ELLMENT: decision support that cites its sources, at altitude, with no network.
The E-2D Large Language Model Entity is an Azcuna Dynamics program built for the crews of the E-2D Advanced Hawkeye and proposed to NAVAIR PMA-231 under Navy SBIR Phase I. Naval Flight Officers aboard the E-2D Advanced Hawkeye apply NATOPS procedures, tactics, rules of engagement, and a running record of mission communications in seconds, under pressure, from memory or paper. ELLMENT is an on-demand assistant that reasons over that material in real time, cites the governing passage for every answer, and runs fully disconnected.
Fifty-five seconds, unedited. The prototype answers a carrier NATOPS question with a page-level citation, then refuses a fabricated brevity code instead of guessing. Running entirely on a laptop.
Constraint 1 · Fully disconnected
A carrier-based tactical aircraft is the archetypal denied and degraded environment. Cloud-hosted and API-dependent models are non-starters. ELLMENT runs every component locally, with zero network egress, so it can be certified as stand-alone under information-assurance review.
Constraint 2 · Traceable and explainable
An operator cannot act on an unattributed assertion from a black box. Every answer points to the governing source, exposes the evidence it used, reports a calibrated confidence, and abstains when the corpus does not support an answer.
Azcuna did not propose to build an offline, citation-grounded LLM from scratch. The hard, mission-agnostic core already exists in BABEL: a 100% offline document-intelligence platform that performs retrieval-augmented generation with per-answer source attribution, runs quantized open-weight models on a 16 GB laptop, and is scored by a reproducible benchmark harness. ELLMENT is that core retargeted to the E-2D decision-support domain. The topic's highest-risk requirements, offline operation and explainability, become the lowest-risk parts of the effort. Phase I work concentrates where the real work remains: domain corpus curation, mission-aligned evaluation, and a defensible deployment path.
A modular, self-contained pipeline with no runtime network dependency.
Ingestion & Indexing
Procedures, TTPs, checklists, and communication logs are parsed, chunked to the model context budget, and embedded locally. Every chunk keeps a provenance record down to document, section, and page.
Hybrid Retriever
Dense semantic search fused with lexical search, then re-ranked. Tactical text is full of exact identifiers - system names, procedure numbers, call signs - that pure semantic search misses.
Grounded Generation
A quantized open-weight model served locally generates an answer constrained to the retrieved evidence. The prompt enforces inline citation and forbids claims the passages do not support.
Trust & Transparency
Inline citations, an expandable evidence view of the exact retrieved text, a calibrated confidence score, and an abstention path that returns "insufficient grounding" instead of a guess.
Governance & Audit
Model and corpus versions, configuration, and per-query audit logs are recorded so results are reproducible and the system can be managed across its lifecycle.
Ahead of award and at private expense, Azcuna stood up a working ELLMENT prototype on publicly available naval and aviation doctrine and ran it against a purpose-built adversarial question set, including out-of-corpus questions and fabricated brevity codes designed to bait a confident wrong answer. Figures below are on the held-out test split. All tuning was done on a disjoint development split.
The hardest safety property, refusing to answer when the corpus does not support an answer, already meets the Phase I feasibility threshold. Retrieval recall and abstention calibration are the identified tuning objectives the Base effort is scoped to close.
A six-month Base that delivers an evaluated, fully offline, source-attributed prototype against a quantitative go/no-go metric, followed by a six-month Option that bridges to Phase II.
- Citation accuracy: the inline citation resolves to the governing document and page at or above 90%
- Abstention correctness: explicit refusal on at least 95% of out-of-corpus and fabricated-identifier probes
- Retrieval recall@k: gold page present in the retrieved set, the leading indicator of end-to-end accuracy
- Operational appropriateness: SME-rubric scores for usability under mission tempo, correctness, and relevance
- Confidence calibration: expected calibration error below a pre-registered bound
- Every metric produced by a scripted, re-runnable harness with zero network egress
Fine-tuning bakes knowledge into opaque weights and degrades traceability. RAG keeps authority in inspectable source documents, makes citation native, and lets the corpus be updated without retraining. Fine-tuning is held as a Phase II option for style and format alignment only, never as the system of record for facts.
A summarized answer with inline citations to the governing passages, an expandable view of the exact retrieved text, and a confidence indicator. When the corpus does not support an answer, the system says so rather than guessing. The human keeps decision authority with the evidence in view.
At the architecture level, not with disclaimers. Answers are constrained to retrieved evidence and must cite it. Unsupported queries are abstained on. Confidence is calibrated and exposed. Bias is tracked as a first-class evaluation metric.
Open-weight models Azcuna already runs offline in 4-bit quantization, in the Llama-3.1-8B, Phi-4-14B, and Gemma-3-12B class, chosen for permissive licensing, on-device feasibility within the SWaP envelope, and instruction-following. No Chinese-origin model weights are used. Final selection is decided by evaluation, not asserted up front.
Unclassified surrogate data only: publicly available naval and aviation doctrine, training-style TTP material, structured checklists, and synthetic mission-communication transcripts. The architecture is structured so later classified or export-controlled work is segregable from the Phase I research.
Phase II moves ELLMENT into a stand-alone laboratory environment for operator-in-the-loop evaluation with Naval Flight Officers, adds lifecycle monitoring, and transitions work that may become classified. Phase III fields the capability through Navy acquisition. The same offline, source-attributed pattern applies to other denied and degraded platforms, and the topic itself notes follow-on interest from ONR Code 32 for anti-submarine warfare.
Prototype figures are internally funded results on unclassified, publicly available surrogate data. Phase I scope reflects the proposal as submitted to the Department of the Navy.