Not yet human-reviewed
This source record was written automatically by the research pipeline and has not yet been reviewed by the maintainer or integrated into a topic page. Treat it as a provisional summary and verify claims against the original source.
LAAF: A Layered Accountability Architecture Framework for LLM Applications
Citation. Chaturvedi, P., Ahmad, S., Nowroozi, E., Waqas, M., Loukas, G., Jolfaei, A., Cordeiro, L., & Dantas, P. (2026). LAAF: A Layered Accountability Architecture Framework for LLM Applications. arXiv:2608.27102.
Summary. This preprint reports a PRISMA-guided systematic review of LLM accountability literature and regulatory instruments, then proposes LAAF as a four-layer architecture spanning model provenance, application controls, human oversight, and governance/redress. It treats accountability as an answerability relation requiring traceability, named roles, and continuous monitoring across the LLM application lifecycle. The framework is a synthesis of reviewed evidence, not a validated deployment artefact.
Key claims
- The review reports screening 4,512 records to include 122 primary studies plus 12 regulatory and standards documents, and finds that accountability mechanisms must combine technical controls, human oversight, organizational governance, and documentation/traceability rather than rely on one family alone.
- The authors argue that mechanism maturity is strongest near the model (for example, grounding, retrieval, and logging) but that regulatory obligations are most specific for oversight and post-deployment governance, where empirical evidence remains comparatively thin.
- LAAF organizes accountable LLM applications into four layers—foundation-model provenance, application logic and guardrails, human oversight and review, and governance/audit/redress—cross-cut by traceability, role clarity, and continuous monitoring.
- The review identifies under-specified human oversight, absent shared accountability metrics, disconnection between technical and governance literatures, and limited empirical evaluation as persistent gaps; it frames LAAF as a design response rather than evidence of effectiveness.
Evidence & limitations
- The paper uses a structured review protocol, five academic databases, citation chasing, purposive regulatory-document collection, thematic synthesis, and author-assigned quality scoring; its search cutoff was March 31, 2026.
- It is an arXiv preprint, not a peer-reviewed publication, and the authors acknowledge English-only coverage, an unregistered protocol, scope limitations, author-scored framework comparisons, and limited quantitative comparability across accountability mechanisms.
- LAAF assumes decision-support outputs reviewed before action and explicitly does not cover agentic/tool-calling deployments without adaptation; the proposed architecture has not been validated in deployed systems.