Editorial Policy
How forecasts and commentary are produced at Datum Signals. The short version: the model runs, the model speaks. We do not override its output to match a preferred narrative.
How forecasts are made
The composite score and the race-level forecast probabilities are computed end-to-end by code. The pipeline is: ingest behavioral indicators, normalize against the other fifty states, combine via calibrated weights, project against each forward-looking race. No human editor adjusts the probabilities before publication. If the model says Wisconsin governor is 63% Democratic, the site says 63%, even when a human reader might disagree.
Candidate names, photos, and bios are a separate pipeline from the win probability itself: a weekly automated check against public sources (currently Ballotpedia) that flags any mismatch for a human to confirm before it's published, and never feeds the model's probabilities.
What we publish
Model output (composite scores, race probabilities, dimension breakdowns, candidate-fit rankings). Methodology disclosure (every coefficient, every Brier score, every retraining date). Source attribution (every indicator, every publisher, every refresh date). Corrections when we get something wrong. None of it is punditry.
What we do not publish
- AI-generated headlines or commentary. Claude assists with editorial drafts; the final voice is human-written and human-signed.
- Forecasts derived from our private preferences. The model coefficients are fit to historical outcomes, not to ideological priors.
- Race calls. We report probabilities, not winners. A 75% probability is not a call; it is a 75% probability.
- Anonymous opposition research. Everything published links back to its public source.
Conflicts of interest
Datum Signals is operated by Parallax Advisory LLC. Datum Signals data is licensed to VoteROI.com, a donor intelligence platform built on FEC records. That licensing relationship is commercial only: the model has never been adjusted to favor a candidate, party, or licensee objective, and no licensee has input into model design or model outputs.
Corrections
When we get something materially wrong, we say so publicly at /corrections with the date and the fix. We never silently rewrite a model output to look more correct after the fact; forecasts are preserved as issued, so any historical call can be checked against what actually happened.