Three deterministic integrity checks running against the USCIS public data surface.
eb5-engine ingests the USCIS regional center registry plus the USCIS Data Hub processing-time data. Three production checks fire on a single deterministic pass: terminated regional centers that are still showing processing activity, quarterly form-level denial-rate spikes against the rolling baseline, and long-pending median processing-time anomalies above the 36-month statutory reasonableness threshold. Each finding carries the regulatory basis (INA section 203(b)(5), EB-5 Reform and Integrity Act of 2022, USCIS Data Hub processing-time methodology) and a content-addressable evidence hash for replay.
USCIS public data, two feeds, court-defensible references.
Source 1: regional center registry. The USCIS Immigrant Investor Program Office maintains the public list of approved + terminated EB-5 regional centers at uscis.gov. Each entry carries the regional center name, USCIS ID, state, principal address, and current status. Terminations are dispositive: a terminated regional center cannot file new I-526 petitions and pending petitions associated with the center lose their statutory basis under INA section 203(b)(5).
Source 2: USCIS Data Hub processing-time data. Per-form per-quarter receipts, approvals, denials, pending counts, and median processing months are published at egov.uscis.gov/processing-times/data. The Data Hub also exposes a JSON endpoint (e.g. /api/processingtime/I-526/IOE); the ingest script pulls the JSON when reachable and stores raw payloads in eb5.processing_metrics.raw_payload.
What we ingested. Two tables under the eb5 schema: eb5.regional_centers (name, USCIS ID, status, termination date, raw payload) and eb5.processing_metrics (fiscal year, quarter, form type, category, decisions and pending counts, median processing months, raw payload). Form types covered on the seed: I-526, I-526E, I-829, I-924.
Reality on the wire. The USCIS public-data surface is hardened against anonymous non-browser clients in many cases. The ingest script attempts the live pull, records the live attempt in raw_payload, and falls back to a clearly-labeled synthetic baseline (raw_payload.source = 'synthetic-fallback') so the engine has signal to find on a fresh install. Live and synthetic rows coexist; the engine treats them identically and the audit trail makes the source unambiguous.
What ships when an EB-5 oversight investigator engages.
Each check runs deterministically against the ingested USCIS data and produces sealed CREB output through the same orchestrator and Ava layer that powers the rest of the platform. Customer-profile gated on lob = 'eb5_oversight_investigator'.
eb5_terminated_active
Detects regional centers whose USCIS status is terminated yet whose name still appears in recent processing-metric activity. Activity within the last 365 days is dispositive: a terminated center cannot file new I-526 petitions, so continued processing points to either a stale USCIS register, an unauthorized successor, or an integrity issue at the petitioner-intake level.
INA 203(b)(5)RIA 2022SEVERITY high
eb5_denial_rate_outlier
Walks every (form_type, category, fiscal_year, quarter) cohort in chronological order and flags any quarter whose denial rate exceeds 2x the rolling 4-quarter baseline. A spike above the rolling baseline is a leading indicator of USCIS policy change, fraud-pattern detection, or systemic decline in petition quality from a specific source channel.
USCIS Data HubRIA 2022 fraud reportingSEVERITY high
eb5_processing_anomaly
Flags any (form_type, fiscal_year, quarter) where the median processing time exceeds 36 months. Median above the 36-month threshold breaks the practical reasonableness standard underlying INA section 203(b)(5) and the USCIS service-level commitment. Petitioners exposed to this cohort have a Mandamus-grade administrative-delay claim.
INA 203(b)(5)Mandamus precedentSEVERITY medium
What the customer takes to a regulator.
A representative finding rendered as a sealed CREB record. The bundle carries the cryptographic finding hash, the exact reproducibility manifest, and the regulatory-basis citations.
Deterministic, reproducible, court-defensible.
Deterministic
Each check is a SQL aggregate or a deterministic walk over the ingested USCIS data. Same registry snapshot + same processing-metric ingest + same thresholds, same findings every run.
No external LLM
The Tier 1 verdict path is rule-based. Ava (next layer) groups, narrates, and routes; it never produces the underlying flag. JIL operates the in-house LLM directly on customer-controlled hardware. No OpenAI, Anthropic, or Vertex API.
Replay manifest
Every CREB carries the source-data pointer, code version, schema migration hash, query plan, and signal thresholds. A third party with the same inputs replays the analysis bit-identically.
One kernel. Eight industries. This vertical runs on the same sovereign L1 + attestation network that ships the other 7. Kernel age: 18+ months. Adding a vertical: ~1 week. Competitor moat: build the kernel first.