How Scout handles your data

VisionFI isn't running things on your data. Your workstation is running things on your data, using software and instructions VisionFI provides. The application and the skills are the rules of the game; the workstation does the work, on the workstation, with the documents the workstation already has.

Scout Harness is the application installed on your workstation. It does the document processing — reading files, extracting fields, applying your QC rules, producing the result.

Scout HQ is the control plane hosted by VisionFI. It provides Harness with the skills (instructions) for processing each kind of document, manages release intelligence, and receives back non-content telemetry. HQ does not see document content.

Yes. Self-hosted HQ — where HQ runs in your own infrastructure — is a supported deployment option. The default Desktop Data Sovereign deployment uses VisionFI-hosted HQ; either choice keeps Customer Content on the workstation.

Documents live on your workstation. They are read in place by Harness, sent transiently to Anthropic for AI reading under Zero Data Retention, and the result is written back to your output folder.

Documents only ever exist in two places: your workstation, and (transiently) Anthropic under ZDR. There is no cloud upload step, no VisionFI-side cache, and no document storage on VisionFI infrastructure.

VisionFI receives a defined, enumerated set of non-content operational metadata called Telemetry. The full schema is published in the VisionFI Subscriber Console and reproduced as Appendix A to the Data Handling Addendum. Permitted Telemetry fields include:

• Transaction counts and timestamps
• Processing durations
• Model identifiers and version identifiers
• Aggregated confidence-score distributions (histograms, not per-field values)
• Categorical exception codes (e.g., document_unreadable, schema_mismatch)
• Inference provider identifier and region
• Harness version identifier
• Deployment Environment identifier (an opaque ID, not hostname or user)

Separately, VisionFI processes Administrative Data you provide to administer the relationship: administrator names and business email addresses, billing contacts, license-key holder records, and support ticket content.

By schema and by contract, Harness-to-HQ traffic cannot include:

• Document content of any kind
• Member or borrower names
• Account numbers, loan amounts, transaction dollar values
• Government identifiers (SSNs, ITINs, driver's license, passport)
• Addresses or contact information
• End-user login credentials

Inference call payloads are sent directly from Harness to Anthropic and do not pass through VisionFI or HQ. The Data Handling Addendum enumerates these categorical exclusions in §3.5.

The architectural guarantee is independently verifiable using standard tools your team already has — endpoint detection and response (EDR), host-based firewall logging, perimeter network monitoring, or a corporate web/SSL proxy.

From a workstation running Harness, you should see exactly two egress destinations:

• The VisionFI HQ endpoint — orchestration and non-content telemetry
• The Anthropic API endpoint — inference traffic, transient under ZDR

There is no third destination to which document content is sent in identifiable form. The Telemetry schema is versioned; any material change requires at least 60 days' written notice under DHA §3.1, with an opportunity to object.

No. VisionFI does not, and will not, use your data — of any kind — to train, fine-tune, calibrate, or benchmark any machine-learning or AI model. This applies to Customer Content (which we don't receive), Telemetry, Administrative Data, Opt-In Feedback (if elected), and any sample obtained through a clean-room engagement.

This is enforced two ways: architecturally, because we cannot train on data we do not possess; and contractually, in §5 (Negative Covenants) of the Data Handling Addendum, which prohibits training even on de-identified data without your express written consent.

No. All inference runs against Anthropic under Zero Data Retention (ZDR). Anthropic processes the request, returns the answer, and keeps no copy. They do not train on the inputs or outputs.

Two paths are supported, and both require ZDR:

• Anthropic via VisionFI's first-party ZDR entitlement (our contractual posture flows through), or
• Your own Anthropic API key under your direct agreement with Anthropic, with the requirement that your agreement include ZDR.

The data-handling posture is identical; the difference is only who holds the Anthropic commercial relationship.

The improvement happens in the skill layer, not the model layer. Skills are software artifacts authored by VisionFI — versioned, signed, and shipped to Harness through HQ the same way any enterprise software receives an update.

VisionFI knows what to improve from three sources, none of which carry document content:

• Telemetry — categorical exception codes reveal which document types are throwing surprises across the customer base.
• Industry monitoring — GSE form updates, state DMV releases, IRS form revisions, watched directly by our product team.
• Opt-In Feedback & clean-room engagements — only with your express, written consent.

When a form changes, we publish an updated skill, HQ pushes it to your Harness, and the next document of that type is read with the new instructions. No retraining cycle, no reconfiguration on your end.

If you choose to enable it in an Order Form, the workstation transmits structured accept / reject / review-required signals to HQ, identified by an opaque finding identifier and an opaque reviewer identifier. It contains no document content and no Personal Data, and it remains subject to the §5 prohibition on training without your express written consent. It can be disabled in the Subscriber Console at any time, taking effect within one business day.

Email sales@visionfi.ai to request the DHA, the Telemetry schema (Appendix A) in machine-readable form, or our SOC 2 Type II report under NDA.

Sub-processors are enumerated in DHA §7. The two that touch the Scout deployment are:

• Anthropic — foundation-model inference under Zero Data Retention.
• Microsoft Azure — hosting for Scout HQ and supporting tooling. Azure does not process Customer Content; HQ is built by schema not to receive it.

Notification, cooperation, and scope are governed by DHA §9. Because Customer Content does not flow to VisionFI, the realistic incident surface on our side is limited to Telemetry, Administrative Data, and the control plane — and the contract is explicit about how each is handled.

DHA §10 covers audit rights, including SOC 2 Type II evidence and regulator access. Reach out to your VisionFI contact to schedule a review or request artifacts under NDA.