What this is, and what it isn’t.
The collaboration
This forecast was built by Claude (Anthropic’s large language model) in iterative collaboration with a Tamil Nadu political analyst. The technical work — data parsing, feature engineering, four-model architecture, backtesting, Bayesian inference, Monte Carlo simulation, documentation — was Claude’s. The domain calls — alliance theses, the 98-AC freeze list, factor magnitudes, F7 conflict overrides — came from the human partner. Neither half would have produced this model alone.
We’re publishing the methodology before the count and the predictions during it because the only way to know if a model works is to commit before the answer is known. Every choice that influenced the prediction is documented in the methodology section.
What this is not
This is not an exit poll. No voters were sampled. No fieldwork was conducted. The forecast is a structural model, not an opinion measurement.
This is not a wager market. No money rides on the outcome. The model has a single purpose: to test whether a transparent, replicable methodology — built mostly by AI — can produce a competitive election forecast.
This is not unique to Claude. The same approach — multi-cycle structural modeling, ensemble triangulation, Monte Carlo — could be built by any analyst. What’s unusual is that one human and one AI built it conversationally over a few sessions, and the AI did most of the technical lift.
Tools
Python (pandas, numpy, scikit-learn, PyMC, recharts), TypeScript, Next.js, Tailwind CSS, Vercel for hosting. Data sourced entirely from the Election Commission of India, the Tamil Nadu Chief Electoral Officer, and publicly available demographic surveys. All data files and the source code repository are public.
If we’re wrong
Election forecasts go wrong in known ways. The Random Forest tells us that 60–70% of an AC’s outcome is structural — a high baseline. The remaining 30–40% includes things the model genuinely cannot see: late campaign shifts, enthusiasm asymmetries, weather on polling day, individual scandals breaking in the final week.
On May 4 we’ll publish a predicted-versus-actual comparison. If the model called it, good. If it didn’t, we’ll publish what we missed and why. That’s the deal.
Acknowledgements
ECI Statistical Reports (2001–2024 Assembly and Lok Sabha results). Tamil Nadu CEO 2026 elector roll. GELS 2024 demographic profiles. Lokniti-CSDS for methodological inspiration. Anthropic for Claude. The anonymous TN analyst whose domain knowledge made the freeze list possible.
Questions, errors, missed data? Read the methodology first — most answers are there. Otherwise, the source repository contains contact information for the maintainers.