What does a layered accountability framework for LLMs actually trace?

The search covered five databases from January 2022 to March 2026. Four thousand five hundred twelve records yielded 122 primary studies and a dozen regulatory documents.

Not a tool. A synthesis.

Five dimensions and four mechanism families

The review frames accountability as an actor-forum relation split into five dimensions. Mechanisms cluster into four families. Technical controls form one family. Human oversight forms a second. Organisational governance forms a third. Documentation and traceability forms the fourth. Each family carries a maturity assessment so adopters can see where their own practice sits.

The dimensions themselves are not listed in the abstract. The paper says they emerge from the sociotechnical account. What the abstract does name are the four cross-cutting concerns: traceability, role clarity, and continuous monitoring. Three threads that run through every layer.

Four layers from provenance to redress

Layer one covers provenance. Data lineage, model cards, supply-chain attestations — anything that lets an auditor ask where a weight or a prompt came from. Layer two addresses application logic. Prompt orchestration, retrieval pipelines, guardrails, and the runtime decisions that turn a raw model into a deployed service. Layer three is human oversight. The review finds this layer under-specified across the entire corpus. Who reviews, when, with what authority, and what happens when they disagree — the surveyed instruments leave those questions open. Layer four handles governance and redress. Incident reporting, audit trails, regulatory notifications, and the paths a harmed party can follow.

The layers are not sequential. They operate simultaneously. A single output can trigger obligations in all four at once.

Mapping onto law and standards

The architecture is mapped onto the EU AI Act, whose high-risk obligations have applied since 2 August 2026. It is also mapped onto the NIST AI RMF with its Generative AI Profile, ISO/IEC 42001, and sectoral guidance in healthcare, consumer finance, education, and the public sector. Cybersecurity controls align to the OWASP LLM Top 10 (2025).

Four persistent gaps appear in every mapped instrument. Human oversight remains under-specified. Shared accountability metrics do not exist. Disciplinary disconnection separates legal, technical, and operational communities. Empirical evaluation of any accountability mechanism is limited. Five structural tensions sit beneath those gaps. The paper does not name the five tensions individually in the abstract. It states that no surveyed instrument resolves them.

What LAAF is and is not

LAAF consolidates the classification device into an integrated accountability architecture. The authors call it a synthesis of the surveyed evidence rather than a validated artefact. That means it has not been stress-tested in a live deployment. No hospital, court, or bank has run a pilot against it. The framework exists on paper as a reading of what the literature and the regulations already require.

For a team building an LLM system today, the value is a checklist that connects a model card to an incident report through the same vocabulary. The gap is that the checklist does not yet tell you how to staff the oversight layer or what metric proves the oversight worked.