Synachron has a functioning simulation engine. Predictive validity is not yet proven.
The first formal benchmark is a transparent diagnostic backtest. It measures how an existing synthetic political survey differs from an external polling reference, while explicitly separating raw agent responses from post-calibrated results.
What this benchmark does not prove
It does not prove that Synachron predicted an election result or outperformed a real survey. The evaluated run used political calibration inputs and was executed after the reference poll was public. The result is useful for diagnosing model error and designing the next controlled experiment, not for making a marketing accuracy claim.
Mean absolute error across 14 comparable party buckets
Largest absolute party-level deviation
Party buckets inside the diagnostic tolerance band
Not a prospective or independent forecast
Benchmark record
- Simulation
- Test vs MARC 25/7 — 240 agents
- External reference
- MARC estimate of vote, fieldwork 20–23 July 2026
- Evaluation design
- Retrospective diagnostic; non-holdout; non-pre-registered
- Stored evidence
- Immutable prediction snapshot, actual snapshot, aggregate metrics and party-level errors
Raw synthetic result vs external reference
Selected party buckets. Undecided responses are excluded from both distributions.
| Party bucket | Raw synthetic | Reference | Error |
|---|---|---|---|
| New Democracy | 27.75% | 30.80% | -3.05 pp |
| EL.A.S. | 11.48% | 16.80% | -5.32 pp |
| PASOK | 11.96% | 11.00% | +0.96 pp |
| Greek Solution | 10.53% | 7.50% | +3.03 pp |
| Elpida | 11.48% | 7.40% | +4.08 pp |
| KKE | 0.96% | 6.70% | -5.74 pp |
| Plefsi Eleftherias | 0.96% | 5.50% | -4.54 pp |
| Foni Logikis | 6.70% | 3.50% | +3.20 pp |
| MeRA25 | 6.22% | 3.10% | +3.12 pp |
| Other / unmatched | 1.91% | 3.60% | -1.69 pp |
Known limitations
The simulation was completed after the referenced poll had already been published. It is therefore a retrospective diagnostic, not a forecast.
The poll was not held out from the calibration process. July 2026 polling targets, migration rules and qualitative political context were already embedded in the system.
IPF weighting deliberately moves the calibrated topline toward a target distribution. Calibrated closeness cannot be presented as independent predictive accuracy.
The simulation and poll did not use an identical party taxonomy, so unmatched categories had to be aggregated.
This first benchmark evaluates aggregate party error only. Segment validity, uncertainty coverage, temporal stability and cross-model reproducibility remain untested.
A pre-registered prospective benchmark with a genuine holdout outcome
Inputs, model versions, population, prediction and scoring rules must be frozen before the real-world result is available. Only then can Synachron make a defensible predictive-validity claim.
Propose a controlled pilot