes observed across HRTech and HealthTech deployments, and proposes a three-layer resilience model: input guardrails, runtime circuit breakers, and post-hoc evaluation replay.

• •Six recurring failure modes: tool timeout cascades, silent hallucination, context bleed, wrong-entity merges, prompt injection, and drift without detection.
• •A three-layer model — guardrails, runtime checks, eval replay — reduced incidents 40% in a live ATS agent deployment.

Closed consultant networks cannot rely on open marketplace search. Talent pools flatten relationships; talent graphs preserve context — who worked with whom, on what skills, under which constraints. This publication compares SQL search vs. graph traversal for niche staffing requests and outlines privacy-preserving integration patterns for HR tech platforms.

• •Linear talent pools break down for multi-skill, relationship-aware searches.
• •Graph models encode engagements, referrals, and skill adjacency without exposing raw PII.

AI-powered hiring tools can amplify historical bias unless fairness is engineered into the release process. This paper outlines evaluation frameworks for candidate ranking — demographic parity checks, explainability requirements, and recruiter override loops — drawn from production ATS work at Noble House Consulting.

• •Bias mitigation is a release gate, not a one-time audit.
• •Explainable ranking signals improve recruiter trust and catch drift early.