Institutional Trust

Sheet 1 of 3·Failure modes

Seven failure modes. Seven stops.

Agent systems fail in known ways: they reach data they shouldn't, invent a figure, run on, or act without asking. Each row names one failure, the control that stops it, and where on the path that control sits.

The HarnessDWG 004 · 1/3
Sheet
Failure modes
Drawn by
AgentSmyth
Checked by
Your risk team
01RegisterFailure · control
Pick a failure to see where it stops.
Configured·Platform code, prod2 Nov 2026
02Where it stopsOne question, top to bottom
The Harness
You ask
  1. ScreenStops here
  2. Tools
  3. Agent
  4. Sandbox
  5. Sources
  6. Approval
  7. Record

Each message is screened before the agent starts. Requests to run unrelated computation, look inside the sandbox or copy data out of it are refused.

Configured·Platform code, prod2 Nov 2026

Sheet 2 of 3·The wall

Every tool call that leaves is checked and logged.

Run the Harness inside your network and the agent, your sign-in and the run record live there with it. When the agent needs market data, a tool call comes out to us carrying the tool, its parameters and your user ID. Before it leaves, the parameters are checked for personal data and the call is logged with its verdict. What comes back is inspected before the model reads it.

The HarnessDWG 004 · 2/3
Sheet
The wall
Drawn by
AgentSmyth
Checked by
Your risk team
01One call, out and backIn-network build
AgentChecksYour wallMarket data
Tool call
Egress checkPassed · Logged
Tool · parameters · user ID
Data and sources
Inspected
To the model
Tool call · 09:41:07Passed
tool
Price history
ticker
NVDA
from
2026-01-02
user
u-4821
Configured·Platform code, prod2 Nov 2026
02Egress logEvery call, its verdict
TimeCallVerdict
Pick a call to send it.
Depiction·Illustrative run
Security

Sealed keys. Boxed code.

Sign in with SSO through your identity provider. Data is encrypted at rest with AES-256, and connections to the platform use TLS 1.2 or later. Credentials you connect are sealed with their own KMS data keys, and code the agent writes runs in its own microVM.

  • SSO: SAML THROUGH YOUR IDENTITY PROVIDER
  • ENCRYPTION: AES-256 AT REST, TLS 1.2+
  • AGENT CODE: ITS OWN MICROVM
Privacy

Your data, not our training set.

We don't train models on customer data, and under our DPA the model providers we use may not either. Run it inside your network and tool calls come out to us carrying the tool, its parameters and your user ID, screened for personal-data patterns before they leave and logged.

  • NO TRAINING ON YOUR DATA
  • EGRESS SCREEN: PERSONAL-DATA PATTERNS
  • OUTBOUND CALLS LOGGED WITH THEIR VERDICT
Model Governance

Evidence in. Citations out.

Answers are built from what the data tools return and cite it inline, back to the source documents. Where a deployment sets entitlements, a tool someone isn't entitled to is never built for their run. Every message is screened before the agent starts.

  • CITATIONS BACK TO SOURCE DOCUMENTS
  • ENTITLEMENTS SET PER DEPLOYMENT
  • INPUT SCREEN ON EVERY MESSAGE
Reliability

Built ahead. Routed around failures.

Our cloud is committed to 99.5% availability and runs across availability zones. If a model provider fails, the call moves automatically to the next model in line and the failed one sits out for 90 seconds. Scheduled reports are built before you open them. Every run has a step limit, and every data call a time limit.

  • Scheduled reports are built ahead of time, so they open as soon as you ask for them.
Compliance

Audited controls. Provenance on every run.

An independent auditor's SOC 2 Type II report covers Security, Availability and Confidentiality from May 1 to July 31, 2026, with no exceptions noted. We hold ISO/IEC 27001 and ISO/IEC 42001 certification. Our DPA is written for regulated firms: GLBA and Reg S-P safeguards, 72-hour breach notice, and examination by your regulator where the law requires it. Every run is stored with its question, its answer, its sources and their citations.

  • SOC 2 TYPE II: NO EXCEPTIONS NOTED
  • WEB APP PENTEST: 2026, NO CRITICAL OR HIGH OPEN
  • RUN RECORD: QUESTION, ANSWER, SOURCES
Deployments

The same Harness, hosted or in your network.

Start hosted, or run the agent, your sign-in and the run record inside your network, with market data coming from us through checked, logged tool calls.

Our cloud

Hosted by us on AWS in the US, across availability zones. The fastest way to start.

Your network

The agent, your sign-in and the run record on your Kubernetes, from our Helm chart.

Your models, our data

Model calls can go through your own gateway and contracts. Market data comes from us in tool calls, each screened and logged.

Sheet 3 of 3·The record

Every run leaves a record.

Each run is stored with its question, its answer, the sources the tools returned and the citations that tie them together. A change to your book waits on a card until you accept it, and a linked brokerage account can be read but not traded.

The HarnessDWG 004 · 3/3
Sheet
The record
Drawn by
AgentSmyth
Checked by
Your risk team
01Run recordOne run, as stored
Run record · run_7f3a91Stored
Question
How exposed is the book to a 50bp rise in the 10-year?
Answer
Down 2.1%, mostly through duration in the bond sleeve.
Sources
  1. [1]10Y yield · daily · 04 Nov 2026
  2. [2]Holdings · Core Book · 04 Nov 2026 close
  3. [3]Bond sleeve · durations · 04 Nov 2026

Each document as the tool returned it, with its as-of time and IDs.

Run
Its ID, when it ran and how long it took.
Pick a citation to see its source.
Configured·Platform code, prod · values illustrative2 Nov 2026
02ApprovalYours
Book change · proposedCore Book
  • Add200 NVDA at 118.40
  • CloseTLT

Review and adjust the quantities and prices, then Accept. Nothing is committed until you do.

Waiting for you
Brokerage link

Reads holdings and transactions. Can't place an order.

Configured·Platform code, prod · values illustrative2 Nov 2026

Frequently Asked Questions

Your environment. Your controls. The same Harness.

Ready when you are

On request: our SOC 2 Type II report, under NDA, and our DPA.