Documentation / examples/blueprints/verifier/travel-document-evidence/README.md

Travel document evidence Verifier

This reference Blueprint compares one set of normalized travel-form values with one set of already extracted document values. It performs only verification: extraction, normalization, validation, approval, persistence, and routing remain separate operations.

Contract

Six canonical fields are compared in a fixed order. Every extraction result is one strict branch:

• found carries a typed value and page/text evidence;
• not_found carries neither a value nor evidence;
• unreadable carries no value but preserves page/text context.

Each output field is matched, mismatched, or unverified. A mismatch always contains the submitted value, extracted value, and source evidence. An unverified result distinguishes a missing submission, absent source field, and unreadable source. It never turns insufficient evidence into a mismatch.

Strings and other non-numeric values use exact JSON value-and-type equality. Two numeric values may match within the versioned absolute tolerance. No trimming, case folding, parsing, translation, or normalization occurs inside this Blueprint.

The aggregate overall_status uses explicit precedence: any mismatch produces mismatch; otherwise any unverified field produces incomplete; all matched fields produce verified.

The evaluation's first case embeds both travel-document-evidence-v1.yaml and verification-policy-v1.yaml. The offline contract test checks those standalone fixtures for exact drift.

Evaluation

The eight-case dataset covers exact matches, a numeric mismatch, a non-equal numeric value within tolerance, absent source evidence, unreadable evidence, a missing submitted value, mismatch precedence over incomplete evidence, and complete source absence. Its canonical target is 8/8 exact schema and semantic passes:

pixiecore blueprint eval \
  examples/blueprints/verifier/travel-document-evidence/evaluations/travel-document-evidence.yaml \
  --seed=travel-document-verifier-reference-1

The checked-in test uses only public runtime and evaluation APIs with an offline fixture provider. Incidental negative-test data uses the run-wide seed. Live-provider evaluation is opt-in and must preserve provider/model metadata, temperature, seed-equivalent value, timestamps, dataset version, and artifacts.

Known non-supported behavior

• Reading PDFs or screenshots directly; use an Extractor first.
• Parsing amount strings or normalizing dates, names, identifiers, or currency.
• Fuzzy, locale-aware, or case-insensitive comparison.
• Assigning a policy level, deciding whether a mismatch is acceptable, or approving the application.
• Persisting evidence, notifying users, or advancing the workflow.