How VoxSim works

Primary axis: 12 political clusters + echo chamber

VoxSim's primary axis is political tendency, not demographics. Turkish society is split into 12 political clusters (social democrat, religious conservative, nationalist conservative, Kurdish political, Gen-Z undecided, etc). Each cluster is fed by its own echo chamber.

For each cluster, news sources are scanned across four time windows; the AI reads this time-weighted news environment and distils the cluster's current mood, grievances, and dominant narrative frames. Discourse evaluation runs against this live context.

The population share (% figure) shown for each cluster is an estimate derived from cross-tabulated national public-opinion surveys and election results. A lifestyle typology from Türkiye's leading independent public-opinion research institution (referenced as LKA under a commercial data licence) provides the core framework; the institution's name is kept confidential per licence terms, raw tables are never republished — only statistical inferences are carried over.

Bourdieu 3-Capital × LKA Subtypology

Faz 1.8 C4 (2026-05-27): each cluster now carries Bourdieu's three-capital framework (Distinction, 1984 — economic · cultural · social) plus a subtype from LKA's publicly available October 2024 subtypology study (Modern Secular, Conservative Modern, Religious Conservative, etc.). The table reflects aggregate cluster behaviour; never an individual persona. SyntheticPollPrompt optionally injects these profiles to improve the LLM's Turkey-specific habitus modelling. As of 2026-09-17 the profile is no longer injected into production queries; its ablation contribution could not be distinguished from noise.

1 = very low, 5 = very high. Navy shades = high capital, amber shades = low capital.

Academic references: Bourdieu (1984) — Distinction: A Social Critique of the Judgement of Taste (Harvard UP); Çarkoğlu & Kalaycıoğlu (2021) — Fragile but Resilient? Turkish Electoral Dynamics, 2002–2015 (U. Michigan Press); Çarkoğlu & Kalaycıoğlu (2020) — Elections and Public Opinion in Turkey (Routledge); Çarkoğlu & Toprak (2007) — Religion, Society and Politics in a Changing Turkey (TESEV).

Information ecology mapping (3-tier hybrid sourcing)

The natural quality filters of major news services exclude some of the most defining sources for specific clusters. Niche publishers — especially ideologically focused outlets — are absent from global news catalogs, yet they reflect the actual mood and concerns of certain political clusters far more accurately than mainstream sources.

VoxSim closes this blind spot with a multi-layer sourcing approach: publishers' own RSS feeds form the foundation of a layered news base (mainstream + niche + echo chamber). Each piece of content is classified by source quality and type.

Academic references: Sunstein (2017) — #Republic: Divided Democracy in the Age of Social Media (Princeton UP); Bail (2021) — Breaking the Social Media Prism (Princeton UP); Levy (2021) — Social Media, News Consumption, and Polarization (American Economic Review); Cinelli et al. (2021) — The Echo Chamber Effect on Social Media (PNAS); Bakshy, Messing & Adamic (2015) — Exposure to Ideologically Diverse News on Facebook (Science); Garrett (2009) — Echo Chambers Online? Selective Exposure Among Internet News Users (JCMC).

Daily echo-chamber pipeline

Each cluster's snapshot is rebuilt nightly through a four-stage pipeline. Total runtime ≈ 30 minutes; happens between 02:30 and 03:30 Istanbul time. Daytime simulations always read the previous night's context. Important constraint: breaking crises that emerge during the day are not reflected in the snapshot until the same-night pipeline completes — Module B should be paired with Event Injection (Module E, Karargah+) for same-day crisis scenarios.

Exponential forgetting curve (w = exp(−t/τ))

Each article receives a weight <strong>w(t) = exp(−t / τ)</strong> based on its age; bucket weights are derived deterministically by integrating this curve over window boundaries. τ (tau) is the single source of truth — half-life is τ·ln(2). Grounded in Murre &amp; Dros (2015 PLOS ONE) Ebbinghaus replication and Wu et al. (2007 PNAS) news attention decay.

Source anonymity

All public-facing surfaces — config files, API responses, PDF reports, audit logs, this page — refer to news sources by codename only (e.g. ULS_01, MRS_03). The real source URI mapping lives in an encrypted file accessible only to Talivio compliance staff.

Backtest temporal pin

When VoxSim simulates a historical event (e.g. February 2023 earthquake), it would be misleading to use today's cluster snapshot — clusters' beliefs in 2023 were different. Backtest selects the snapshot version active on or before the event_date.

Academic references: Murre & Dros (2015) — Replication and Analysis of Ebbinghaus' Forgetting Curve (PLOS ONE); Tversky & Kahneman (1973) — Availability: A Heuristic for Judging Frequency and Probability (Cognitive Psychology).

Secondary axis: demographic nuance

Alongside the political-cluster axis, a secondary demographic axis runs as well. Türkiye's adult population is split across three independent dimensions to build a statistical model: geography (NUTS-2 regions), socio-economic tier (SES), and cultural-lifestyle cluster. Each combination forms a distinct "cell" in a 26 × 4 × 3 matrix.

Within each cell, synthetic agents inherit statistical distributions derived from publicly available demographic surveys and census data. No real individuals are represented — every agent is a procedurally generated statistical composite.

The three lifestyle clusters reflect research on Turkish society's self-identification patterns with respect to secularism, tradition, and religiosity. The approximate national-level proportions are:

Academic references: Beckman, Baggerly & McKay (1996) — Creating Synthetic Baseline Populations (Transportation Research A); Müller & Axhausen (2011) — Hierarchical IPF: A Synthetic Population for Switzerland (ERSA); Park et al. (2023) — Generative Agents: Interactive Simulacra of Human Behavior (UIST); Argyle et al. (2023) — Out of One, Many (Political Analysis); Aher, Arriaga & Kalai (2023) — Using LLMs to Simulate Multiple Humans (ICML).

Historical event calibration

A simulation engine is only as good as its calibration. VoxSim validates its population model against historical events where an objective outcome — an election result or a credible polling snapshot — provides ground truth. This process is called backtest calibration.

1. 1

   Collect historical events

   Each event is documented with date, category, a plain-language description, and a URL pointing to public sources (Wikipedia, official gazette, press archives).

2. 2

   Assign a measurement anchor

   Only events with a qualifying anchor — a national election result, a credible nationwide poll, or official statistics — enter the calibration set. This is the distinguishing criterion.

3. 3

   Run population simulation

   The event description is passed to the simulation engine, which predicts emotional response (anger, fear, hope, apathy, support) across all population cells.

4. 4

   Compare against ground truth

   Predicted distributions are compared against the anchor measurement. For elections, vote-share shifts are mapped to segment-level support estimates. For polls, direct sentiment questions provide validation.

5. 5

   Iterate and re-calibrate

   Systematic deviations reveal model biases (e.g. underweighting rural religiosity). Parameters are adjusted and the calibration set is re-run until mean absolute error stabilises.

The anchor condition is strict: an event enters the calibration set only if a qualifying measurement (election or poll) falls within approximately six months of the event. This ensures the predicted societal response can be compared against an independently observed signal.

Academic references: Brier (1950) — Verification of Forecasts (Monthly Weather Review); Tetlock (2005) — Expert Political Judgment (Princeton UP); Hyndman & Koehler (2006) — Forecast accuracy measures; Gneiting & Raftery (2007) — Strictly Proper Scoring Rules (JASA).

Accuracy score calculation

Each backtest run produces a per-event accuracy estimate. The overall accuracy score reported in VoxSim is the mean across all calibration events, weighted by event reliability.

The score is decomposed by topic category (economic, political, security, etc.) so users can see exactly which policy domains the model performs best and worst on. Every score links back to the specific events that produced it — full transparency, not a black box.

Academic references: Willmott & Matsuura (2005) — Advantages of MAE over RMSE (Climate Research); Hyndman & Koehler (2006) — Another look at measures of forecast accuracy (Int'l J. Forecasting); Gneiting & Raftery (2007) — Strictly Proper Scoring Rules (JASA).

Consistency of cluster seed distributions with YSK (2002–2024, 8 anchors)

Each cluster's hand-curated historical vote distribution is weighted by its population share, aggregated nationally, and compared against the actual YSK result. This is NOT an independent validation: because the distributions were curated by looking at these very elections, the result is an in-sample consistency measure — it shows that the clusters sum arithmetically to the official result, not that the model has forecasting power. It involves no AI call. The 8 anchors below (2002, 2007, 2011, June 2015, November 2015, 2018, 2023, 2024 local) produce the composite score; pre-2014 anchors are back-projected onto today's cluster structure.

Scores are recalculated automatically every day and update instantly when model parameters change. These computations require no AI inference — they are purely deterministic mathematical operations.

Replication variance & algorithmic fidelity

Two foundational academic warnings are addressed with explicit system safeguards. Both tests below run on live data.

Prompt instability — Bisbee (2024) replication sweep

Bisbee et al. (2024, Political Analysis) warn that LLM-based synthetic surveys exhibit low variance + high prompt instability. Relying on a single shot is dangerous; the same prompt must be repeated N times. VoxSim runs an N=5 replication sweep per template, measures the coefficient of variation (CV), and classifies into a three-tier band. The weekly automated sweep measures the production template on a sample basis; separately, the 2026-08 measurement program found the run-to-run noise floor of a single-shot poll to be 4.7–5.8 points (see §04.7).

Monte Carlo uncertainty band & LLM stochasticity

The ± values in report PDFs are not an internal confidence measure of a single run; they are derived from the standard deviation of N independent Monte Carlo runs of the same simulation. Large language models are not deterministic: due to sampling temperature and context effects, the same input produces somewhat different output on every run. Rather than suppressing this inherent stochasticity, VoxSim measures it and reports it as an uncertainty band.

A narrow band is not a guarantee of accuracy — it only shows the model is internally consistent; systematic bias does not show up in band width. A wide band signals the discourse sits in a "sensitive zone" for that cluster. The band should therefore be read as a model-consistency indicator, not as a real-world margin of error.

Algorithmic fidelity — Argyle (2023)

Argyle et al. (2023) central claim: synthetic population models must reproduce real-world demographic subgroup distributions. VoxSim aggregates its 12 cluster priors into the three LKA lifestyle segments (modern / traditional / religious) and compares them against published cross-tabs from LKA's monthly public opinion survey. Fidelity score is based on Total Variation Distance (TVD): 1.0 = perfect, 0.0 = completely off.

Cultural bias mitigation — Atari (2023)

Atari et al. (2023) demonstrate that LLMs (GPT/Gemini included) lose cross-cultural validity in non-WEIRD societies (r=-0.70 for collectivist cultures). Türkiye sits in the non-WEIRD bracket, making this drift critical. Mitigation: 36 Turkish aggregate behaviour patterns (drawn from publicly available YSK + LKA aggregates; never individual personas) are injected into the prompt as in-context examples (Brown et al. 2020). Mood and vote-rationale generation are calibrated to the Turkish context.

Academic references: Bisbee et al. (2024) — Synthetic Replacements for Human Survey Data? The Perils of LLMs (Political Analysis); Argyle et al. (2023) — Out of One, Many: Using LLMs to Simulate Human Samples (Political Analysis).

Measurement integrity program — what does the score actually measure?

Module K's calibration score can be misleading on its own, because the seed distribution given to the model is derived from the official result: even a model that simply echoes the seed back would score well. In August 2026, four tests were built to measure this question directly, and were run under a single pinned model across two independent production polls. Results and the decisions taken are below; the measurement files are kept under version control.

Evidence hierarchy — how is each claim supported?

All measurement commands and decision tables are documented in the product repository. Measurements are valid under a single pinned model; they are re-run whenever the model fingerprint changes.

Academic references: Bisbee et al. (2024) — Synthetic Replacements for Human Survey Data? (Political Analysis); Dominguez-Olmedo, Hardt & Mendler-Dünner (2024) — Questioning the Survey Responses of Large Language Models (NeurIPS); Distribution-First Population Simulation: Collapse, Calibration, and Recall in Non-WEIRD LLM Persona Modeling; Murphy (1988) — Skill Scores Based on the Mean Square Error (Monthly Weather Review 116:2417).

Historical event coverage

The calibration set consists of hand-curated events spanning 1999 to 2025; the current count is computed live below.

150 events total

Earlier events (pre-2010) use electoral data as their primary anchor, since polling archives are sparser. Post-2013 events benefit from richer survey research, producing more granular per-segment ground truth.

The calibration set is continuously expanded. New events added in each release are noted in the changelog. The target is 500+ events by end of year two.

Academic references: King & Lowe (2003) — Automated Information Extraction for Conflict Data (International Organization); Schrodt (2012) — Precedents, Progress & Prospects in Political Event Data (Int'l Interactions); Boschee et al. (2015) — ICEWS Coded Event Data (Harvard Dataverse); Leetaru & Schrodt (2013) — GDELT: Global Data on Events, Location and Tone.

In its default mode, Module K generates an intra-cluster party distribution for each of the 12 political clusters in a separate AI session; the national estimate is the population-weighted sum of these distributions. Sampling across 26 NUTS-2 × 4 SES × 3 lifestyle = 312 demographic cells is used only in the cell-level mode available on the Karargah+ plan. It provides a decision-support signal based on synthetic simulation for the question "what would happen if elections were held today?" — not a public opinion poll, but a strategic internal assessment tool.

1. 1

   Cluster-based parallel analysis

   Each of the 12 political clusters — given its cluster identity, lifestyle anchor, ideology axes, and that cluster's echo-chamber snapshot — is evaluated in a dedicated session. The model revises the cluster's seed distribution based on the live agenda in its echo chamber.

2. 2

   Recall calibration

   Each cluster is also asked for its recalled distribution in the 2024 local, 2023, 2018, and November 2015 elections; the weighted national total is compared against actual YSK results. This comparison is NOT a corrective, and since the 2026-08 measurement program it is no longer treated as a standalone confidence signal either: alongside the score, a skill value is reported that measures what the model gains relative to the seed distribution it was given — if skill is negative, the confidence band stays low regardless of the score. Every poll holds out one anchor: the model is given neither its seed distribution nor its result; the confidence band is primarily derived from this anchor.

3. 3

   Population-weighted national aggregate

   All segment outputs are weighted by population share and combined into a national estimate. The national total is not anchored to any polling firm's results or to a poll average. Historical election results are used only as a calibration-quality indicator (recall); they do not pull the estimate toward past outcomes. Invalid or missing segments are automatically excluded.

Until 2026-09-21 the national total was anchored to a consensus of published poll averages via Iterative Proportional Fitting (raking). That method has been removed: the synthetic result no longer targets any polling firm's results or uses a poll average; the published figure is the raw population-weighted total. The blog, Pulse page and home page all show this single series. Client reports produced before that date remain archived under the method of their time and are labelled "consensus-anchored (legacy method)".

Academic references: Argyle et al. (2023) — Out of One, Many: Using Language Models to Simulate Human Samples (Political Analysis); Horton (2023) — Large Language Models as Simulated Economic Agents: Homo Silicus (NBER WP 31122); Bisbee et al. (2024) — Synthetic Replacements for Human Survey Data? (Political Analysis); Aher, Arriaga & Kalai (2023) — Using LLMs to Simulate Multiple Humans (ICML); Deville & Särndal (1992) — Calibration Estimators in Survey Sampling (JASA 87:376-382); Deville, Särndal & Sautory (1993) — Generalized Raking Procedures in Survey Sampling (JASA 88:1013-1020).

Known limitations

VoxSim is a decision-support tool, not a prediction oracle. Honest communication of limitations is central to responsible use.

• Synthetic agents ≠ real people

  Responses come from statistical agents, not actual citizens. The model cannot capture individual stories, micro-community dynamics, or real-time mood shifts.

• Calibration lag

  Model parameters reflect the most recent calibration cycle. Rapidly evolving situations — breaking crises, sudden leadership changes — may not yet be reflected.

• Correlation, not causation

  A high anger signal does not mean your policy caused that anger. Other concurrent factors may explain the historical pattern; the model attributes to the event, not to external context.

• Coverage gaps

  Pre-1999 events have no polling anchor and are excluded. Regional events affecting only specific NUTS-2 zones may be under-represented in aggregate accuracy.

• LLM variability (mitigation: Bisbee + Atari)

  Simulation outputs are generated by large language models and carry inherent LLM limitations: hallucination risk, sensitivity to prompt wording, lack of genuine world-understanding. Faz 1.8 introduced active mitigations — Bisbee replication sweep (N=5 σ measurement, see 04.6) and Atari WEIRD-bias mitigation via a Turkish few-shot library (12 clusters × 3 aggregate patterns). Single-shot output remains limited; if the production replication detects an unstable band, the PDF embeds a warning.

• A synthetic poll is not an opinion poll

  Module K produces an LLM-based distribution estimate across 12 political clusters; it is NOT a poll of real voters and can only be shared under the synthetic-simulation framing, with a watermark. Past-election recall distributions are used solely as a diagnostic indicator — never as a corrector that drags predictions back to the last election. Module K is auto-disabled during the 10 days preceding any announced election.

• Municipal polls do not localise cluster shares per province

  In the municipal poll variant, the national shares of the 12 political clusters are not re-estimated per province (a deliberate decision driven by sample-sparsity risk in small provinces). Province context is injected into the LLM prompt and the model revises within-cluster behaviour accordingly; in atypical provinces (where a cluster's true local share deviates strongly from its national share) the error margin of this approach grows. Calibration is measured per province against 2019 + 2024 YSK anchors; anchors for 51 provinces are compiled from secondary sources and low-confidence provinces are flagged UNCERTAIN in config.

Academic references: Bisbee et al. (2024) — Synthetic Replacements for Human Survey Data? The Perils of LLMs (Political Analysis); Santurkar et al. (2023) — Whose Opinions Do Language Models Reflect? (ICML); Park et al. (2023) — Generative Agents: Interactive Simulacra of Human Behavior (UIST); Atari et al. (2023) — Which Humans? (WEIRD bias in LLMs).

Compliance & ethical framework

VoxSim is designed with EU and Turkish data protection, AI transparency, and ethical requirements in mind.

• EU AI Act (High-Risk Classification)

  AI systems that may influence democratic processes fall under high-risk classification. VoxSim addresses this with transparent accuracy scoring, human review gates, and ethical filters against misuse.

• GDPR (General Data Protection Regulation)

  The system uses synthetic (artificial) data — no real individual personal data is used. Customer documents are processed under a DPA. Data retention is 12 months; right to erasure is provided via self-service.

• KVKK (Turkish Personal Data Protection Law)

  A KVKK-compliant data processing agreement applies for Turkey-based customers. Talivio Technology OÜ acts as data processor; the customer is the data controller.

• Synthetic data & privacy

  Generated simulation outputs are not based on any real person's or survey respondent's data. Statistical distributions (TÜİK and public research) are used only at aggregate level.

For compliance and legal enquiries: [email protected]