Surveon finishes every survey with synthetic completes that are flagged, labeled and statistically transparent. The same model previews a survey before it fields, answers the question you didn't ask, and finds the headlines in a fresh upload.
Current model: large-v6 · 298M parametersCost per completed interview. Panel and phone figures are what Marshall pays today.
700 real completes are in. The last 300 are the hard ones: the small county, the age cell, the ethnic quota the panel can't fill. Finish them with…
The field window runs as it does today. Synthetics fill what is still open at the deadline.
At any point, you can see how accurate a given number of synthetics will be.
Every synthetic carries the isSynthetic flag in the data file. See synthetic, authentic or the composite.
Surveon runs on our own hardware. We adjust the weights and train on new data as needed.
Not a chatbot, and not a general model with our data pasted in. A language model that has read every Marshall survey since 2014 in the survey's own format, and learned what a market's numbers look like and how they move from year to year. Tested blind on 24 surveys it never saw.
large-v6 · 298M parameters · 24 held-out surveys
All answers ±2.0 · correlation 0.99
The honest summary: on a repeat question, Surveon behaves like a 175-person survey of that market. Far too loose to replace fielding. Plenty tight enough to finish a survey, catch a bad upload, or steady a small cell.