VoxSim: A Political Simulation Platform for Turkey, Built on Echo Chambers and Historical Calibration

VoxSim is a political simulation and decision-support platform built specifically for Turkey. It takes a policy text, a campaign message, a draft law, or a public announcement — and simulates how twelve calibrated political clusters across the country would react before the message reaches the real public.

The platform is not a polling tool. It doesn't survey real people. What it does is maintain living, continuously updated models of how each major segment of Turkish political society currently thinks — built from their actual information ecosystems — and then run those models against new inputs using large language models in parallel.

We built it because there was no tool that answered the specific question political communicators actually need answered: not "what do voters think today?" but "how will this specific message land with these specific segments tomorrow?"

Twelve Political Clusters, Not Demographics

Most public opinion tools segment populations by demographics: age, gender, income, region. VoxSim segments by political identity — which is how Turkish political behaviour actually works. A 45-year-old Istanbul professional and a 45-year-old Istanbul factory worker may share the same demographics but live in completely different information ecosystems, hold different values, and respond to the same message in opposite ways.

VoxSim's taxonomy defines twelve clusters, each with a distinct ideological profile. The Nationalist Left cluster holds state-centrist economic views combined with secular nationalism. The Social Democrat cluster is centre-left, urban, and oriented toward European integration. The Religious Conservative cluster combines market-friendly economics with strong Islamic identity. The Kurdish Political Movement cluster is geographically concentrated and has a specific political history that shapes how any national policy is interpreted.

Each cluster is defined not just by what it believes, but by where it gets information — the specific media sources, social media communities, and opinion leaders that shape its interpretation of events. This is what makes the difference between a demographic model and an echo chamber model.

The Echo Chamber Pipeline

The core technical work in VoxSim is maintaining what we call the echo chamber pipeline. Each of the twelve clusters has ten anonymized RSS news sources — the actual outlets, blogs, and media channels that cluster genuinely consumes. These sources are kept confidential (we refer to them internally by codenames) to prevent gaming, but they're real, active, and monitored daily.

Every day, the pipeline ingests articles from all 120 sources (12 clusters × 10 sources each), deduplicates them by canonical URL, and stores 150-word copyright-safe excerpts. A summarization pass then generates 200-word LLM summaries of each article. By midnight, every cluster has a fresh corpus of what its information ecosystem produced that day.

Each night at 3:00 AM Turkey time, the snapshot builder runs. It takes each cluster's article corpus, slices it into four time windows, and applies Ebbinghaus-inspired weights: the last two weeks count for 50% of the cluster's current mood, the last three months for 30%, the last six months for 15%, and the last twelve months for 5%. The recency weighting reflects how political memory actually works — recent events dominate, older events fade but don't disappear.

Gemini 2.5 Pro reads the weighted corpus and produces a structured cluster snapshot: current grievances, recent wins, enduring anxieties, narrative frames, active lexicon, and a weighted mood vector across five emotions (anger, fear, apathy, hope, support). This snapshot is what makes VoxSim's simulations current — not a static dataset from six months ago, but a model of what each cluster thinks this week.

The Crash Test: Twelve Parallel Simulations

The primary simulation module takes a policy or message text and runs it against all twelve cluster snapshots simultaneously. Rather than processing clusters sequentially, VoxSim uses Laravel's HTTP pool to fire twelve Gemini API calls in parallel — what would take ~36 seconds serially completes in approximately 5 seconds.

Each cluster simulation receives two context blocks: the cluster's current snapshot (its information ecosystem, current grievances, narrative frames, mood) and its demographic profile (the distribution across Turkey's 26 NUTS-2 regions and three socioeconomic strata). Gemini reads the policy text through the lens of that specific cluster's current reality and returns a five-emotion response distribution plus a sample quote in the voice of a typical cluster member.

The output shows, for each of the twelve clusters, how strong each emotional response would be and what a realistic cluster member might actually say. A minimum wage announcement might produce strong support in the Young Undecided cluster, strong hope in the Social Democrat cluster, and significant anxiety in the Religious Conservative cluster concerned about inflation implications — all from the same policy text, all in the same run.

Historical Calibration: 50+ Real Events

A simulation that can't be verified against reality isn't useful — it's just speculation with better production values. VoxSim's accuracy score, called the Doğruluk Skoru, is derived from backtesting against more than fifty historical political events where actual public reaction was documented.

The historical event library includes election campaigns, major economic announcements, constitutional changes, security crises, and foreign policy decisions — events stretching back to 2002, covering Turkey's full modern political period. Each event has documented ground truth: what actually happened in terms of vote shifts, polling changes, or measurable behavioral responses.

When a user runs a backtest, VoxSim reconstructs the cluster snapshots as they existed at the time of the historical event — not using current snapshot data, which would be cheating. The pipeline stores snapshot history, so a 2023 event is evaluated against 2023 cluster moods, not 2026 moods. This temporal pinning is what makes the calibration honest.

The vote recall data is anchored to YSK (Turkey's Supreme Electoral Council) results going back through the 2002, 2007, 2011, 2015, 2018, and 2023 elections. Each cluster carries a calibrated historical vote distribution that reflects its actual documented behaviour across these elections, not modelled assumptions about it.

Module K: Synthetic Polling

Beyond crash testing individual messages, VoxSim includes a synthetic polling module that generates full electoral scenarios. There are three poll types: parliamentary, presidential, and municipal.

Parliamentary polls simulate national seat distributions across Turkish parties, roll up alliance-level totals, and are validated against YSK recall data from three election cycles. Presidential polls handle two-round logic for up to five candidate configurations. Municipal polls operate at the NUTS-3 level (Turkey's 81 provinces, identified by license plate code) and are calibrated against 2019 and 2024 local election results for the thirty largest provinces.

LLM-generated poll outputs are corrected for known systematic biases before delivery — early testing revealed that Gemini consistently overestimated CHP support by 8–10 points relative to actual results. The calibration layer applies Deming-Stephan iterative proportional fitting to align outputs with the historical baseline before the results reach the user.

Parliamentary polls are blocked in the ten days before any announced election — a deliberate design constraint built into the backend, not just a policy. The system checks the configured next election date and returns an error if the timing would make the poll output potentially misleading during an active campaign period.

EU AI Act Compliance

VoxSim handles political information in a high-stakes context. We took this seriously from the start, which is why the system was built with EU AI Act compliance as an architectural constraint rather than an afterthought.

The platform is classified as a potential Annex III Article 8(a) system — one that could affect democratic processes. This classification triggers a set of mandatory requirements: human oversight mechanisms, technical documentation, bias reporting, and audit logging.

Every LLM call writes a mandatory audit record: workspace, purpose, model, token counts, latency, cost, prompt hash, and response hash. The audit log is append-only with six-month retention. Source brands are never exposed in simulation outputs — only internal codenames appear, so outputs cannot be traced back to specific media outlets. Workspace identifiers are never included in LLM prompts, isolating the cache layer from any customer data.

The system also runs twenty-eight adversarial tests in CI covering prompt injection, workspace data leakage, malicious input, and disinformation framing attempts. All 118 tests pass at launch.

What VoxSim Is Not

VoxSim doesn't micro-target individuals. A module that would have generated cluster-specific persuasion messages was designed, scoped, and then permanently removed before launch. The platform produces analysis, not manipulation tools. The distinction matters, and we were deliberate about it.

VoxSim is also not a prediction system. It produces probability distributions conditioned on current cluster moods and a specific input — the accuracy score tells you how well the model has performed on similar historical inputs, not that any specific prediction is correct. Users are expected to treat the output as structured intelligence, not as oracle results.

The platform doesn't track or profile real individuals. The cluster models are built from aggregate information ecosystems, not from individual user data. Organizations using VoxSim to test messages are not contributing surveillance data to the system.

Access and Availability

VoxSim is currently available only in Turkey — the clusters, calibration data, and source ecosystem are all Turkey-specific. Access is application-based rather than self-serve: organizations apply, go through a use-case review, and receive workspace credentials if approved. Academic and media researchers have a dedicated tier with institutional verification requirements.

If you're working in political communication, public policy, strategic communications, or academic research on Turkish political behaviour and want to explore what the platform can offer for your specific context, the application process is the starting point.