The Living Archive (Nemotron edition)

Inspiration

Small organizations - clinics, legal aid, local government, unions - own the documents and the responsibility, but not a research team. Generic assistants answer fluently from memory; in these settings that is the failure mode, not the feature.

What it does

The Living Archive answers only from the corpus you give it, attaches the source page to every claim, and refuses with a reason when the corpus is silent. It is local-first: sensitive material never has to leave the building. It ships a deterministic citation verifier (cit_mbeshtet - citation-supported) plus a corpus-independent harness: CLI or HTTP adapter, a JSONL question set, and a CI gate (exit 0 = pass, 2 = fail).

The Nemotron lane - a substantial update

Same guard, new engine. The verifier is unchanged and model-independent: it checks evidence, not prose - which is exactly what makes swapping a local model for a frontier cloud model safe. Generation is served by NVIDIA Nemotron via Nebius Token Factory. Sensitive material stays local; the cloud lane is a choice, and it is used only with public or synthetic corpora.

How we measure

Identical JSONL question set, identical thresholds, per-response citation support, refusal correctness, hallucination rate, p50 latency and measured tokens per query (the cost basis) - one harness, four engines, printed side by side. No cherry-picked prompts. Numbers are published, not promised: we report accuracy, supported-citation rate, correct refusals and hallucination rate, and we print the cases where we are wrong. Measured matrix, n=50, 28.09.2026 - one harness, four engines (rows are comparable only inside the same set):

Configuration (engine + set + date)	n	Accuracy	Supported citations	Verifier pass	Correct refusals	Hallucination	p50 (s)	Tokens in/out (measured)
Nemotron-3-Ultra-550b-a55b via Nebius Token Factory - set-50 - 28.09	50	35.6%	33.3%	57.8%	100%	0%	2.59	96,992 / 23,428
Nemotron-3-super-120b-a12b via Nebius Token Factory - set-50 - 28.09	50	31.1%	31.1%	53.3%	100%	0%	3.97	94,032 / 29,649
Local qwen3.8-liruar (27B) + nomic-embed-text - set-50 - 28.09	50	28.9%	24.4%	44.4%	100%	0%	1.65	- (local)
Nemotron-3-Nano-30B-A3B via Nebius Token Factory - set-50 - 28.09	50	20.0%	6.7%	8.9%	100%	0%	9.28	120,062 / 49,412

What holds on every row: the deterministic citation guard refused 100% of the unanswerable questions with 0% hallucination - cloud models and local alike. The cloud engines lead on accuracy (35.6% / 31.1% vs 28.9%), but at n=50 that gap sits inside one standard error, so we do not claim a winner - we publish the table, the measured token counts per query (the basis for cost per query) and the raw JSONL.

Jury-proofing footnote: every row carries its engine, question set and date, and rows are compared only inside the same set. The set-50 index was built with nomic-embed-text; an incompatible embedder (bge-m3) measures 0.72 vs 0.00 similarity on the same fragments and is therefore not comparable - documented, not hidden. Cost per query is a direct multiple of the token column (list price per 1M tokens).