Enterprise AI governance at runtime

Your AI.
Your corpus.
Your rules.

Run open-weight models you control, on premises or in the cloud. TrustHarness is designed to check requests, outputs, and agent commands against your approved documents, policies, and guidelines before an answer reaches a customer or a command runs.

Request a Demo See How It Works

On premises or in the cloud.
Priced per request, capped monthly per user.

Illustrative workflow
Customer response

Your statement is available in the client portal.

Corpus-grounded checks

  • Approved sourcesSupported
  • Policies and guidelinesMet
  • Human approvalNot required
Passes configured checks. Delivered to the customer.Fictional examples. No model is running.

Five questions, answered

Where does the AI run?
In your environment: on premises or in a private cloud dedicated to your organization.
What information governs it?
Your corpus: your approved documents, knowledge, policies, and guidelines.
What gets checked?
The requests, outputs, and agent commands you designate.
Who controls execution and release?
Your organization, through configured checks and human approval.
What evidence is retained?
A signed record of the rule versions, check results, and disposition for each item.

Solutions

One control layer for the open-weight models you run.

TrustHarness is designed to sit between your AI model and the people or systems that rely on it. Your AI proposes. TrustHarness checks each output against your corpus before it is released.

Your AI proposes

Open-weight model

Run through Ollama, with llama.cpp planned.

TrustHarness checks

Runtime gate

Checks each output against your approved rules before it is released.

ReleaseReviseBlockEscalate

You rely on it

People and systems

Receive a checked result rather than raw output.

Conceptual view. Where an output can be revised, it goes back to the model and is checked again.

  • Runtime gate

    Holds each output, checks it against your approved rules, then releases it, returns it for revision, blocks it, or escalates it.

  • Your corpus, your rules

    Designed to turn your approved documents into rules your own reviewers read and approve.

  • Model calibration

    Designed to calibrate to the open-weight model you run, including its internal measurement.

  • Gate API

    Designed to sit on top of your AI, so applications get a checked result rather than raw output.

  • Model connectors

    Designed to plug into Ollama, with a llama.cpp connector planned.

  • MCP server

    Designed to let AI agents ask the gate before they act. Anything short of an authorized answer counts as not authorized.

  • Server deployment

    Designed to run the gate as a service on your own server, reachable from other machines only over TLS.

  • Open-source edition

    Planned

    A self-hosted edition you can run yourself, with security controls you can inspect.

Open-weight models are models whose weights you can run yourself, through tools such as Ollama. Capabilities are described as intended functionality. Items marked Planned are not available today. Supported models and runtimes are confirmed for each deployment.

What it is

The enterprise adapter for governed AI.

TrustHarness is designed to help corporations and trusted institutions deploy AI at scale—on premises or in a private cloud. It brings organizational ethics, policies, and subject-matter requirements into the runtime decision process, governing individual requests and recording how each was handled.

What “corpus” means

Your corpus is your organization’s approved documents, knowledge, policies, and guidelines. It is the material your AI is held to.

  • Documents
  • Knowledge
  • Policies
  • Guidelines
X marks held inside a blue bracket, with check marks emerging below it.
Conceptual representation of review and controlled release.

Four core use cases

From a single response to a running agent.

01

Review AI responses before customers see them

Check AI-generated responses against approved sources and organizational requirements before delivery. Release outputs that pass configured checks, request revisions, block violations, or escalate unresolved cases.

Illustrative applicationsCustomer support, client communications, and document summaries.

Draft reply to a customer

Illustrative

Thanks for reaching out. Withdrawals are free at any time. Your balance is shown in the client portal.

  • Approved sourceProduct sheet lists a withdrawal fee.
  • Required formatGreeting and portal reference present.
Held before deliveryBlock

02

Match outputs to incoming information

Compare AI-generated conclusions with inbound documents, internal information, and applicable guidelines. Identify unsupported claims, inconsistencies, and departures from encoded requirements.

Two distinct inputsA submitted document supplies facts. The approved corpus supplies the rules used to evaluate them.

Comparing a conclusion

Illustrative

Supplies the facts

Submitted document

Contract value $180,000. Insurance certificate expired June 30.

Supplies the rules

Approved corpus

Contracts above $100,000 require a current insurance certificate.

AI-generated conclusion

“Vendor meets onboarding requirements.”

  • Departs from an encoded requirement
  • Inconsistent with the submitted document

03

Govern batch commands and outputs

Apply a consistent review process across batches of AI commands, responses, or decisions. Check each item before execution or release, route exceptions, and record its disposition.

Item by itemEvery item gets its own status. One passing item does not approve any other item or the batch.

Batch B-204

Illustrative
Illustrative batch with one status per item
ItemTypeStatus
001Refund commandRelease
002Customer noticeRelease
003Refund commandEscalate
004Account updateBlock
005Customer noticeRevise
Exception queue 3 items
  • 003 Amount needs human approval
  • 004 Account is closed
  • 005 Wording only partly supported

04

Control agentic workflows

Build stable, monitored, and controlled agentic flows. Check proposed actions at defined control points, apply organizational requirements, and stop or escalate steps that fail configured checks.

Repeatable processes. Consistent review. Traceable actions.

Where repeatability comes fromControlled workflow steps, versioned rules, defined output formats, and consistent review criteria. Generative models can still word answers differently from run to run.

Agent workflow

Illustrative
  1. Retrieve account recordControl point passed. Step proceeds.
  2. Draft customer noticeControl point passed. Step proceeds.
  3. Send payment instructionControl point failed. Stopped and escalated.
  4. Update ledgerWaiting. Not authorized to run.

How it works

Five steps from corpus to record.

  1. 1

    Connect your corpus.

    Establish the approved documents, knowledge, policies, and guidelines your AI is held to.

  2. 2

    Define your controls.

    Specify what gets reviewed, which requirements apply, and when human approval is needed.

  3. 3

    Evaluate at runtime.

    Check candidate outputs and proposed commands against relevant information and configured requirements.

  4. 4

    Control the next step.

    Release, revise, block, or escalate before delivery or execution.

  5. 5

    Record the outcome.

    Sign a record of the rule versions, check results, and disposition, chained to the one before.

The AI model proposes

A model drafts a response or suggests an action. A proposal changes nothing on its own.

The system authorizes

TrustHarness determines what happens to the proposal, using your configured checks and approvals, before anything is delivered or executed.

Review coverage

Review selected workflows—or every designated output.

Choose where TrustHarness applies: customer-facing responses, internal decisions, batch processing, agent commands, or all designated material in a workflow.

  • Customer-facing responses
  • Internal decisions
  • Batch processing
  • Agent commands
  • All designated material

100% review coverage means every designated item passes through configured checks. It does not mean every error will be detected.

Review coverage
Which items go through configured checks.
Not the same as accuracy
Whether a check finds a given error is a separate question.

Interactive demonstration

Illustrative workflow

Select a scenario and follow it from request to record. All data is fictional. No live model is running, and nothing here shows measured performance.

Customer response

1Incoming request

“Is there a fee if I take money out of my savings account before day 90?”

Customer message. Fictional.

2Relevant corpus excerpt

Product Sheet ES-3, version 3.2, section 4

Withdrawals before day 90 carry a $25 fee. The fee amount must be stated in any reply about early withdrawal.

3Candidate response

“No fee applies. You can withdraw from your savings account at any time.”

AI draft. Held, not delivered.

4Check results

  • Source match“No fee” conflicts with section 4.
  • Required disclosureFee amount is missing.
  • Approved formatMeets the reply template.

5Disposition

Blocked because two rules failed. A new draft must state the $25 fee and is checked again before release. If no draft passes, a person decides.

6Audit record

Record
AR-1024
Rule version
ES-3 v3.2
Checks
2 failed, 1 passed
Disposition
Block
Integrity
Signed and chained

Guideline check

1Incoming information

Vendor onboarding form

Vendor: Fictional Supply Co.
Contract value: $180,000
Insurance certificate: expired June 30

Submitted document. Supplies the facts.

2Relevant corpus excerpt

Guideline PG-12, version 2.1, section 3

Contracts above $100,000 require a current insurance certificate. Exceptions need Risk Committee approval.

Approved corpus. Supplies the rule.

3Candidate conclusion

“Fictional Supply Co. meets onboarding requirements. Proceed to contract.”

AI conclusion. Held, not acted on.

4Check results

  • Rule applies$180,000 is above the threshold.
  • Current certificateDocument shows it has expired.
  • Exception approvalCannot be determined from the record.

5Disposition

Routed to a risk reviewer with the reason stated. The conclusion is not acted on until a person decides.

6Audit record

Record
AR-2048
Rule version
PG-12 v2.1
Checks
1 passed, 1 failed, 1 unresolved
Disposition
Escalate
Integrity
Signed and chained

Agent batch

1Incoming request

Month-end refund workflow

An agent prepares refunds for batch B-1042 and proposes five commands.

Workflow trigger. Fictional.

2Relevant corpus excerpt

Payments Policy PP-4, version 5.0

Section 2: refunds up to $5,000 may proceed. Section 3: refunds above $5,000 need human approval. Section 4: no refunds to closed accounts. Section 5: every refund needs a case reference.

3Proposed commands

  1. 1refund 1182, $420, case 771
  2. 2refund 2044, $1,250, case 772
  3. 3refund 3310, $7,800, case 773
  4. 4refund 4127, $95, case 774
  5. 5refund 5063, $310, no case

4Check results

  • 1Within section 2.
  • 2Within section 2.
  • 3Above limit, section 3.
  • 4Account closed, section 4.
  • 5Case reference missing, section 5.

5Disposition by item

  • 1, 2Release
  • 4Block
  • 3, 5Escalate

Commands 1 and 2 pass on their own merits. The next workflow step waits until the exception queue is resolved.

6Audit record

Record
AR-4096
Rule version
PP-4 v5.0
Items
5 reviewed individually
Exceptions
3 in queue
Integrity
Signed and chained

Deployment

Run within boundaries you set.

On premises

Run the model, TrustHarness, your corpus, and your records on infrastructure your organization manages.

Private cloud

Run the same setup in a cloud environment dedicated to your organization.

Why controlled infrastructure matters

TrustHarness is designed to run where your model runs, so the model, your corpus, and review records stay inside boundaries your organization sets. You decide where workloads run and who can change the rules.

Your organization stays in charge

In both options, your organization chooses the model, approves the rules, and keeps the records. Supported models are confirmed for each deployment.

Why open-weight models

Trust built from your checks, not the model.

Open-weight models are models whose weights you can run yourself, through tools such as Ollama. You choose the model, run it on premises or in the cloud, and decide when it changes. TrustHarness is designed to hold its work to your rules, so trust rests on the checks you configure, not on any one model.

  • You hold the model

    It runs where you put it and changes only when you update it.

  • Your rules, not a vendor’s

    A closed model follows its provider’s safety policies. TrustHarness is designed to apply yours: your corpus, your thresholds, your approvals.

  • Still checked at runtime

    Open weights alone do not make a model transparent, safe, or correct. That is why every designated output and command is still checked.

Economics

Greater control over AI—and its economics.

TrustHarness is priced per request checked, with a monthly maximum for each user, so spending has a ceiling you know in advance. On our estimate, running an open-weight model can cost about 85% less per token than paying closed model API prices, before hardware and hosting.

Discuss Your Deployment

What drives your total cost

  • TrustHarness usePriced per request checked, up to the monthly maximum for each user.
  • WorkloadThe volume and complexity of requests.
  • UtilizationHow fully dedicated capacity is used.
  • Model selectionThe models you run and what they require.
  • HostingThe cost of the environment you choose.
  • Verification overheadThe compute and review time checks add.

Enterprise value

What runtime governance adds.

  • Corpus-grounded review

    Evaluate outputs against approved organizational material.

  • Runtime control

    Check responses and proposed actions before release or execution.

  • Consistent workflows

    Apply defined rules and review criteria across repeated tasks.

  • Managed exceptions

    Route failures and uncertainty to the appropriate review step.

  • Auditable records

    Signed, chained records an auditor can recheck without trusting TrustHarness or the model.

FAQ

Straightforward answers.

Capabilities are described as intended functionality. Specifics are confirmed for each deployment.

What is TrustHarness?

TrustHarness is an enterprise AI adapter from Black Lake Digital Markets. It is designed to check requests, outputs, and agent commands against your approved corpus and configured rules at runtime, then control what happens next.

What is a corpus?

Your corpus is your organization’s approved documents, knowledge, policies, and guidelines. AI work is evaluated against that material. You decide what belongs in it and which version applies.

How does it review customer-facing AI output?

Before a response is delivered, TrustHarness is designed to compare it with approved sources and your organizational requirements. The response is then released, returned for revision, blocked, or escalated to a person, and the outcome is recorded.

What do Release, Revise, Block, and Escalate mean?

Release: every rule passes and no other check raises a concern. Revise: the rules pass, but another check falls short, so the model tries again. Block: a rule fails, so the output or command is stopped and redrafted. Escalate: a fact is missing or a result is unclear, so a person decides. If redrafting does not produce a passing result, a person reviews it.

Can it compare incoming information with internal guidelines?

Yes, that is an intended use. A submitted document supplies the facts, and your approved corpus supplies the rules used to evaluate them. TrustHarness is designed to identify unsupported claims, inconsistencies, and departures from encoded requirements in an AI conclusion.

How does it govern batches and agent commands?

Each item is reviewed individually, so one passing item does not approve any other item or the batch. Exceptions go to a queue and each disposition is recorded. For agents, proposed actions are checked at defined control points, and steps that fail configured checks are stopped or escalated. Agents can reach the gate through its MCP server before they act.

Can an auditor check the records?

Yes. Each decision carries a signed record of the facts, rule results, and outcome, and each record holds the fingerprint of the one before, so changing any record breaks the chain. An auditor can recheck a record without trusting TrustHarness or the model.

What does 100% review coverage mean?

It means every designated item passes through configured checks. It does not mean every error will be detected. Review coverage describes what is checked, not how accurate the checks are.

Does it guarantee identical or error-free outputs?

No. Generative models can produce different answers to the same request, and no review process finds every error. Repeatability comes from controlled workflow steps, versioned rules, defined output formats, and consistent review criteria. TrustHarness does not guarantee compliance or accuracy.

Can it run on premises or in the cloud?

It is designed for both: on premises, on infrastructure your organization manages, or in a private cloud environment dedicated to your organization. Requirements and supported models are confirmed for each deployment.

Which models and runtimes does it work with?

TrustHarness is designed for open-weight models, whose weights you can run yourself through tools such as Ollama. Its model connector is designed to plug into Ollama, with a llama.cpp connector planned. Calibration is specific to the model you run, so supported models are confirmed for each deployment.

Why open-weight models instead of a closed model API?

You run the model where you choose, it changes only when you update it, and it is held to your rules rather than a vendor’s. Open weights alone do not guarantee safety or correctness, which is why TrustHarness is designed to check each designated output at runtime.

How is TrustHarness priced?

Per request checked, with a monthly maximum for each user, so the cost per user has a set ceiling. Rates are quoted for your deployment, and model and hosting costs depend on where you run it.

What determines deployment cost?

Our estimate that an open-weight model can cost about 85% less per token than closed model API prices covers model tokens only, before hardware and hosting. Your total also depends on workload, utilization, model selection, hosting, and verification overhead, so it is reviewed for your deployment.

The company behind TrustHarness

About Black Lake

TrustHarness is built by Black Lake Digital Markets, which connects institutions, technology, and trust—transforming how mortgage assets are originated, traded, and transferred.

Black Lake’s mission is to make mortgage markets faster, fairer, and empirically intelligent.

“We believe execution speed should never come at the expense of transparency or control.”
Black Lake mission statement
Office
985 Lake St. East, Suite 100
Wayzata, MN 55391

Black Lake is the marketing name for a group of affiliated companies providing a range of services. All services are provided through separate agreements with each company. This material is for general informational purposes only and is not intended to provide specific advice or a specific recommendation. Black Lake does not provide tax, legal, or accounting advice. Black Lake is not a licensed insurance company and does not underwrite or sell insurance products. Black Lake Investment Solutions LLC wholly owns Black Lake Services LLC d/b/a Black Lake Digital Markets and Black Lake Investments LLC. Black Lake Digital Markets offers proprietary technology services for the pricing, trading, transfer and tokenization of mortgage assets. This website highlights services provided by Black Lake Digital Markets. Black Lake Investments LLC is an Investment Advisor registered under the Investment Advisers Act of 1940 that provides both investment advisory and non-advisory services. Registration does not imply a certain level of skill or training. If you are interested in learning more about Black Lake Investments, please contact info@blacklakeinvestments.com or visit blacklakeinvestments.com. Additional applicable regulatory information is available upon request. Case studies are provided for informational purposes only and do not constitute specific advice or a recommendation. Opinions, results, or data presented are not indicative of future performance. Actual results may vary. Read Black Lake’s disclosures.

Put your AI workflows under your rules.

From customer responses to batch commands and agentic flows, bring corpus-grounded review, runtime control, and auditable decisions into your AI deployment.

Form preview. This form is not connected yet. Nothing you enter is sent or stored.

Our Privacy Notice explains how form details are handled.