Research System Architecture

V5.0: Evidence-Based Multi-Model Jury

Three independent models with evidence-weighted scoring and jury consensus logic. Current production system adopted 2026-06-01.

V4.0's anchor overcalibration and single-model bias revealed that rigor requires: (1) evidence-based scoring (not exemplar anchoring), (2) multi-model consensus (not single-point estimates), (3) formal inter-rater validation, and (4) human governance gates.


Organization + Evidence Package8-Tier Evidence HierarchyPrimary Sources → Secondary → Grey → AI-DerivedIndependent Jury ModelsClaude SonnetNo shared contextGPT-4oNo shared contextGemini ProNo shared contextConsensus Logicjury_mean • jury_median • jury_spread • consensus_strongSpread Thresholds: 0-5pt strong | 6-10pt accept with notes | 11-20pt review | >20pt reviseHuman Review Gate≥2/3 Jury Required • Governance Decision • Acceptance/Modification

Three-Model Jury

V5.0 uses three independent LLMs (Claude Sonnet 4.6, GPT-4o, Gemini Pro) without shared context. Each model scores independently on the same evidence package.

Evidence Framework

Evidence classified into 8-tier hierarchy: Peer-reviewed journals → Books → Government reports → Interviews → News → Documentary → Grey literature → AI-derived.

Consensus Logic

  • jury_mean: Average of three scores
  • jury_spread: Max - Min (indicates agreement strength)
  • consensus_strong: Boolean (spread ≤ 5 points)

Human Review Gate

Jury proposals require human review before acceptance based on spread thresholds.

Scoring Outputs (Generated in Parallel)

V5.0 generates two scoring outputs simultaneously for each organization:

  • Young's Original Score: 0-10 binary checklist across the 10 criteria. How many of Young's indicators are present?
  • Composite Score: 0-100% formula-based: (Breadth ÷ 10) × (Mean Intensity ÷ 10) × 100. Measures evidence-weighted prevalence and intensity.

Both scores are generated from the same jury consensus and evidence package. Users receive both perspectives on each organization.


Adoption Date2026-06-01
Organizations Scored565+ (all with jury provenance)
Status✓ Current Production