Glaus Thorne Systems, Inc.
Beyond the Black Box.
AI systems can produce sophisticated, convincing answers through internal processes that are difficult to interpret or independently examine.
That opacity presents a problem when people are expected to rely on those answers. A conclusion may sound authoritative without establishing whether the evidence supports it, whether important information has been overlooked, or whether the reasoning behind it withstands scrutiny.
Complexity does not justify unaccountability.
The inability to inspect a model's internal decision-making processes does not mean its outputs should be accepted without examination. We can still investigate whether the evidence supports a conclusion, identify inconsistencies, challenge assumptions, and evaluate the reasoning presented.
Glaus Thorne Systems, Inc. develops software that systematically evaluates AI-generated outputs and works alongside human decision-makers to strengthen the evidentiary support, reliability, and defensibility of conclusions throughout the decision-making process.
Understand, create, and check AI-assisted work.
GT Analyst is a research, writing, and reasoning assistant that brings evidence review into the work itself. It helps you understand complex material, develop ideas, draft and revise documents, compare sources and options, and pressure-test assumptions and conclusions.
Its evidence-grounded fact-checking examines whether claims are supported by, contradicted by, or extend beyond the available sources. It helps preserve meaning and qualifications as material is revised, and makes conflicting evidence and unresolved questions visible.
Supported workflows preserve the sources used, findings, and review history for later inspection and export. You can examine the basis of an answer—not just read its conclusion—and keep the final judgment in human hands.
Request a GT Analyst demonstrationNot just an AI assistant. A governed one.
An AI assistant can help you produce an answer.
GT Analyst is built to help you question it.
Governed means the assistant works within defined rules for evaluating evidence, recording findings, and keeping decisions in human hands. Its answers aren't treated as proof simply because the AI produced them.
The assistant can help. It doesn't get to decide.
Ask questions, analyze documents, explore alternatives, and challenge conclusions.
GT Analyst helps you do the work. You decide what to accept, what needs another look, and what happens next.
Evaluation is part of the workflow.
Defined checks look for gaps in supporting evidence, reasoning problems, and conclusions that go beyond what the sources establish.
The resulting findings are available for review.
The assistant's confidence is not a substitute for those findings.
The review doesn't disappear with the conversation.
Findings, disclosures, and review history are retained as part of the governed review process.
Return to saved work, inspect what was flagged, and review the recorded findings, disclosures, and evaluation history.
When someone asks why a conclusion was accepted, there should be more to consult than the final answer.
An assistant that helps. A process that checks. A person who decides.
Make the conditions, evidence, and behavior inspectable.
On supported workflows, Glaus Thorne makes AI-assisted system behavior under defined conditions and the resulting evidence record reviewable—so organizations can assess what the evidence supports.
Set controlled conditions
Protocols can freeze a failure or evaluation condition before execution and specify an expected result or comparison oracle where applicable. Oracle independence is identified rather than assumed.
Keep the run reviewable
Supported workflows preserve execution conditions, evidence lineage, and results in a versioned record. This gives reviewers a basis for examining what happened.
Keep conclusions bounded
Attribute outcomes only when the change under study is isolated. Replay and comparison require compatible records; interpretation stays within the conditions exercised.
What a governed evaluation can establish
Glaus Thorne’s evaluation model separates the condition being tested, what ran, the evidence recorded, interpretation, and human disposition. Selected workflows preserve frozen cohorts and versioned run records; replay or comparison is limited to compatible histories. Expected outcomes or oracles, controlled-change isolation, and independent adjudication depend on the specific protocol. A passing test does not establish broader performance.
What the test runs show.
Architecture describes the evaluation model; implementation varies by workflow; API behavior was exercised in bounded tests; practitioner outcomes and independent domain validation are not established.
In bounded API validation, 477 test cases passed with no failures across distinct test scopes. Of these, 470 ran against disposable PostgreSQL; seven chain-monitor cases passed separately against live PostgreSQL in rollback transactions. This combined count is not a single 477-case run. Results describe behavior on exercised cases and do not establish independent accuracy, operational effectiveness, or customer outcomes.
Evidence stage: exercised API behavior. Field observations and independent validation are separate evidence stages.
Contact
Enable JavaScript to use the contact form, or email us directly to request a GT Analyst demonstration.
Contact