Buyer Guides
Clairic’s Preview Changes What an Extracted Value Can Mean
Cloudflare’s October 19 preview puts source text, computed values and predictions behind typed fields. Choose the evidence each field must supply before use.
Cloudflare announced Clairic as a hosted Workers AI preview on October 19, 2026. Open weights were described as a later intention, not part of that release. [1] For application buyers, the consequential change is the range of evidence that can sit behind a returned value: the endpoint can identify something stated, derive something from supplied facts, or predict something still unknown.
Before connecting a field to an application, choose its evidence contract: what must support this value, and what may the application conclude from it? A source excerpt, a calculated quantity and a forecast answer different questions. Treating all three as extracted facts can turn a useful prediction into a false record of an event.
One equipment note, three different purchases
Consider a fictional equipment record whose entire note is: “R17: Inspection began on October 18, 2026.” Stipulate October 20, 2026 as the comparison date. There is no completion record, inspection method, clock time or actual QA result. The following questions illustrate application requirements; they are not outputs from a model run.
The first question is which record the note identifies. Returning R17 is a source-identification task. Clairic’s documented text representation includes a value and a span locating text in the input, with confidence metadata. [3] A buyer can require the application to retain the note revision and show the selected characters in context. That makes the selection inspectable; it does not establish that an inspection occurred merely because the note says it began.
The second question is how many calendar days separate the stated start date and the stipulated comparison date. October 20 minus October 18 is two calendar days. The number 2 need not appear anywhere in the note. A numeric field can represent that computed result with value and confidence metadata, without a literal source span for the number. [3] Its evidence is the supplied dates and the specified operation. No times or working calendar were supplied, so this is neither measured operating time nor two completed workdays.
The third question is when the inspection will finish. Neither the note nor the subtraction determines that outcome. If the application asks for an estimate of calendar days from October 20 to completion, the model is being asked to predict an unknown quantity. The documented estimate representation includes a median, p10 and p90. [3] A completion date displayed from that predicted duration would remain a prediction, even if its typography matched the recorded start date.
A missing completion date and an estimated one carry different evidence
Cloudflare’s concepts documentation distinguishes single-value extraction and computation from estimation: null means the model considers the input insufficient, possibly because information is absent or was not identified. Estimates and forecasts can still return predictions with little information. [2]
For R17, a completion-date extraction request therefore needs a way to preserve “not determined.” An estimation request needs a way to preserve “predicted.” A populated prediction cannot resolve whether the original note contains completion evidence. Changing the question has changed the product being delivered.
The model reference also illustrates a missing field whose final value is null while candidate guess and confidence metadata remain present. [3] An interface that substitutes the guess for the null silently changes the result’s meaning. For this buyer decision, the application should keep final values, candidate metadata and predictions distinguishable through storage, display and later reuse. Otherwise, a downstream report can lose the very distinction the preview makes available.
The requested window constrains the answer space
Suppose the fictional application permits estimated completion only from zero to seven calendar days after October 20. That gives the permitted predictions an upper date of October 27. It supplies no evidence that the inspection will actually finish by then. A real completion on October 30 would remain possible; no completion outcome has been stipulated.
Cloudflare explicitly separates output constraints from correctness. [2] Choosing a window because a dashboard requires a date this week does not make that window plausible. The application owner must explain why the permitted range fits the intended task and what the interface means when the available evidence cannot support the requested planning use.
Extending the permitted range to fourteen days changes which predictions can be returned. It adds no inspection progress, staffing information or completion record. The buyer should therefore compare the proposed field definition and its supporting inputs together. A narrower answer space alone is not an improvement in the evidence.
The middle of a distribution is not demonstrated coverage
The documented p10-to-p90 interval describes the middle 80% of the predicted distribution. It is not a guarantee about the actual outcome, and coverage can vary by task and dataset. [2] In the R17 application, displaying that interval would not establish that 80% of comparable inspections finish inside their displayed windows.
That latter claim would need evaluation against relevant completed cases: predictions recorded before outcomes were known, an agreed definition of completion, and a comparison of the intervals with the eventual outcomes. The evaluation owner must also explain which inspection population those cases represent. A result from familiar equipment or an established workflow leaves a changed population unresolved.
The existing guide to training and validation group overlap explains a related evidence question: whether held-out records represent the subjects on which performance is claimed. It cannot supply a coverage result for this preview. Nor should a prediction interval inherit the meaning of uncertainty in a physical measurement report; those are separate models and evidence bases.
Buy validation for the operation the field performs
For source identification, ask the application owner to demonstrate that the selected text belongs to the intended record and remains linked to the relevant input revision. For computation, ask how the inputs, date basis, units and arithmetic will be checked. For prediction, ask what evaluation supports the stated use, including interval misses and cases unlike the evaluation population. These are different demonstrations, even when their results arrive through one endpoint.
Use the schema-validation guide for the separate checks on types, business rules, source context and permissions. Here the prior decision is what the field is supposed to establish. A perfectly typed predicted date is still the wrong deliverable when the requirement is evidence of completed inspection.
Cloudflare documents typed results and operation-specific uncertainty metadata. [1] The buyer’s next step is to assign each proposed field an operation, its supporting evidence, its application check and its unresolved-result handling. For R17, that preserves three useful states: identified record R17, a derived two-day calendar difference, and an unknown completion outcome. A proposed estimate may inform a separately evaluated planning use; it becomes neither a completed inspection nor a supplier promise.
01Footnotes
Footnotes
- Cloudflare, Introducing Clairic, a new System One class of extraction models, October 19, 2026. Hosted preview announcement and intended later open weights; capability descriptions are vendor statements, not independent performance findings.
- Cloudflare Workers AI, Extraction models, updated October 19, 2026; checked October 20, 2026. Operation definitions, missing results, constraints and uncertainty interpretation. Documentation describes intended behavior; it does not establish accuracy or interval coverage for an application.
- Cloudflare Workers AI, Clairic model reference, checked October 20, 2026. Current text, numeric, null and estimate response vocabulary; the documented examples describe response structure rather than evidence for equipment-inspection predictions.
Contact
Discuss Your Infrastructure Project
Tell us about your project, and our team will follow up to discuss the scope, requirements and availability.