Control
The model may organize evidence and prepare work. Coverage, settlement, payment, external communication, filing, and signature stay with an authenticated person.
Trust
Built for a regulated industry, from the database up.
How Sentanis keeps models in bounds and claim data protected: the AI safety and security controls a CISO or CTO will ask about, named and inspectable.
See the mechanismsAI safety and security
AI safety
Prompt injection and hallucination, engineered against.
How a hostile letter is kept from steering the model, and how a fabricated citation is kept from reaching counsel's signature. Every control named, and the limits stated in the specifications' own words.
- One gateway and one immutable policy for every model call
- Documents, tool results, and prior output enter as labeled evidence
- Seven conditions before a citation can be called verified
- Proofs bound to the exact revision counsel signs
Security
Claim data, from intake to decision.
Seven stages from a repository review: who decides, who can open a claim, how it is stored, what leaves the boundary, where generated analysis runs, how a decision is reconstructed, and how failure is detected. Plus what is still being proven.
- Claim-scoped authorization, checked in the application and again in the database
- Field-level encryption and tenant guards in the schema
- A no-network sandbox for generated analysis
- Append-only, content-free audit tied to a named person
The mechanisms, by name
Shipped controls, not policy statements.
The things a security review or a market-conduct exam actually asks about.
- Private deployment
- Runs on the carrier's own infrastructure; claim data never leaves it
- Row-level security
- Postgres RLS plus named permissions and explicit per-claim grants — the database enforces access even if the UI doesn't
- Append-only audit
- Every decision is recorded against an authenticated person, and the access trail is append-only — enforced by a database trigger
- Field-level encryption
- Names, contacts, source records, and document text encrypted at the field level in the system of record
- Untrusted-evidence rule
- Imported documents are evidence, never instructions — a demand letter cannot steer the model that reads it
- Redaction before egress
- PII redaction runs before any external model call sees claim text
- Login admission control
- Account and aggregate rate limits on password work throttle credential attacks at the door
- Audited analysis access
- Analyst and agent data access runs read-only, permission-scoped, and audited — content-free logs, bounded queries
One demand, end to end
Follow a time-limited demand through the controls.
The first segment makes the architecture concrete: a represented auto bodily-injury claim, a hostile deadline, and a response that still belongs to an adjuster or counsel.
Demand arrives
The document enters as untrusted evidence—not executable instruction.
Access is checked
Workspace and claim scope are established before the file is opened or processed.
Assertions are deconstructed
Demand assertions are separated from the medical records, bills, facts, and authority that may support them.
TLD Sentinel starts the clock
The deadline is captured with its source, jurisdictional basis, and the uncertainty a professional needs to resolve.
A response is prepared
Generated work stays linked to the evidence set and cannot send, sign, deny, settle, pay, or file itself.
A person decides
The adjuster or counsel reviews, revises, and authorizes the consequential step.
The record remains
The actor, revision, evidence set, provider, and outcome can be reconstructed without copying claim content into the audit log.
Assurance, honestly stated
What we can show. What we are still proving.
We do not treat a source file as proof of a running environment—or a model-generated answer as proof of a human decision.
Implemented in source
A control or technical gate can be inspected in the product source.
Designed
The intended boundary is specified; implementation or test evidence is incomplete.
Validation in progress
The control exists, but adversarial, route, or integration proof is still being assembled.
Operating evidence required
Effectiveness depends on deployed configuration, logs, people, vendors, or recovery exercises.
Human review begins here
Review the controls on a fictional claim.
Bring claims, legal, compliance, and security. We will trace one demand from intake to a human-authorized response and identify which evidence can be shared publicly or through a protected review.