About the practice
When AI acts,
who answers?
Node & Norm studies the evidence and human authority behind consequential AI decisions.
The decision pathEvery decision is a Node.
Every standard is a Norm.
Responsibility needs a way to act.
Consider a team using AI to assess a claim, recommend a payment, or trigger an operational action. A reviewer needs the evidence, authority, and time to intervene. The system must also carry that intervention through to execution. A missing link can leave an institution accountable for a decision it cannot effectively control.
Node & Norm develops methods for examining those links. Its focus is the institution’s decision: the evidence used to authorize an action, the person able to intervene, the result that reached execution, and the responsibility that follows.
Without that record, a team may struggle to explain an action to a customer, reproduce a result, or establish whether a correction took effect. A missing correction can allow the same failure to recur.
The gaps are already visible.
A gap in control can expose other people’s systems. A failure of safety or disclosure can put an institution’s permission to operate at risk.
Anthropic · 9 September 2026
4reported incidents of
unauthorized access
During cybersecurity evaluations, models accessed real third-party systems without authorization. Misconfigured environments had allowed internet access; an earlier search had missed one incident.
The tests ran without the cyber safeguards supplied with released models. Frequency in ordinary use remains unknown.
Read Anthropic’s assessment →California DMV · October 2023
Deployment
suspended.
California’s DMV suspended Cruise’s deployment and driverless-testing permits, citing unsafe performance and misrepresentation of safety information among its grounds.
The consequence reached beyond the vehicle: the institution lost permission to continue the operations covered by those permits.
Read the DMV’s statement →These records give the research concrete questions: what could the system actually do, what did oversight detect, and what evidence supported the response? The Incident & Assurance Cases collection preserves source dates, reported findings, and unresolved questions as disclosures change.
Case records are provisional documentary reviews. Node & Norm has not independently investigated these incidents. The sources above were checked on 13 September 2026.
One decision, examined end to end.
Node & Norm asks what would justify an institution relying on an AI system for a particular decision, under the conditions in which it will actually be used.
Three permanent programs examine the same decision from different directions.
Human Control & Decision Authority
Can a person change
what happens?
Trust, Autonomy & Evidence examines the conditions that give authority practical force. Human Influence Telemetry examines how reviewers assess documented human influence.
Decision Evidence & Governed Action
Can the institution explain
why it acted?
Governed Decision Intelligence specifies a record of evidence, alternatives, authority, and obligations. Regulated Gate Decision Support explores this through a regulated-biopharma reference implementation.
AI Evaluation & Assurance
Does the evaluation
justify reliance?
This program examines judges, methods, scoring rules, and test conditions. Catholic Doctrinal Fidelity Index connects evaluation rules to their research basis and makes the scoring contract available for inspection.
Evidence for the result. Authority to act. A record of the decision.
Together, these inquiries help leaders examine how a test result becomes institutional action.
The work continues after harm.
Two emerging areas extend these questions. AI Incident Reconstruction & Assurance examines what happened and which controls held. Institutional Responsibility & Repair asks what an institution owes the people affected, and what must change in the decision process.
The researcher
“Trust in AI rests on evidence people can examine, decisions they can challenge, and corrections that reach those affected.”
Mark Julius Banasihan
Founder & Independent Researcher
An open research record.
Readers can examine the work through its Canonical Registry, including the assurance-case records. Research notes expose methods, source materials, findings, and limits. A technical test, a historical case study, and an observed improvement in a live institution each support different claims.
The research-integrity commitments call for disclosure of funding, evidence access, review restrictions, and responsibility for conclusions. External work retains its own identity; SAICRED is an external relationship.
Start with one consequential decision.
Choose an AI workflow in your institution. Trace the evidence behind an action, identify who can stop it, and check whether a correction reaches execution.
Use the research to structure that inquiry. Its methods and reference implementations require evaluation in your setting; demonstrated institutional benefits remain an open research question.
Explore practical human control