Technical Security, Confidentiality & Vertex AI Architecture
Designed for California criminal defense panels, solo trial practitioners, and civil/family litigation teams operating under Bus. & Prof. Code §§ 6068(e)(1) & 6068.1 and CRPC Rule 1.6.
In Draft Master v5.0, production LLM inference is strictly locked to Google Cloud Vertex AI (us-central1 primary / global Vertex failover) via Cloud Run service-account IAM authentication, with zero silent fallback to consumer AI Studio API keys in production.
Under the Google Cloud Data Processing Addendum (CDPA), customer inputs, grounded legal search queries, and outputs are contractually excluded from foundation model training. Per Google's published Vertex AI governance documentation: (1) flagged prompts may be retained up to 90 days for abuse monitoring (with potential human review) unless the Google Cloud account holds an approved Abuse Monitoring Exemption under a Cloud Master Agreement (30 days for Frontier/Advanced AI models); (2) implicit caching is enabled by default while request-response logging is disabled by default; and (3) Grounding with Google Search retains search query logs for 3 days, so Draft Master restricts search queries strictly to abstract legal authorities and pre-scrubs client PII and medical PHI in your browser prior to cloud transmission.
Attached PDFs (including 1,000+ page discovery volumes processed in segmented batches), Word documents, and pasted narratives are pre-scanned locally in your browser before upload. Detected SSNs, dates of birth, California Driver's Licenses / CDLs, bank accounts, minor schools, and substance-rehabilitation PHI are replaced with tracked [INSERT: REDACTED ...] tokens.
Whenever PII is detected in a PDF's embedded text layer—or when an un-OCRed image-only PDF lacks explicit opt-in confirmation—our backend strictly blocks raw PDF binary slices so unredacted visual text never bypasses text scrubbing.
Pass 1 deterministic discovery redaction runs 100% offline in your browser using disableCombineTextItems: true, 50/100-page segmented volume indexing for 1,000+ page productions, and proportional glyph advance measurement. If you opt into Pass 2 contextual entity detection, Pass 1 hits are masked out locally before text windows leave your browser.
On export, every redacted page is rasterized at 2.4x DPI into a clean pdf-lib container while stripping /AcroForm dictionaries, XMP/Info metadata, embedded page thumbnails, and annotations—permanently destroying underlying vector characters on redacted pages.
When verifying citations against the Free Law Project / CourtListener Citation Lookup API, Draft Master transmits only the extracted reporter citation strings (e.g., 13 Cal.4th 497) — zero client names, docket numbers, or case facts are ever shared with external legal indexes.