Who holds decision authority?
Separate human approval, delegated execution and autonomous discretion without assuming that a model's presence implies independent authority.
When automated financial decisions propagate faster than institutions can identify, contest or reverse them, operational efficiency becomes a question of systemic control. This independent research programme investigates the boundary between machine execution and meaningful intervention, publishing experiments, methods and evidence as the field evolves.
Financial institutions are adopting increasingly automated workflows, but automation is not the same as autonomy. This project studies what can be observed, what remains undisclosed, and which safeguards may fail when decision speed exceeds intervention capacity. Its laboratory is one method within a broader research programme—not a claim of access to private bank systems.
Separate human approval, delegated execution and autonomous discretion without assuming that a model's presence implies independent authority.
Test the time and technical conditions required to detect, contest, interrupt or reverse machine-driven outcomes.
Examine correlated exposures, shared providers and infrastructure dependencies under explicit experimental assumptions.
Automated execution does not mean autonomous governance. Research must distinguish systems that recommend, systems that act within fixed rules and systems with delegated discretion. Where the decision chain is unknown, the project labels it unknown rather than inventing evidence.
Common infrastructure can transmit a single defect across many institutions. The experiment contrasts independent dependencies with a common-mode shock; the displayed probability is a mathematical scenario, not a measured frequency of banking incidents.
A meaningful control requires an accountable owner, an observable event, an available interruption mechanism and an auditable record. Faster execution may improve service while simultaneously narrowing the time available for human intervention.
The first experimental protocol defines the effective intervention margin as the time until irreversible effect minus detection and intervention latency. A positive margin indicates temporal feasibility under the selected assumptions, not operational safety. Intervention coverage divides a positive margin by the time horizon; negative margins yield zero coverage.
The correlated exposure illustration uses p + (1 − p) × [1 − (1 − q)n], where p is the assumed common-mode failure probability, q is an explicitly illustrative 7% independent dependency failure probability, and n is the number of dependencies. It is a teaching model; it does not estimate real-world institutional risk. No bank-specific evidence is asserted.
Primary research starting points: Bank for International Settlements, Financial Stability Board, Bank of England, European Banking Authority. These are reference institutions, not automatically ingested datasets. Continuous institution-level ingestion and editorial verification remain separate implementation stages.
Research limitations. The public version is an interactive reproducible scenario engine, not a live monitoring service. Continuous institutional ingestion, proprietary decision traces and empirical calibration are not represented as implemented. On-demand Crossref discovery is browser-based and does not create a persistent archive.
Studies, institutional investigations, unresolved questions and newly indexed literature are kept separate. Every external finding retains its source and publication status. The project does not republish copyrighted research.
Curated primary-source records. Original short descriptions by BondStats. Publication dates and research classifications are not evidence of independent replication.
Institutional publication updates, not a republished newswire. Entries link directly to the original research and preserve its publication status.
Project Agenti examines the feasibility of using AI agents to investigate payment timing, liquidity demand and potential gridlock.
Read original projectA new occasional paper examines how frontier AI changes vulnerability discovery, remediation and shared-provider exposure.
Read original paperThe FSB consultation sets out twelve proposed practices for responsible AI adoption; these are proposals, not binding rules.
Read consultationCrossref bibliographic metadata is requested directly from the public Crossref API in your browser; availability depends on the provider. Results are discovery candidates, not vetted studies or news articles. When Crossref is unavailable, the curated publications remain available above. Editorial selection is not independent verification.
Use “Refresh research index” to request newly indexed publications.
When decision execution accelerates, effective human oversight may decline unless detection and interruption mechanisms accelerate too.
Test: intervention-latency sensitivity · Status: model hypothesisDependence on common models, infrastructure and data providers may produce correlated failures even when institutions appear operationally independent.
Test: dependency and concentration scenarios · Status: openAutomated audit logs alone may be insufficient if no identifiable authority can contest, stop or reverse a consequential decision.
Test: traceability and control ownership · Status: openAttribution. We link to original institutional publications and DOI records; no third-party full texts, images or figures are copied into BondStats.
Evidence hierarchy. Curated institutional material, unreviewed Crossref discovery, and synthetic BondStats experiments have visibly different statuses.
Reproducibility. Experimental assumptions and outputs can be exported. The literature monitor retrieves public Crossref metadata directly in the browser; the curated evidence catalogue is machine-readable.
Limitations. Discovery does not imply peer review, empirical validation, endorsement, accuracy of metadata or comprehensive coverage of the literature.
Explore twelve independent research protocols with sixty calculated outputs. Every control changes a reproducible scenario; no synthetic result is presented as observed banking data. The programme also publishes an evidence register, methodological limitations and machine-readable definitions. Select a laboratory to work with one question at a time.
The sources below identify official research institutions and documentation channels, not live feeds. No numerical claims are extracted from these websites in this release. Source retrieval is deliberately separated from synthetic model output.
| Institution | Research relevance | Evidence status | Primary source |
|---|---|---|---|
| BIS | Payments, liquidity, central-bank innovation | Reference only | bis.org |
| Financial Stability Board | Systemic vulnerabilities, third-party concentration | Reference only | fsb.org |
| Bank of England | AI adoption, financial stability, operational resilience | Reference only | bankofengland.co.uk |
| European Banking Authority | Model governance, banking supervision | Reference only | eba.europa.eu |
| European Central Bank | Payments, oversight, financial stability | Reference only | ecb.europa.eu |
| International Monetary Fund | Macro-financial stability and digital finance | Reference only | imf.org |
Institutional statements, dated primary publications, transparent definitions and reproducible methods. A simulation result is not an observed frequency, forecast or institution-specific claim. Missing evidence remains unknown.
Each protocol exports its assumptions, five derived outputs, the model version and generation time. Sensitivity CSV files change one input at a time while holding other assumptions constant.
Vary the assumptions, inspect formula definitions in the machine-readable research manifest, test extreme inputs and compare alternative causal mechanisms. These are deliberately simplified teaching models and not validated production risk engines.
Correlated model failure, hidden provider concentration, weakened accountability, algorithmic discrimination, cyber exploitation, operational outages, energy demand and faster liquidity spirals. The research laboratories make these trade-offs explicit rather than presenting automation as universally beneficial.
Better anomaly detection, more consistent audit logs, quicker fraud screening, continuous control testing, faster settlement and more efficient resource allocation. Real-world benefit requires empirical validation and appropriate governance.
Continuous institution-level publication ingestion, empirical calibration, BondStats peer-reviewed publications, secure institutional log import, live incident alerts and third-party model validation. Browser-based Crossref discovery is available on demand, but is not a continuously maintained news archive. The research manifest publishes these limitations so machines and readers cannot mistake plans for deployed functionality.
Browse dated institutional source records, compare their research limitations and follow each publication to its original publisher. This catalogue complements the curated research library with source-specific observations and searchable evidence questions.
Each proposition is paired with a falsification condition, making it possible to challenge rather than merely repeat the research premise.
Public data: evidence · publications · model definitions · release integrity