Methodology
Neutrality is a structure, not a slogan.
Republiq's analysis stack runs entirely on xAI's Grok. Neutrality isn't enforced by asking the model to “be fair” — it's enforced by the shape of the outputs the model is required to produce.
The four production pipelines
Distinct surfaces, one provider, the same neutrality constraints applied throughout.
News Lens
Sentiment, framing, and bias analysis on news coverage.
Production
Each article is scored for hype index (0–10), an emotion vector (fear, anger, joy, disgust, sadness, surprise), and a spectrum lean from −1.0 to +1.0. Outputs include three Challenge Perspectives and three Probing Questions to surface assumptions the framing depends on.
Bill Enrichment
Deep legislative analysis of incoming bills.
Production
Bill text is structured into a deep summary, fiscal-impact note, and constitutional implications. Provisions are flagged by type (earmark, rider, sunset, preemption). Triggered automatically on Congress.gov ingest; admin re-enrichment is available.
Research Suite (4-agent orchestrator)
Multi-step reasoning for complex civic questions.
Production (Premium)
Complex queries route through a Research Agent → Legal Agent → Writing Agent → Logic & Language Agent pipeline. Implemented as a native Python state machine with explicit handoffs — no third-party agent framework. Premium tier; 50 chains per user per month.
Politician Mindset Snapshot
Multi-source view of a politician's positions and patterns.
Production
Combines voting record (Congress.gov + OpenFEC), public statements (X), and stance positions (Vote Smart) into a radar chart, trait summary, and cited narrative.
Neutrality guardrails
Six structural controls applied across every analysis surface.
Single-provider AI stack
All AI features run on xAI's Grok (default: grok-4-1-fast-reasoning, with explicit reasoning enabled for complex analysis). Concentrating on one provider keeps the system auditable: one prompt library, one cost ledger, one evaluation surface.
Source-anchored prompts
Bill text and government records are passed in as the grounding context for analysis, with truncation thresholds chosen to keep model behavior predictable. Citations always link back to the full upstream document, not the truncated excerpt.
Challenge Perspectives
Every analysis output includes three Challenge Perspectives — alternative readings of the same record. The structural requirement to produce them prevents single-framing outputs from being published as the only reading.
Probing Questions
Each analysis ships with three Probing Questions designed to surface the assumptions the framing relies on. Each question carries an implied-reasoning field and a broader-application field — the analysis exposes its own scaffolding.
Spectrum lean as an explicit field
News Lens outputs include a spectrum-lean score from −1.0 (left) to +1.0 (right) as a first-class field, surfaced alongside the analysis rather than buried in tone.
Versioned prompts and usage tracking
Prompts are version-controlled in an internal Prompt Studio. Token usage and model cost are logged per feature so the cost and behavior of every analysis surface can be reviewed.
What we don't claim — yet
A formal third-party bias audit pipeline is on the roadmap and is not yet running in production. The neutrality guarantees on this page describe the structural controls that are running today — primary-source grounding, Challenge Perspectives, Probing Questions, explicit spectrum-lean detection, single-provider auditability, and versioned prompts.
When the formal audit cadence ships, it will be published here with results.
Read the records. Read the analysis. Verify both.
The product is built so you don't have to take our word for anything.