Compliance is a constraint, not a feature.

“Who saw which patient field, and when?”

Patient.ssn declared with acl: { read: ["compliance_officer"] } returns Null for an analyst’s MATCH (p:Patient) RETURN p.ssn — enforced by the engine, not the resolver. Every denied read appends to acl_events.log. db.aclEvents replays it; db.verifyAclChain proves no record was tampered.

“Patient asks us to forget them. Can you prove erasure?”

Per-subject AES-GCM keys make right-to-erasure mathematical, not hopeful. Destroy the key (db.destroySubjectKey('patient_42')) and every encrypted property for that subject becomes unrecoverable ciphertext — retroactively across every snapshot, backup, and audit replay. db.proveErasure('patient_42') returns a signed receipt your DPO hands the patient.

“What did the patient’s chart say on the day of treatment?”

MATCH (p:Patient {mrn: 'M12345'})
  AT VALID    '2024-03-15'
  AT RECORDED '2024-03-15'
RETURN p

AT VALID asks “what was clinically true that day?” AT RECORDED asks “what was in the chart that day?” — not the same question when a result was backdated. The legal record replays verbatim for malpractice, billing review, or public-health audit.

Research data sharing without DUA pain.

Issue a research collaborator an API key with privacyMode: {epsilon_budget: 1.0, period: "Daily"}. The planner refuses any non-aggregate return — they can compute count, sum, avg for publication, never see individual records. Laplace noise + ε budget + dp_events.log audit replay give the IRB an actual proof.

“Reconstruct the trader’s view at 14:32:17.4”

AT RECORDED '2024-10-15 14:32:17.4' returns the exact graph the firm’s risk engine operated against at that moment — not a reconstruction, the actual state. For SR 11-7 model-risk audit: “the same Cypher query against the same RECORDED timestamp returns the same answer forever” is the property a regulator wants to hear.

“Prove these trade records weren’t modified after submission.”

WORM mode (irrevocable) refuses every mutation except audit events. The hash chain links every committed snapshot to a SHA-256 of its predecessor; GET /graphs/{name}/audit/verify re-walks the chain and reports tamper-evidence. SEC 17a-4(f) “non-rewriteable, non-erasable” — answered.

“Where did this credit decision come from?”

MATCH (a:Application {id: $appId})
CALL db.propertyProvenance(id(a), 'approved')
YIELD value, validFrom, version, txnAt,
      actor, source, reason, confidence
RETURN value, actor, source, reason,
       confidence, validFrom
ORDER BY version DESC

Every property version records who, when, from what source, and with what confidence. Adverse-action explanations, model-risk audits, ECOA fair-lending defensibility — one walk of the provenance log.

“Quants want to compute portfolio statistics on real data.”

A research API key in private mode lets the quant team compute avg(position_size) and count(...) over the firm’s real positions without ever exposing per-position records — the planner refuses individual returns, Laplace noise is calibrated to the dpClamp bounds you declared, and the ε budget gives a math-backed guarantee. Pairs with the inverse: a production-trader API key with full access, gated by capability tokens with short TTLs.

“Show the policy + coverage + claim graph as of the loss date.”

InvariantDB’s hosted demo is an insurance graph — 500K nodes across the full lifecycle. For a rate-filing review, the same MATCH (p:Policy)-[:HAS_COVERAGE]->(c) AT VALID '<filing-effective-date>' RETURN … query reproduces the filing’s exhibit dataset from the live database, weeks or years later. No ETL pipeline; no separate immutable archive.

“Adjuster changed a reserve estimate. What was it before?”

MATCH (r:Reserve {id: $reserveId})
CALL db.propertyHistory(id(r), 'amount')
YIELD value, validFrom, validTo, version, txnAt
RETURN version, txnAt, validFrom, validTo, value
ORDER BY version

“Underwriting bound a risk. Who approved? What was the basis?”

Every binding decision records actor, source (model run id, or underwriter login), reason, confidence. Capability tokens give each underwriter their own time-limited credential — the audit log records which capability authorized the binding. NAIC Model Audit Rule § 6 mandates controls testing. The audit chain is the test artifact.

“Carrier sharing claim statistics with a state DOI.”

Issue the state’s data partner an API key with privacyMode configured. They get the aggregate statistics needed for the rate-review filing without ever seeing individual claimant records.

“This graph never leaves us-east-1.”

InvariantDB’s HA cluster (Raft, 3 nodes default) can be constrained to a single AWS region or sovereign jurisdiction. The graph’s residency schema field is enforced by the cluster’s membership controller. (Status: design complete, implementation in flight. The single-region single-node deployment satisfies most FedRAMP-moderate boundaries today.)

“Every property read by every credentialed user must be loggable.”

Mint a capability token with the agent’s PIV cert principal claim, a session id, an expiry. Every Cypher query under that capability appends a row to the audit chain naming the principal, property, timestamp, result. WORM mode prevents an insider from rewriting that history.

“Suspect-data crypto-shred at end of authorized retention.”

db.destroySubjectKey runs on a cron schedule keyed to the retention table. The receipt (db.proveErasure) goes to the authorizing official.

Publish statistics from sensitive registries without re-identification risk.

Same answer as healthcare / finance. The US Census Bureau pioneered DP for the 2020 decennial; state and local agencies asking the same question get the same answer.