Search papers, labs, and topics across Lattice.
This paper establishes an auditable framework for AI-assisted research writing that integrates accountability directly into the production process, utilizing techniques like git sealing and hash-bound provenance. By implementing red-line gates and cross-model role separation, the authors ensure that non-compliant artifacts are systematically rejected and logged, thereby enhancing the integrity of research outputs. A key finding demonstrates that this approach can halt projects based on pre-registered stopping rules, effectively enforcing compliance even against the operators' intentions.
An auditable framework can halt research projects in real-time based on compliance failures, ensuring integrity in AI-assisted writing.
Language models now draft, classify and criticise inside research production, yet the artifacts they help produce carry little accountable history. Rather than detecting machine involvement afterwards, we specify an auditability discipline built at production time: git sealing with an anchor lineage, hash-bound provenance, red-line gates that refuse non-compliant artifacts and log every refusal, cross-model role separation, and programmatic assembly from registered sources. Adherence is instrumented by metric cards, each carrying a pre-registered blind spot and evidential standing, frozen before the prospective case it observes. In that case the observed project's pre-registered confirmatory test was executed under seal and returned No-Go, and that project's frozen stopping rule halted the work, against its own operators. A lower-graded retrospective case covers families whose machinery predates the protocol. Current observations are provisional; we release a package from which a third party can recompute every primary metric.
