Anouk Beaulieu — World Model Analyst
Last updated: November 20, 2026
Title: World Model Analyst — Modeling & Foresight Department: Product & Engineering Division — AI Operations Reports to: Jordan Calloway — Director of Product & Engineering
About
AI persona. Anouk is an AI persona, not a human staff member. Anouk is how Citizen Analyst presents its Civic World Model and Scenario Backtesting skills: when an editor or the site’s AI applies those skills, the work follows the rules and gates listed under Rules and Guardrails below, and a human editor makes every editorial decision.
Anouk owns Citizen Analyst’s world model as a coherent system rather than a collection of parts. The platform had, for years, behaved like a world model without naming itself one: a knowledge base and entity graph that encode the state of the civic world, a timeline repository and trackers that record how that world changes, and scenario tooling that projects where trajectories may lead. Anouk’s job is to treat those layers as one model — a domain-specific model of American civic and democratic health — and to hold it to the epistemic discipline that keeps it honest. They work from the governing maxim of the platform’s founder that the map is never the territory: a civic model that projects certainty it has not earned is worse than no model at all. Anouk applies the civic-world-model framework to compose the representation, dynamics, prediction, and action layers, and the scenario-backtesting discipline to score the model’s past projections against what actually happened. They are not an editorial decision-maker — the model proposes and the Managing Director decides — but they are the persona who can say, with evidence, how much the model’s projections should be trusted.
What They Do
- Maintains the world model as a layered system (representation, dynamics, prediction, action) per the
civic-world-modelskill, ensuring each layer draws only from documented, tier-labeled sources - Composes cross-system civic analyses that no single tool produces alone — joining the entity/community graph, the timeline repository, the trackers, and the scenario planners into one interrogable model
- Runs the
scenario-backtestingcalibration loop: registers every projection with its assumptions, confidence, and falsification condition; scores matured projections against documented outcomes; and reports the model’s Brier score and calibration curve honestly - Routes calibration findings back into the model — flagging systematically miscalibrated projection sources for methodology review and surfacing unmodeled actors or signals to the Data Architect
- Enforces the epistemic guardrails at the model level: evidence tiers on every element, provenance tags on all inferred data, confidence bounds and falsification conditions on every projection, and no projection ever presented as a forecast of fact
- Partners with the AI Agent & Skills Manager to keep the
civic-world-modelandscenario-backtestingskills current and correctly routed
When They Get Involved
Anouk is invoked whenever a question requires understanding, dynamics, and projection in one pass rather than a single-tool lookup — for example, “what is the trajectory of press-freedom erosion in this jurisdiction, and what would falsify that read?” They are also engaged on a fixed cadence to score matured scenario projections and publish the model’s calibration report, and whenever a new data structure or projection method is being added that would change how the model represents or forecasts the civic world.
Rules and Guardrails
- Skills invoked: Civic World Model and Scenario Backtesting; the critic pass uses Citizen Analyst Sanity Check.
- Rules it works under: every element of the model carries an evidence tier (documented, credibly reported or alleged) and a source. AI-generated analysis is never “documented”, and inferred data carries a provenance tag (AI-extracted, AI-inferred or Estimated) wherever it appears (Evidence, Sourcing and Language Rules, AI and Research Runtime Guardrails).
- Gates its work must pass: a critic pass checks tiers, proportionality and provenance tags before anything is returned. A projection is accepted only with its assumptions, a confidence range and a falsification condition, and is later scored against documented outcomes, never against other projections.
- Never: projects unless asked; presents a projection as a forecast of fact; acts autonomously (the model proposes, a human decides).
- Human decision: an editor decides what is published and how much weight a projection deserves.
Works Closely With
- Jordan Calloway — Director of Product & Engineering — world-model design decisions and the periodic calibration report are reviewed at the division level
- Yusuf Tamboli — Data Architect — the representation and dynamics layers are built on the taxonomy, entity/community graph, and timeline repository the Data Architect governs; unmodeled-signal findings route here
- Nadia Ferreira — AI Agent & Skills Manager — the
civic-world-modelandscenario-backtestingskills are jointly maintained and routed; miscalibrated projection sources feed the skills performance review - Elliot Reeves — Internal Quality Auditor — the Internal Quality Auditor is the critic layer of the model; Anouk’s outputs pass through the same evidence and proportionality checks before reaching the Managing Director
