LLM-based agentic forecasting system for binary questions. Supports backtesting on AIBQ2 (Metaculus) and ForecastBench (markets + datasets), plus live competition submission.
docs/workflow.md— full pipeline walkthroughdocs/TODO.md— tracked TODOs for paper and codesrc/MODULES.md— categorized guide to all source modulesdocs/shrinkage.tex— logit-space shrinkage derivation../nips26/main.tex— NeurIPS 2026 paper draft (moved to top-level)
src/— all Python source (seesrc/MODULES.mdfor module guide)data/— inputs: questions, exams, tags, cachesexperiments/— xid definitions + outputs (forecasts, eval)- Configs use hyphenated directory names (e.g.
flash-high-brave-c1-t1); delta strings (e.g.flash/thk:high/crowd:1) are used on the CLI but resolved to directory names internally