Skip to main content

Trust Center · AI transparency

AI assistance with a visible authority boundary.

ValorAI uses AI to organize and prepare reviewable work. People remain responsible for reviewing information and choosing next actions.

Updated September 16, 2026

What AI does

  • Summarizes documents
  • Extracts structured information
  • Retrieves relevant guidance
  • Generates reviewable checklists and drafts
  • Identifies possible information gaps
  • Organizes questions and next steps

What AI does not do

  • It does not determine VA ratings or guarantee eligibility.
  • It does not diagnose medical conditions or provide legal representation.
  • It does not autonomously submit claims or make final agency decisions.
  • It does not replace human review or accredited assistance.

Model and provider disclosure

Reviewed source configures Grok models from SpaceX Super Intelligence (formerly xAI) directly or through Vercel AI Gateway, including a moving latest alias and model overrides, and the Jev decision model (typesafe-ai/jev) through Vercel AI Gateway for flag-gated classification and safety checks. This inventory describes code paths, not verified deployed settings or workflow availability.

Grok language and multimodal models

Purpose
Chat assistance, source-grounded guidance, summarization, structured extraction, checklist and draft preparation
Data sent
Prompts and workflow text or document content; this inventory does not verify data minimization or authorization across all workflows
Retention
Subject to applicable Grok API terms and configured controls; this source review does not verify account settings or provider terms
Training use
ValorAI does not authorize uploaded veteran documents or VA-connected records for model training
Fallback
Bounded retry, a configured alternate Grok family, or a clear unavailable/error state depending on workflow
Source disclosure reviewed
2026-10-11

Configured routing for Grok models

Provider
Vercel AI Gateway
Purpose
Routing, provider availability, usage metering, and controlled fallback for selected language workflows
Data sent
Request content required to route the selected AI workflow
Retention
Subject to gateway and downstream provider terms and configured controls; these settings are not verified by this source review
Training use
Gateway routing does not change ValorAI’s prohibition on training with uploaded veteran documents
Fallback
Configured direct provider route or a clear unavailable/error state when allowed by workflow policy
Source disclosure reviewed
2026-10-11

Local privacy classifier

Supported browser assistance flows use Rampart, a local MiniLM classifier with deterministic rules, to replace detected identifiers before sending entered text. Placeholder mappings stay in browser memory so responses can be restored locally. Detection is imperfect, and model failures block assistance sends. Supported UTF-8 chat text attachments up to 64 KB are also filtered locally. When a text attachment is saved to Documents, the saved copy contains the redacted text and a generic document name. Its original values cannot be restored from that saved copy. Keep your original file separately. Other original document uploads follow a separate processing flow and are not covered by this text-redaction protection. Voice and dictation send original audio to the speech provider for transcription before any text redaction can run.

Human oversight

Users can edit, reject, or replace generated output. Extracted facts should be checked against original records. Uncertain or high-risk outputs should direct users to official sources, qualified staff, or accredited assistance. Organizational staff retain final workflow authority.

Source grounding

Answers can use official VA, federal, and configured organizational sources. Source freshness and conflicts are reviewed by content owners; conflicting or uncertain guidance should be surfaced rather than silently resolved. Users can report incorrect information through product feedback or support.

Evaluation program

Document-extraction accuracy
In progress. Fixture-based extraction checks plus planned reviewed samples
Citation quality
In progress. Source presence, relevance, and unsupported-claim review
Hallucination testing
In progress. Adversarial prompts and unsupported benefits-guidance checks
Safety testing
Implemented and expanding. High-risk request boundaries and refusal behavior
Benefits-guidance review
In progress. Official-source and domain-review workflow
Accessibility testing
In progress. Keyboard, screen reader, zoom/reflow, reduced motion, and Axe evidence
Bias and disparate-impact review
Planned. Cohort and workflow review before consequential organizational use

No unsupported numerical performance score is published.

AI incident reporting

Output feedback and safety reports are triaged by severity. The response process supports investigation, model/configuration rollback, affected-workflow suspension, corrective evaluation, and public notice when an incident materially affects users or requires disclosure.

Known limitations

AI can omit details, misunderstand documents, cite stale guidance, or generate plausible but incorrect text. Scans, handwriting, poor images, uncommon record formats, conflicting sources, and policy changes can reduce reliability. Every consequential output requires human review.

Discuss a responsible deployment with ValorAI.

Tell us the mission workflow, users, data types, and review requirements you need to evaluate.

Valor AI LLC · UEI R49VCKHH9N95 · CAGE 22Y86