Skip to content

How can people retain practical control of AI-assisted decisions?

Eight projects follow an AI-assisted decision through three moments: the opportunity to intervene, the grounds preserved for review, and the institution’s response after harm. Select a project to trace its evidence to the place where the record stops.

node&norm’s active research is focused on revocation in delegated AI workflows: when permission is withdrawn, does further protected action stop? Further research and development on the other projects is paused. Their published methods, code and evidence remain available, with findings and limitations preserved.

Explore the evidence

Click a large dot to trace its evidence · Five kinds of evidence, each read on its own terms

Three questions

Each project sits under one question. Counts describe the program’s projects. A filled mark in one kind of evidence carries no weight in another.

Read the same evidence as a table

Positions support navigation, not measurement. Evidence kinds are independent, not stages of maturity. Held marks include authored or illustrative work as specified in each record; they do not establish field effectiveness.

Where each project’s evidence stops

Five kinds of evidence, each read on its own terms. A filled square names what the retained record holds. An open square names evidence still needed; it is not an invitation or authorization to begin a study.

Held in the retained record Open Outside this project’s design
Evidence held and open for each node&norm project, by kind of evidence
Project Method or specification written Applied to cases or records Agreement or software checks Independent replication Effect in a deployed setting
Trust, Autonomy, and EvidenceMethods phase complete Held: Reconstruction method and preprint Held: 3 historical cases, all Unresolved Open: Independent reliability and validity Open: Scorers outside the program Open: Transfer to contemporary AI systems
Human Influence TelemetryMaintenance Held: Documentary scoring rules Held: 1 preserved case packet Held: 2 scorers agreed on 7 of 7 items, earlier rules Open: Replication under the current rules Open: Use in a live review process
Control Evidence CorpusDevelopment Held: Record structure with preserved gaps Held: 3 authored synthetic demonstrations Open: Human validity and reliability Open: Use by readers outside the program Open: Empirical pilot pending
Revocation in delegated AI workflowsCurrent study Held: Test contract, version 0.2 Held: 18 configured cases: 9 schedules × 2 permission rules Held: Offline checker, 3 corrections, replay matched 18 of 18 Open: Justified external replication Open: Autonomous agents and institutional oversight
Governed Decision IntelligenceSpecification and reference implementation Held: Specification Held: Demonstration receipts Held: Structure, routing and receipt-integrity checks Open: Implementation built outside the program Open: Reviewer understanding and effective control
Regulated Gate Decision SupportReference implementation Held: Record format Held: 6 illustrative records Held: Validation tools Open: Use by a team outside the program Open: Regulatory acceptance and workflow effect
Catholic Doctrinal Fidelity IndexProvisional development Held: Provisional scoring rules and source-to-method record Open: Scored response set after reconciliation Open: Qualified theological validation Open: Scorers outside the program Open: Use within SAICRED evaluations
Poenitentia InstitutionumForthcoming chapter Held: Approved chapter, forthcoming Outside design: Conceptual and normative study Outside design: Outside this design Outside design: Outside this design Open: Measured improvement in institutional repair

18 cells held. 19 open. 3 outside a project’s design.

Independent replication remains open in all seven projects designed to admit it, and effect in a deployed setting remains open in all eight. Those two columns mark the program’s frontier; outside work there adds the most information. Counts come from the project records on this page, and a filled cell in one column carries no weight in another.

Practical human control

These projects examine what a person could do at a consequential moment, and what records allow a reviewer to establish.

Trust, Autonomy, and Evidence

Methods phase complete
Historical documentary study

What shows that a person could understand, intervene and affect a decision?

The method reconstructs control conditions from three historical cases. The approved reassessment leaves all three Unresolved, conditional on the retained judgments: the assessed record leaves open whether practical control was present or absent.

Limits. Independent reliability, validity and transfer to contemporary AI systems remain unestablished.

Evidence still neededIndependent reliability, validity and transfer to contemporary AI-assisted decisions remain unestablished and require separate evidence.

Human Influence Telemetry

Maintenance
One human scoring exercise

Can reviewers agree on the evidence of human influence?

The method organizes documentary evidence of human authority and its exercise. In one preserved case packet, two scorers agreed on all seven items under an earlier set of scoring rules. Both used the same rating category throughout.

Limits. Broader reliability and replication under the current scoring rules remain open.

Evidence still neededCurrent-protocol approval and independent reliability evidence remain open. Candidate-protocol scoring remains prohibited while the release gates are unresolved. The earlier 7/7 agreement exercise does not establish reliability under the current rules.

Control Evidence Corpus

Development
Synthetic demonstrations

What survives in the record when a control is supposed to act?

Three authored demonstrations separate an instruction to act, evidence that the action reached its target and evidence of mitigation. The record structure preserves the gaps that block a conclusion.

Limits. The empirical pilot is pending. Independent human validity and reliability remain unestablished.

Evidence still neededAn empirical pilot, independent coding, validity and reliability remain unestablished. The retained demonstrations are authored synthetic records, not empirical validation.

When permission ends, does action stop?

The current study, Revocation in Delegated AI Workflows, examines withdrawn authority. Queued work, handoffs and retries create specific points at which a control must reach the action.

Decision evidence

Decision records and evaluation rules make different judgments available for scrutiny. Each project has its own method and evidence.

Governed Decision Intelligence

Specification and reference implementation
Software checks

What would let someone reconstruct the grounds for an institutional action?

The specification connects a decision to its evidence, authority and obligations. A reference implementation checks record structure, selected routing rules and the integrity of demonstration receipts.

Limits. Factual truth, reviewer understanding and effective control require evidence beyond those checks.

Evidence still neededIndependent evidence that the record improves reconstruction of decision grounds, reviewer understanding or effective control remains absent; bounded implementation checks do not establish those effects.

Regulated Gate Decision Support

Reference implementation
Six illustrative records

What must happen before a conditional approval permits action?

A record format and its validation tools make permission, outstanding conditions and closure evidence explicit. Six illustrative records show how a team can document a decision to proceed, stop or wait.

Limits. Regulatory acceptance and effectiveness in deployed workflows remain unestablished.

Evidence still neededEmpirical evidence about condition closure before action, regulatory acceptance and workflow effects remains unestablished.

Catholic Doctrinal Fidelity Index

Provisional development
Domain-specific evaluation method

Which failures should disqualify an otherwise persuasive answer?

The framework develops evaluation rules for SAICRED, a project examining Catholic reliability and doctrinal fidelity in AI responses. Its source-to-method record traces selected AI safety research into scoring rules and review procedures.

Limits. The scoring rules, configuration and calculator still require reconciliation. Qualified human theological validation remains open.

Evidence still neededRule reconciliation and qualified theological validation on held-out responses remain open.

Responsibility and repair

Correcting a decision can leave its consequences unresolved. This strand examines what an institution owes to the people affected.

Poenitentia Institutionum

Forthcoming chapter
Conceptual and normative research

What does an institution owe after an AI-assisted decision causes harm?

The approved chapter examines acknowledgment, responsibility and changes to institutional practice. It is forthcoming in AI and the Catholic Perspective, published by En Route Books and Media.

Limits. The full manuscript remains private. Publication date and chapter number are unconfirmed. Measured improvements in institutional repair remain unestablished.

Evidence still neededAn observable measure of institutional repair and empirical evidence linking changed practice to the harm it addresses remain open.

The same questions, in eight kinds of institution

Defense commands, federal agencies, trading firms, hospitals, enterprises, small firms, frontier developers and mission-driven institutions each delegate authority to AI-assisted or automated work, keep records of it and answer for its consequences. The sector map traces where a named directive, statute or enforcement record already asks one of these eight questions.

Open the sector map →