MiroFish · Swarm Intelligence ResearchThe Simulation of Democracy
45 AI agents. 100 simulation rounds. 1,965 discrete social interactions. A full multi-agent AI simulation of Hungary's 2026 parliamentary election — with no polling data used as input. The prediction emerges entirely from collective behaviour.
Polling captures a snapshot. This simulates the process.
Standard polls cannot model how a rural pensioner's vote shifts after weeks of state television, or the strategic calculation of a DK supporter in a marginal Budapest constituency who knows their candidate cannot win. MiroFish addresses this gap by instantiating AI-powered agents representing distinct voter archetypes, letting them interact, form opinions, and arrive at vote intentions through simulated social dynamics — with no polling data used as an anchor.
Hungary's April 2026 election is the most competitive since 2010 — a genuine contest between Fidesz and TISZA (Péter Magyar's party) on a structurally asymmetric field of gerrymandered constituencies and state-media dominance. The question posed to the simulation: “How do different voter segments form their final voting decision — and what aggregate outcome does that produce?”
“The experiment was run by a 21-year-old student with no data-science background, in the middle of LSE exam revision, on less than $10 in API credits.”
Architecture & stack
The simulation runs on MiroFish — a platform built on CAMEL-AI and OASIS for the agent layer, Gemini 2.5 Flash Lite via OpenRouter for all reasoning calls, and Neo4j AuraDB as the shared knowledge graph. Agents were seeded with six structured documents covering the Hungarian electoral system, party platforms, the media environment, economic grievances, and demographic profiles — but explicitly not with polling results.
After the run, InsightForge (MiroFish's reporting layer) produced an initial output. This was followed by a deep-interaction phase — structured surveys sent to individual agents to extract vote-share estimates by segment. The raw output then passed through eleven documented bias corrections under a two-layer epistemic recalibration framework before final vote shares were accepted.
Constituency-level seat modelling was handled by Chronicler-v2, calibrated to a strategic transfer rate (lambda = 0.58) from DK and MKKP voters to TISZA in competitive single-member districts.
Final 199-seat parliament
TISZA wins a governing majority by a single seat. Fidesz drops 51 seats from 2022. The result lands within 1pp of independent polling — without having been anchored to it at any point.
| Party | SMD | List | Total | Share |
|---|---|---|---|---|
| TISZA | 66 | 35 | 101 | 50.8%Majority |
| Fidesz-KDNP | 38 | 46 | 84 | 42.2% |
| DK | 1 | 6 | 7 | 3.5% |
| MKKP | 0 | 5 | 5 | 2.5% |
| Other | 1 | 1 | 2 | 1.0% |
| Total | 106 | 93 | 199 | 100% |
TISZA reaches the bare majority threshold of 100 seats — enough to form a government and pass ordinary legislation, but not enough to change the constitution (which requires 133). Fidesz remains the largest party by vote share at 41.0% but loses their supermajority era. TISZA wins 66 SMDs not because it is nationally more popular, but because opposition votes consolidate behind a single candidate in enough constituencies.
The decisive variable: lambda
Lambda (λ) is the strategic transfer rate — the share of DK and MKKP voters who tactically back TISZA in competitive single-member districts. It is the single variable that most determines the outcome.
TISZA gains a majority only when lambda crosses 0.57. At λ = 0.55 they finish at 99 seats — one short.
TISZA 101 seats. Bare majority, one-seat margin. Fidesz loses their majority once lambda exceeds 0.20.
TISZA falls short of a majority. Hung parliament — without precedent in modern Hungary.
A similar 4–6 seat shift. Hung parliament becomes the central scenario.
Research by Matteo Mio · Forward College / University of London (LSE programme) · April 2026 · Built with MiroFish, CAMEL-AI, OASIS, Neo4j AuraDB, Gemini 2.5 Flash Lite · Total cost: under $9 · Part of the SCALEFORM builder-studio research track.