saved
Introducing Clairic, a new System One class of extraction models
Hraness wrote this summary from a saved copy of the source. Quotations are taken word for word from the source.
gist
Cloudflare previews Clairic, a model on Workers AI that pulls typed values out of text and images in one forward pass, with no output tokens. It returns extracted spans, computed numbers and dates, estimates with an 80% interval, and forecasts. In Cloudflare's damage-claim test, chaining Clairic, its Clef decision model and GLM-5.3 took 19.3 seconds against 61.7 for GLM-5.3 alone, and cost $6,032 per million claims against $11,740.
ideas
- Extraction joins decision models. Cloudflare frames Clairic as a new kind of System One model beside TypeSafe's Jev and its own Clef: those pick among listed options, while Clairic answers with values.
- Every answer is typed or null. Fields declare a type such as text, number, date, duration, list, estimate or forecast, and each answer carries a confidence and a null probability, with character spans for extracted text.
- Estimates come as distributions. Given five comparable home sales, Clairic estimated a sale price with median $415,000 and an 80% range of $370,000 to $470,000, inside the bounds the caller set.
- Forecasts were roughly calibrated. Over 21 days of next-day TSA checkpoint counts, its forecasts were off by 3.3% on average and 17 of 21 actuals fell inside the 80% interval, with misses clustered on the Columbus Day weekend.
- The checkpoint has limits. Cloudflare says multi-step arithmetic, estimates without evidence and long-horizon forecasting are weak spots, and it plans to release open weights after more work.
quotes
“Extraction models turn raw input into typed values, decision models turn values into decisions, and LLMs turn decisions into words.”
“The architecture makes every answer well formed, not necessarily correct.”