DAI: A Domain-Specific LLM with Model-Excluded Authority, Capability-Disjoint Verification, and Recomputable Receipts

Abstract

DAI (Decripte Artificial Intelligence) is a proprietary domain-specific LLM within a governed autonomous defensive architecture. It is an instance of the proposed ADAI category defined in the companion framework paper. The paper describes incident-response sessions spanning signal curation, structured target extraction, reviewed playbook selection, declared action contracts, deterministic admissibility, bounded network-edge execution, capability-disjoint verification, hash-linked evidence sealing, scheduled review, reconciliation and undo. The learned model is trained on the system’s own outcome-verified transcripts and narrates the result; admissibility and execution authority remain structurally outside the model. The paper addresses per-principal containment within shared enforcement domains, rule-budget aggregation, element-identified reconciliation, independent receipt recomputation and exact localization of controlled tampering. Reported measurements refer to the version and period described in the manuscript and are not a claim of universal performance.

Problem and approach

How can a defensive system act on production without making model confidence the authority for action? These papers place enforceable boundaries before execution and verifiable evidence after it. The LLM is part of the system, while the decision function is pure in its declared contract inputs.

Inside the manuscript

• Architecture and episode flow
• The action envelope
• Evidence and independent recomputation
• Principal-scoped containment
• Review, reconciliation and testimony
• Tokenizer, causal Transformer and training
• Observed failure classes
• Production measurements and limitations

Companion paper

Twelve mechanisms, precisely named · AC / MEA / ACO / EBT / PPB / DSC / SAV / SSC / PPR / ELR / RBT / SRL

Cite this paper

BibTeX · RIS · Zenodo