Independent AI research lab
Measurement science for trustworthy AI.
We help AI labs and organizations in high-trust domains understand and build trustworthy AI systems.
Featured research
When AI Is Your Pastor
Do clearer instructions improve how AI handles theological, moral, and pastoral-adjacent questions?
What we found
Guided responses scored higher in every question category.
- Average guided improvement
- +3.96
- Pastoral application, largest gain
- +6.62
Why it matters
System instructions and product design changed the behavior users would encounter, even when the underlying model stayed the same.
- 14 frontier models
- 8,792 scored items
- 120 scenarios
Read how we tested it ↗
What we measure
Trustworthy AI requires more than correct answers.
Fide AI studies whether complete AI systems use evidence well, respect the limits of their authority, and keep people in control. We turn those questions into benchmarks and evaluation methods.
01
Evidence
Does the system use authoritative sources and represent them faithfully?
02
Authority
Does it understand what it may claim, recommend, or decide?
03
Oversight
Does it defer, escalate, and preserve meaningful human control?
High-trust domains
Where trust carries consequences.
High-trust domains are settings where people make consequential judgments from context, take high-stakes action balancing risk vs reward, and shape relationships and deliver care.
These settings include enterprise, cybersecurity, finance, law, healthcare, education, and faith.
First domain
Faith
Faith concentrates the questions we care about: authoritative sources, legitimate authority, and human formation. We begin there because the methods can transfer wherever people depend on AI for consequential guidance.
How we measure
We evaluate the whole system, not just the model.
Real-world behavior emerges from the model and everything around it. We measure the assembled system under stated conditions so findings reflect what people actually encounter.
01
Model
The underlying capabilities, tendencies, and limits.
02
Instructions
System prompts, policies, roles, and behavioral constraints.
03
Sources and tools
Retrieval, databases, APIs, and the authority behind them.
04
Interface and workflow
How the system is presented and used in practice.
05
Runtime and records
Operational traces, incidents, decisions, and outcomes over time.
06
Human oversight
Review, escalation, accountability, and final authority.
Trustworthiness is a property of the system people use, not a score attached to a model in isolation.
Agent alignment and runtime assurance
Trustworthy AI must remain trustworthy while it works.
Agents do more than generate answers. They plan, use tools, delegate tasks, access information, and take actions across systems. We develop methods to test whether that behavior remains aligned with human intent and institutional policy during real-world use.
Read the research call →01
Observe
Capture the task, sources, tool calls, permissions, delegation paths, actions, and outcomes needed to understand what occurred.
02
Interpret
Distinguish expected behavior from error, drift, manipulation, unauthorized action, and emerging patterns of misalignment.
03
Intervene
Give responsible people the evidence and controls needed to correct, constrain, pause, or stop an agent before a failure compounds.
Published research
01/03Clearer system guidance improved scores across all 14 tested models.
Because the instruction changed performance across all 14 models, an evaluation should test the model together with its system prompt, sources, interface, and escalation rules.
Guided response score by question type
0-100 scale · improvement vs raw model condition
Pastoral application
The largest improvement came when models were given clearer guidance for pastoral-adjacent situations.
+6.62
92.30
Primary doctrine
Models performed better when the task called for doctrinal clarity instead of vague balance.
+3.51
88.00
Secondary doctrine
Guidance helped models represent disagreement without flattening real theological differences.
+2.64
91.40
Tertiary questions
Less central questions benefited from humility, uncertainty, and careful framing.
+1.62
91.70
Research transparency
Trustworthy AI requires transparent research.
Every research program reflects judgments about what matters, what counts as harm, and how evidence should be interpreted. Establishing trust requires making those presuppositions known.
Fide AI discloses the assumptions, prior commitments, and worldviews that shape each project. Readers can examine the lens, test the methods, challenge the interpretation, and decide what the evidence supports.
01
State the lens
Disclose the assumptions, prior commitments, and worldviews that materially shape the research.
02
Show the evidence
Publish methods, data, limitations, and uncertainty so others can inspect the work.
03
Separate evidence from judgment
Distinguish measured findings from interpretation, recommendation, and moral judgment.
That transparency strengthens the research by making disagreement precise and conclusions accountable. The independence policy explains the rules that protect the integrity of the work.
From research to action
We provide assurance, not insurance.
We investigate how complete AI systems behave, where they fail, and what needs to improve so organizations can make deployment decisions with confidence grounded in evidence.
From models before release to agents and deployed applications, our measurement science helps organizations understand and build trustworthy AI.
01
Research
Publish papers, methods, data, and open questions for others to inspect.
Browse published research →
02
Assurance
Provide independent evidence for release, deployment, and continued-use decisions.
Explore evaluation and assurance →
03
Public guidance
Explain what the findings change for institutions, builders, and the public.
Read analysis and guidance →
Bring us a system or consequential question you need to understand.
Fide AI works with AI labs and organizations in high-trust domains on scoped evaluations and public-interest research. We define the question, make assumptions visible, and produce evidence others can inspect.
Newsletter
Read new work by email.
Substack carries new essays and research notes. The permanent archive stays here on Fide AI.