Input

A decision call has one required field, input: the body of the capability you are calling. Everything else is optional.

{
  "type": "statement",
  "input": { "file_base64": "...", "file_name": "march.csv", "profile": { "monthly_income": 3200 } },
  "subject_ref": "cust_1042",
  "subject_consent": true,
  "session_id": "review_771"
}

Type inference

Omit type and Ward infers it from the input:

Input contains	Type
`applicant`	`kyc`
`disputed`	`ato`
`name` with `owners`, `mcc`, `merchant_id`, `observed`…	`merchant`
`files`, or a `.csv` / `.xlsx` / `.tsv` file	`statement`
`mrz`, `claimed`, or an image file	`document`
`session_id` + `intent_type`	`intent`
`amount`	`payment`
`name`	`screen`
only `agent_id` / `agent_did`	`agent`

A PDF that could be either an ID or a statement needs an explicit type (otherwise 422). Send type in production: inference is a convenience for exploration.

Give context, not history

For real-time types (payment, intent, agent, screen) send the event plus a profile summary, not the full history: a cardholder's usual countries and average ticket, not two years of transactions. Real-time calls take up to 32,000 input tokens; file types take up to 1,000,000 per call. Larger inputs answer 413 call_too_large; split files across calls in the same session.

Files

Files go base64-encoded in file_base64 (with file_name), or as a files array for several statements of one account. Ward reads CSV, TSV, XLSX and PDFs with a text layer in any column layout; see Files in any format. Files are analysed in memory: for documents and KYC nothing is stored; statement analyses keep the result, not the file.

Your references

subject_ref: your customer id. Links decisions about the same customer (Subject memory).
external_ref (inside most inputs): your transaction, applicant or case id, echoed back and written to the audit log.