Daatan is a suite of tools designed to bring objective, mathematical measurement to public predictions — turning the unaccountable world of expert opinion into a scored, ranked, and queryable dataset.
For decades, Israeli journalists, analysts, and commentators have shaped public understanding of some of the most consequential events in the region — elections, wars, diplomatic breakthroughs, economic shocks. Yet there is no mechanism to track whether they were right.
A columnist who confidently predicted the judicial reform would collapse, a security correspondent who dismissed the Hamas threat as managed, a financial analyst who called the Moody’s downgrade “alarmist” — none of them are ever scored. Their reach and reputation are determined by their platform, not their track record.
This is not only a problem of explicit predictions. Much of the signal is hidden: in the framing of a headline, the certainty of a source quoted, the sentiment underneath a neutral-sounding analysis. Daatan extracts and scores all of it — not just clear forecasts, but implied direction, vague sentiment, and the story being told between the lines.
Daatan addresses this problem through three interconnected products:
Daatan News — A news portal covering Israeli and regional media. Readers don’t just consume news: they engage with it by joining forecasts, staking positions, or flagging claims as wrong. This gamification layer passively generates reputation data — every interaction is a signal about source credibility, by topic, by domain, over time.
Daatan Forecast — An active prediction market where users, analysts, and bots make explicit forward-looking claims and stake reputation on them. Forecasts are resolved against real-world outcomes, building verifiable track records.
Retro Analysis (בדיעבד) — The retroactive audit engine. Rather than waiting for sources to opt in, Retro processes the public trail of what media already published — and scores it. It bypasses the cold-start problem entirely: reputation scores are generated from history, not from future participation.
Retro Analysis is built around The Factum Atlas, a definitive map of historical reality against which all media claims are measured. It cross-references major events against media outlets, journalists, and think tanks.
For every article in scope, a multi-stage LLM pipeline extracts a forensic prediction vector:
Each prediction is matched to a seed event, scored against historical ground truth, and used to update the source’s Brier Score (calibration accuracy) and ELO rating (relative rank).
The resulting dataset — thousands of scored predictions, ranked sources, and verified outcomes — becomes the training corpus for The TruthMachine.
Because we mathematically understand how past language correlates with actual outcomes, The TruthMachine can read a cluster of articles published today and ask: given who is saying what, and how accurate they have historically been, what is the probability this event occurs?
This methodology, Outcome-Supervised Narrative Calibration (OSNC), allows the TruthMachine to identify the specific “linguistic signatures” of accuracy.
Output examples:
"Chances Netanyahu survives no-confidence vote: 0.83""Probability of IDF ground operation in Lebanon within 30 days: 0.61""Likelihood of Moody's second downgrade by Q2: 0.44"This shifts Daatan from a media-auditing tool into a forward-looking intelligence product. Primary clients:
Daatan’s goal is to create a definitive reliability layer for the information ecosystem — starting with Israel. By objectively measuring who was right, who was wrong, and who saw it coming when no one else did, we surface hidden analytical talent, penalize confident noise, and provide the most honest probabilities available for the region’s most consequential events.
Hybrid architecture: High-volume filtering uses cheap or free models (Nemotron Nano); nuanced forensic extraction uses mid-tier models (DeepSeek V3.2 at $0.25/1M tokens). Heavy models are called only when necessary.
Multilingual extraction: The pipeline natively processes Hebrew and English, capturing signals in Israeli media before they surface in international coverage.
Scope: Every article that passes the prediction filter is processed — not just op-eds with explicit forecasts, but news analysis, commentary, sourced claims, and any text that implies a directional view on a future outcome.
Brier Score — measures calibration: how close the extracted probability was to the actual binary outcome. Lower is better.
ELO Rating — zero-sum relative ranking: sources that correctly predict events their peers got wrong absorb significant rating gains, instantly surfacing hidden analytical alpha regardless of platform size or brand recognition.
Compute: AWS EC2 t4g.small (Graviton, ARM) — free-tier eligible through 2026. 40% better price-performance than T3 for this workload.
Storage (MVP): Filesystem JSON — one file per event/source pair (~3,000 files). Migrates to PostgreSQL + pgvector as data volume and API demand grows.
Orchestration: OpenClaw / custom Python async workers.