We never take a model at its word.

CASEY navigates, applies and follows through for people who cannot afford a mistake. The safety is not in the model. It is in the harness around it.

SOC 2 Type 2 Independently audited

CASEY

Text message · 6:15pm

  1. Last step for the childcare subsidy: the program needs proof of income.
  2. i have pay stubs, bank statements and my taxes. send everything?
  3. Just your latest pay stub, nothing else. I'll delete it when the application is decided.

Pay stub deleted at the decision.

Four gates.

What any model inside CASEY may say, ask, send and keep. Enforced in code, not by a prompt.

  1. 01 Grounding

    Every fact traces to a provider.

    What CASEY says about a program comes from the vetted provider that runs it. If the provider did not say it, CASEY does not say it. No ads, and no sponsored answers.

  2. 02 Minimization

    Only the data the task needs.

    CASEY asks for a fact about the person or a document, at the step that needs it rather than at the start. Nothing more is taken.

  3. 03 Consent

    Sent only when the person agrees.

    The model cannot send anything. CASEY shows the person what will go out and to whom, and waits for their answer. A send that differs from what was shown is blocked.

  4. 04 Deletion

    No document outlives its task.

    Every document is deleted when its task completes. The audit trail stays: what was read, asked, allowed, sent and deleted, in order.

One electric bill, through all four gates.

CASEY

Text message · no app required

  1. Heads up, the electric company letter you uploaded last week says the power goes off on the 15th unless the bill is paid. A local hardship fund can cover it. Want me to apply?
  2. yes what do they need
  3. Proof of income and the letter, which I already have. Your latest pay stub covers the income. Upload a photo of it here: Upload your pay stub textcasey.com/u/4k7m2x
  4. Got it. Your income is under their limit. Okay to send the stub and the letter to the fund?
  5. yes
  6. Sent. They usually answer within a day and pay the electric company directly. I'll tell you as soon as they do.
  7. Approved this morning. The fund paid the bill in full. The power stays on.

System activity

  • Letter $312 owed, power off on the 15th read
  • Pay stub IMG_4127.jpg, secure page received
  • Stored Humanitas systems, not the thread encrypted
  • Consent stub and letter, to the fund granted
  • Sent hardship fund, 2 documents delivered
  • Employer no documents, no conversation private
  • Outcome $312 paid to the electric company confirmed
  • Documents stub and letter, task closed deleted

Complete. Nothing else was collected.

Why a one-time link

A text passes through the carrier and stays on the phone. So the thread carries the conversation only, and documents go by a one-time link that expires in 24 hours.

Who holds a copy

Humanitas's audited systems, until the task completes and the copy is deleted. The provider it was sent to, as with any application. No one else.

Every lab is building a smarter model. We build the layer that finishes a task with any of them.

A wrong answer can be caught. A wrong send is acted on.

The same four gates stand between CASEY and every institution it reaches: a hardship fund on paper, a utility on a portal, an employer on email. They do not depend on which model is inside, so each step runs on the smallest one tuned to do it accurately.

Who sees what.

Person

Everything. Any detail CASEY holds about them, shown on request and corrected on request. They can leave at any time, and everything is deleted.

Employer or public agency

The aggregate. Need by cohort, where it goes unmet, which pathways work, and what they return. De-identified and released only above a cohort size set by re-identification risk. Never the record of a conversation, and never used to decide what anyone qualifies for or pays.

Provider

One request at a time. The need, the timing, the funding, and the requirements CASEY has already gathered. The decision on it stays theirs.

What CASEY learns about one person serves only that person. What it learns across many improves it for all.

The controls.

Models
Chosen for the task, open and closed, each under a zero-retention contract: the model reads the request, stores nothing and trains on nothing.
Audit
SOC 2 Type 2, covering every control below. Additional frameworks are scoped to what your programs require, as part of pilot design.
Data
Encrypted at rest and in transit. Labeled by sensitivity.
Access
MFA on critical services. Production access restricted and audit logged.
Endpoints
Anti-malware and encryption on every device, under mobile device management.
Vulnerabilities
Scanned and patched. Penetration tested, findings remediated.
Infrastructure
Web application firewall, infrastructure as code, changes reviewed. Backups automated and recovery plans tested.
Incidents
Incident response policy established. Infrastructure and network monitored.
Vendors
Inventoried and classified by data sensitivity and risk.
People
Security training, confidentiality agreements, onboarding and offboarding process.
Open the trust center

Questions or a security report: [email protected]