How DeepV reviews a UAD 3.6 appraisal

The customer uploads an appraisal payload through the DeepV portal. Accepted formats follow the URAR 3.6 runtime contract: 1004, 1004UAD, 1073, MISMO 2.6 GSE, MISMO 3.x, and TOTAL/alamode PDF. Supporting exhibits and any prior demand correspondence may be attached.

• Accepted: pdf.1004, pdf.1004UAD, xml.mismo.26, xml.mismo.3x, pdf.total_alamode
• Intake preflight runs the 5-gate security model before any review begins
• No console-only success — uploads either land in the intake bucket or fail closed

The intake service normalises the payload, validates the MISMO 3.6 namespaces, and produces an immutable AppraisalState with provenance for every extracted FieldValue. Legacy 1004 / MISMO 2.6 payloads flow through the legacy bridge before review.

• XML and PDF preflight, deskew/dewarp/contrast where needed
• AppraisalState is frozen — version control by content hash
• Legacy bridge maps MISMO 2.6 GSE into UAD 3.6 canonical fields

DeepV parses every UAD field, runs the deterministic rule pack, and emits structured rule outcomes. Rule IDs are stable across runs so findings can be reconciled against prior reviews. Misparses route to Section 9 for reconciliation.

• Stable rule_id and rule_family on every outcome
• Validation engine integrated with the field catalog
• No fallback narrative; structured output enforced

The parser output renders back into a structured form view so reviewers see exactly how DeepV read each field. The render is built from the same FieldValue records the brains consume, so the reviewer view, the audit trail, and the brain inputs stay in sync.

• Form view matches the appraisal grid exactly
• FieldValue provenance is visible per cell
• No silent inference — every read shows its source

Prime Brain orchestrates the canonical 20 DeepV brains across the 21-step URAR 3.6 runtime sequence. Every brain produces structured output for every review — either a finding, a "searched but not found" report with attempted sources, or a "not applicable" status.

• 21-step runtime sequence per DEEPV_RuntimeBundle_v2_FULL_EMBED.json
• Always-active brains; no silent passes
• Output behavior pinned: fully_expanded_if_triggered

Findings roll up into the RISC bar by Risk Band Conductor. Materiality weighting (value-impacting 1.0, compliance-only 0.6, cosmetic 0.3) and severity bucketing (0.90 / 0.70 / 0.40 / 0.15) are pinned from the runtime configuration. Hard fails surface a black-ball indicator regardless of composite score.

• Six-level palette: Low, Mild, Moderate, Elevated, Critical, Unknown
• Lower-is-better composite, 0–100
• Black ball = hard fail; surfaced visibly

Every finding carries evidence references (page, field, image, comparable, supplement). Every brain logs the public-record sources it queried with timestamps — even when it found nothing. LangSmith / glass-box trace IDs are pinned per stage in the audit trail.

• Per-finding evidence: {kind, uri}
• Searched-but-not-found registry with attempted sources + timestamps
• LangSmith trace_id and span_id in the audit_trail block

Report Assembly composes the deterministic JSON report and the human-readable narrative. Audit trail (rule_pack_version, brain_version_ids, prompt_version, model_id, temperature, seed, langsmith_trace_id) is pinned in the report so the same input plus the same configuration produces the same artifact.

• Report JSON schema versioned r79.v1
• Eight required disclaimers
• deterministic_hash excludes generated_at and storage paths

PDF rendering produces one artifact per review_id with SHA-256 pinned. The downloadable copy is gated behind the payment / report-completeness gate inside the portal. The public site shows only the labeled R81 sample for walkthroughs. No live paid report download is implied until a payment provider is configured.

• One PDF per review_id, SHA pinned
• Public-site sample download labeled SAMPLE; not a live paid run
• Live download enabled only when payment provider is wired