Skip to content

DATA-CARD.md

Latest commit

 

History

History
89 lines (73 loc) · 4.59 KB

DATA-CARD.md

File metadata and controls

89 lines (73 loc) · 4.59 KB
Raw

Data card: what crosses the optional Bedrock extraction seam

This card documents the data that leaves the machine when the optional cloud seam is enabled, who receives it, what governs it once it has left, and what never leaves under any configuration. It is the companion to the model card. Source of truth for every claim: src/constituent_reconciler/extract/seam.py, src/constituent_reconciler/pipeline.py (read_pdf_records), and the [extract] recipe section in src/constituent_reconciler/config.py.

What crosses when the seam is enabled

One page at a time. A page from a constituent intake PDF is offered to the seam only when its offline extraction confidence falls below the recipe's confidence_threshold (default 0.5) and the seam's gates all hold. The design sends that page as an image for field extraction.

An intake page can contain any personally identifying information the form collects: names, dates of birth, home addresses, phone numbers, email addresses, household details, consent notes, case information. Assume the whole page crosses, because it does: the selected page is rendered to a 150-DPI PNG without redaction or cropping. The request also includes a fixed instruction listing the six allowed output fields. Bedrock returns extracted field values and confidence scores, so response content can contain the same PII. The default backend remains none; enabling bedrock under a permissive policy is an explicit egress decision.

Who receives it

Amazon Bedrock, in the AWS region the deployer's own AWS configuration selects, under the deployer's own AWS account. The receiving model is Claude (built by Anthropic), served by AWS; the default model id is us.anthropic.claude-sonnet-4-6. This project holds no keys and operates no service; the AWS relationship belongs entirely to the deployer.

Retention and processing once data has left

Handling of a page after it leaves the machine is governed by the deployer's AWS agreement and the AWS Bedrock service terms, not by this project. Once a page image has been sent, this project cannot enforce deletion, retention limits, or access controls on the other side.

Before enabling the seam, a deployer should review the AWS controls that shape what happens to the data, in particular Bedrock model invocation logging (an account-level setting the deployer controls, which if switched on stores request and response content in the deployer's account) and the AWS documentation on how Bedrock handles prompts. Verify the current terms yourself; this card intentionally does not restate them, because they are AWS's to change.

What never crosses

  • Anything under the dv or hipaa policy packs. make_seam() returns a NoOpSeam for those packs at construction time, regardless of the recipe's backend setting, so no page image or field value leaves the machine. The merge-blocking tests tests/test_no_egress.py and tests/test_extract.py assert this.
  • Pages at or above the confidence threshold. They keep their offline extraction result and are never offered to the seam.
  • CSV sources. Only PDF pages route through extraction; rows read from CSV files never touch the seam.
  • Every other stage's data. Normalization, matching, the review queue, the provenance log, and the export all run locally whether or not the seam is enabled.
  • Telemetry content. GenAI telemetry includes model/provider identity, token counts, duration, finish reason, and estimated cost. It excludes the page image, prompt, response, field values, and record ids by default, with a regression test using representative PII.

Provenance of cloud-refined fields

When the seam returns fields for a page, read_pdf_records in pipeline.py replaces that page's offline fields with the returned ones. Each returned field uses the same ExtractedField shape as the offline extractor: a field name, a value, a confidence score, and an optional source span pointing back into the document.

The outputs do not currently mark a field as cloud-refined. A reviewer looking at a record cannot tell whether a value came from the offline extractor or from the seam. This is a known gap rather than a designed guarantee: if your audit trail needs that distinction, keep the seam off, or add provenance tagging when you implement refine().

Training data

None exists. This project trains no model, so there is no training dataset to document. The seeded synthetic fixtures in examples/intake-demo/, scored by the committed eval in eval/, contain planted ground truth and no real personal data; they exercise the pipeline in CI and never cross the seam.