The data doesn't feed the AI. It lives inside it.

Real and measured. The benchmark is 456 questions whose answers were machine-computed from a frozen snapshot and digit-compared: 454 exact. Selection latency, recall and payload are recorded run outputs, not projections. The incumbent baseline was measured live on the same data. The appendix carries the per-operation tables and the methodology, including what was measured and how.

They published the pieces, which is exactly the point of the research map above. DSI encodes a corpus into a Transformer, KBLaM injects knowledge bases, Memory Layers ship parametric memory, RETRO integrates trillion-token corpora. The closest published system, Parametric RAG (SIGIR 2025), injects LoRA per document, per query, for open-domain QA. No published work makes an entire multi-domain corpus native as the primary answer path, cold-started from computed ground truth, with a deterministic exactness arm. NCL is that synthesis, and it is benchmarked, not proposed.

No, and this is the easiest claim to verify. Cache-augmented generation and long-context caching preload documents into a frontier model's context window, on a third party's servers, inside a token limit. NCL has no context window in the loop: the substrate is its own 2.7M-parameter model, trained from scratch on the corpus, running on the data owner's hardware. The corpus lives in its weights, not in a rented cache. A language model appears only at the narration step, reading figures that were already computed, and that step is swappable, including for local models. The 6.4 ms selection and $0 answer path exist precisely because there is no API call inside the answer loop.

The source systems remain the system of record; NCL absorbs from them and replaces the answer path, not the storage. Nothing is cached on third-party servers, so data control never leaves the owner's contracts or perimeter. Changes, including record removals for right-to-erasure requests, land through re-absorption: the substrate is retrained from the corrected corpus, a laptop-scale cycle, not a frontier-model retrain. What is absent from the corpus is absent from the model.

Because the model never does math. The neural substrate does one job: serve the complete relevant cell surface for a question (0.9998 recall, 6.4 ms). A deterministic arm computes every figure from those cells. The language model only parses the question and narrates over precomputed, certified numbers. Exactness is a property of the architecture, not a promise about LLM behavior.

Their incentive structure points at general-purpose scale: papers that advance one mechanism at a time on public benchmarks. This is a systems synthesis on private, structured, multi-domain corporate data: absorption pipeline, computed ground truth, calibrated selection, deterministic compute and parity validation against a live incumbent, engineered end to end. The labs are publishing the ingredients. This is the dish, shipped and measured.

The pipeline is corpus-agnostic by construction: absorption runs from raw structured exports, ground truth is machine-generated and machine-verified from the data itself with zero human labeling, and training fits on a laptop. It runs entirely inside the data owner's perimeter. Nothing about the architecture is tied to the pilot corpus.

Price the alternative. A conventional stack pays permanently: retrieval infrastructure, per-query LLM spend, integration fragility, and answers in the tens of seconds that still arrive fragment-fed. NCL's deterministic answer path costs $0 in LLM tokens, answers in milliseconds from complete context, trains on commodity hardware, and keeps sovereign data at home. In any domain where answers must be exact, fast, auditable and private, that is not an optimization. It is a different cost structure.