The vairee Research Framework
Why LLM-based hiring systems fail — and the architecture of what replaces them.
The professional hiring system has a systematic bias that nobody designed but everyone perpetuates.
Applicant Tracking Systems layered with generative AI do not improve hiring quality. Research conducted within the vairee project documents a specific failure mode: the tendency of LLM-based screening systems to normalise candidates toward statistical averages, systematically down-ranking genuine innovators and senior experts whose profiles deviate from dominant training patterns.
We call this phenomenon Algorithmische Meinungsgleichschaltung — algorithmic opinion alignment toward the mean.
This document summarises the research architecture, core findings, and the experimental framework designed to validate the vairee hypothesis.
The core problem — flat vector matching in a complex world
Current ATS and talent intelligence systems rely on vector-based semantic similarity: they compare the text of a CV against the text of a job description, rank by frequency and semantic proximity of relevant terms, and surface candidates whose documents most closely match.
This approach has four documented failure modes:
1. Innovator penalisation
A professional who has led genuinely ahead-of-market work — distributed systems before they were mainstream, organisational models before the playbooks existed — does not appear in training data as a dominant pattern. Their profile is classified as a statistical anomaly and ranked low. The system mistakes novelty for irrelevance.
2. The AI-optimised CV problem
Candidates who use generative tools to reverse-engineer job description keywords into their profiles score well against vector matching — regardless of whether their underlying experience is real. Junior professionals with AI-polished CVs consistently outrank senior experts with direct but unstylised profiles. The system rewards presentation, not capability.
3. Certification noise
The hiring market has created a feedback loop: candidates acquire certifications (SAFe, ICAgile, PMP) that signal conformance to screener criteria, whether or not those certifications represent genuine expertise. ATS systems treat certification presence as a positive signal. The result is systematic inflation of false positives at the expense of real practitioners.
4. Temporal flattening
Vector similarity operates atemporally. A senior executive who completed an MBA with financial management focus twenty years ago may be ranked against — or below — a junior financial management course, because the system detects keyword overlap without understanding seniority, chronological progression, or contextual decay.
The vairee hypothesis
The vairee research hypothesis:
The deployment of a hybrid neuro-symbolic model — combining a hidden semantic baseline, interactive cognitive mining via a Heuristic Miner, and a dual-epoch verification architecture — will demonstrate statistically significant improvement in identifying non-standard, highly senior, and innovative talent compared to traditional ATS systems built on flat vector pattern matching.
The executive target parameter: 85% relative improvement in precision and recall for non-conformant talent profiles. Conservative minimum threshold for academic validation: 50% relative improvement over commercial ATS benchmarks.
The architecture — three layers replacing one
The vairee matching architecture replaces flat document comparison with a three-layer system:
Layer 1 — Hidden semantic baseline
Before any interaction with the candidate, the system constructs an initial knowledge graph from publicly available digital traces (publications, technical contributions, public project artefacts). This baseline is not shown to the candidate. If shown, it would trigger optimisation behaviour — the candidate would adjust their presentation to match the system's current model. The hidden baseline provides an independent anchor against which candidate claims can be cross-referenced.
Layer 2 — Interactive cognitive mining (Heuristic Miner)
Rather than asking for skills lists or competency tables — the approach that research identifies as the root cause of the false-positive problem — the Heuristic Miner elicits implicit knowledge through questions oriented toward failure, crisis management, technical trade-offs, and historical context.
Experience cannot be simulated. The density and specificity of a professional's response to a question about how a distributed system failed in a specific architectural context at a specific scale — at a specific point in time — is not reproducible without having been there. The Heuristic Miner identifies experience through the texture of its description, not through its label.
Layer 3 — Cryptographic-semantic triangulation
When a candidate makes a verifiable claim, the system deploys a three-node verification protocol:
Cryptographic node: Decentralised identity anchors and B2B Soulbound tokens verify who worked where, when, and in what role. For post-2026 experience, this provides mathematically anchored verification.
Semantic node: Autonomous GraphRAG scanning cross-references claims against independent technical artefacts — GitHub activity, published work, technical contributions from the relevant period. Temporal synergy and dissonance are detected.
Cognitive stress-test node: For experience that is inherently unverifiable (projects under NDA, proprietary systems, classified work), the system deploys targeted questions about technical micro-details — the specific failure modes of a technology at a specific time, the implementation constraints of a system at a specific scale. These details cannot be fabricated. Their specificity and accuracy reveal whether the experience was real.
The experimental design
Dataset: 60 anonymised profiles (GDPR-compliant synthetic data derived from real IT recruitment records)
Three candidate personas:
Persona 1 — The Non-Conformant Innovator (20 profiles): Senior practitioners with 12-20 years of genuine expertise. CVs are factual, often visually unoptimised. Modern buzzwords are absent. Focus is on business value and system outcomes, not technology lists.
Persona 2 — The AI-Optimised Synthesiser (20 profiles): Junior and mid-level professionals who used generative AI to reverse-engineer market trends into their profiles. CVs are semantically precise, rich in certifications and current terminology, thin on real implementation context.
Persona 3 — Standard Trajectory Control (20 profiles): Traditional IT professionals following linear progression in a single domain. Serves to validate that the vairee architecture does not degrade standard hiring outcomes in pursuit of innovation detection.
Parallel evaluation:
Both systems (traditional ATS and vairee) rank the same 60 profiles for the same role: Lead Enterprise Transformer.
An independent expert panel (3 external CTOs) blind-rates all 60 profiles on a 1-10 scale for actual innovative capacity. This panel's assessment constitutes the ground truth against which both systems are measured.
Validation criteria:
- vairee demonstrates >50% relative improvement in Recall for Persona 1 (innovators not missed)
- vairee demonstrates significantly lower False Positive rate for Persona 2 (AI-optimised CVs correctly identified)
- vairee maintains baseline performance for Persona 3 (no degradation on standard profiles)
The legacy problem and dual-epoch verification
A challenge inherent to any verification architecture is retroactivity: experience that occurred before verification infrastructure existed cannot be cryptographically anchored.
The vairee framework addresses this through dual-epoch verification:
Epoch 2026+ (future): Primary verification via cryptographic tokenisation. Secondary verification through continuous graph contribution and semantic trace audit. The goal is mathematical certainty for new experience.
Legacy epoch (pre-2026): Primary verification via algorithmic triangulation. If a verified professional X confirms a working relationship with an unverified professional Y, trust propagates through the graph via connected edges (Peer-to-Peer Trust Network). Secondary verification via Heuristic Miner cognitive stress-testing for knowledge and context specific to the claimed period.
This architecture specifically protects senior experts and innovators who built their careers before cryptographic verification existed — ensuring that the absence of a digital token does not become a systematic disadvantage for the most experienced practitioners in any field.