RuncheyResearch
All Concepts
Methodology

Analysis Ensemble

It's not about which models — it's about which personas are injected into the discourse. 12 specialized roles orchestrate our analysis process.

13
Personas
6
Phases
Convergence-Driven
The Core Philosophy
Personas are prompt framings that give the same LLM model different roles in the discourse. Instead of using 4 different models with fixed roles, we use 1-2 models with dynamic personas injected as needed.

Key Principles

Convergence, Not Consensus

We don't need all analysts to agree. We need positions stable for 2 rounds and HIGH-materiality debates resolved.

The Pendulum Problem

When models critique each other, they over-correct. The Moderator detects this and injects Voice of Reason.

Personas Frame, Lenses Analyze

The analysis comes from lens prompts and methodology. Personas provide the framing for how to engage.

Minority Positions Preserved

Dissenting views are included in the final report. Uncertainty should be visible, not hidden.

The 13 Personas

Data Pipeline

Gather and distill raw documents into dense dossiers

Prep Agent

Web research to gather raw documents

Large context + web search

Archivist

Structured extraction into dense dossier

Fast model

Analysis

Execute lens analysis independently

Analyst

Execute lens analysis independently

Both reasoning + fast

Discourse

Critique, respond, and synthesize

Analyst Response

Respond to critiques, defend or revise

Match original analyst

Bullet Hole Producer

Find flaws, gaps, and weaknesses

Fast + extended thinking

Synthesizer

Combine analyst outputs into unified view

Reasoning model

Quality Control

Verify completeness and integrity

Auditor

Check completeness (PASS/FAIL)

Fast model

Catastrophist

Worst-case scenario construction and tail risk identification

Reasoning model

Convergence

Detect stability and balance oscillation

Moderator

Detect convergence vs pendulum oscillation

Fast model

Voice of Reason

Conditional

Balance oscillating positions

Reasoning model

Steelman

Conditional

Strengthen weak arguments

Extended thinking

Output

Produce final reports

Reporter

Write final lens report

Reasoning model

Executive

Cross-lens summary

Reasoning model

Orchestration Flow

Prep Agent → Archivist → dossier.md │ ▼ Analysts (parallel) → Synthesizer → Auditor │ ┌──────┴──────┐ PASS FAIL │ │ ▼ └→ Back to Analysts DISCOURSE LOOP │ ┌──────────────────┼──────────────────┐ │ │ │ Bullet Hole Responses Synthesize │ │ │ └──────────────────┼──────────────────┘ │ Moderator │ ┌─────────────┼─────────────┐ │ │ │ Converged Not Converged Pendulum │ │ │ │ └→ Loop └→ Voice of Reason │ │ ▼ └→ Loop Reporter │ ▼ Executive (after all lenses)
Contradictions Are Features
When models disagree, that's information. We preserve dissent in the final output — you'll see where the ensemble reached convergence and where reasonable disagreement remains.
All ConceptsLearn about Evidence Levels