RUNTIME ARCHITECTURE / 02 NS / SCHEMA

Extract the signal.
Keep the shape.

One encoded request can produce extracted fields and a classification. The runtime checks both outputs against your contract before the workflow moves on.

Read the path ↓01 / Router ↗

FORWARD PASS / EXTRACT + CLASSIFYWORKED EXAMPLE · ILLUSTRATIVE

01 / MISSION

“Please send order 12345 to Maya today.”

EXTRACTorder_idcustomerCLASSIFYintent

02 / ENCODEWords become vectors

Text, fields, and intent labels share the encoded context.

03 / SCORETwo heads, one input

CLASSIFICATION HEAD / INTENT

send_order

cancel_order

other

Spans locate values; the classifier ranks intent labels.

04 / DECODE + CHECKReturn typed values

{ "order_id": { "value": "12345", "score": 0.97 }, "customer": { "value": "Maya", "score": 0.95 }, "intent": { "label": "send_order", "score": 0.92 }}

The runtime validates the extracted fields and intent. Scores are illustrative.

ONE ENCODER PASS → SPANS + CLASS LABEL → CONTRACT CHECKVALIDATED OUTPUT

SCHEMA EXTRACTION / FORWARD PATH

One job.
A clear boundary.

Extraction is a structured perception step before reasoning or execution.

Encode text + labels

The input contains source text, fields to extract, and allowed intent labels. One encoded context can support both tasks.

Score spans and labels

The extraction head scores start/end boundaries while the classification head scores the allowed intent labels from the same encoded request.

Decode and validate

Span scores become field values; class scores become a label. The runtime checks required fields and confidence before execution.

EXECUTION PATHS

Where the waiting happens.

For the same extraction and classification task, these paths do different work before an application can use the result. The visual shows operations, not measured response times.

SAME MISSION / FIELDS + INTENTOPERATIONS, NOT A LATENCY BENCHMARK

Extract + classifyspecialized encoder

encode text + labelsscore spans + intentvalidate

ONE PASS

GPT-5 minigeneral purpose generator

prefill promptgenerate token ↻parse JSON

OUTPUT LOOP

Claude Haiku 4.5general purpose generator