Guided path, Step 4 of 4

See the cost

How Nemotron did the work · Recurance

How Nemotron did the work

Three sizes of NVIDIA Nemotron on Nebius Token Factory read 6,454 coroners' reports, judged 4,543 possible recurrences and wrote 25 dossiers, for $5.65 in all.

6,454

reports read, 2013 to now

15,008

paid model calls, every one in the cost ledger

What surprised us

98%of the quotes Nano extracted were found word for word in the source. The rest are never shown.See the funnel
66%of the pairs that looked alike were judged not the same hazard: related or different. That is why a judge is needed.See Super's verdicts
9 of 14test pairs it missed never reached the judge. The next gain is in retrieval, not in the model.See where it fails
$0.0085per verified chain, every stage counted, across 666 chains.See the cost

From archive to evidence

Each step narrows the archive. Small models do the reading, a larger one makes the judgement, and plain code checks every quote against the source.

Step 1Concerns extracted15,366
From 6,454 reports, with OCR for scanned ones. Each concern is an exact quote.
Nemotron Nano
$0.77 · 5.3M tokens in, 1.9M out
Step 2Quotes verified word for word15,044
98% of quotes were found word for word in the source.
Exact-match check
No model calls: checked in code
Step 3Pairs found and judged4,543
Same organisation, at least six months apart, close in meaning or wording. Super decides whether it is the same hazard. The cost includes the evaluation runs.
Qwen3 + pgvectorNemotron Super
$3.38 · 6.6M tokens in, 3.3M out
Step 4Links shown1,360
Identical or closely related, confident, both quotes verified.
Rules
No model calls: checked in code
Step 5Chains666
One organisation, one hazard, over time.
Grouped
No model calls: checked in code

Bar widths use a square-root scale, so every step stays visible. Costs and tokens come from the call ledger.

Alongside

2,276

promised actions taken from 1,763 response letters

Nemotron Nano

Alongside

269

promises followed up on the web, using 801 Tavily credits

Nemotron Nano

Alongside

dossiers with a timeline and open questions

Nemotron Ultra

Alongside

258

repeat requests answered free from the request-hash cache

Cache

What Super decided

Verdicts on 4,543 candidate pairs. 66% were judged only related or different: alike in wording, but not the same hazard. That is why a judge is needed at all. Only confident identical or closely related verdicts with both quotes verified are shown (1,360 links).

Right-sized models

Each task goes to the smallest model that does it well. The big model is kept for the few jobs that need it.

Nemotron 3 Nano

30B parameters, 3B active

nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B

Read every report and response letter. Pulled out each concern (15,366) and each promised action (2,276) as an exact quote, in enforced JSON. Sorted web search results for 269 follow-ups.

Paid calls: 9,593
Cache hits: 47
Tokens in: 8.8m
Tokens out: 2.3m
Reasoning tokens: none
Average latency: 5.9 s
Cost: $1.07
Share of spend: 19%

Concern extraction, per 1,000 reports$0.12

Why this model

Extraction is a narrow task repeated thousands of times. Nano is the smallest model that copies quotes reliably, so reasoning stays off.

Nemotron 3 Super

120B parameters, 12B active

nvidia/nemotron-3-super-120b-a12b

Judged 4,543 candidate pairs: identical, closely related, related or different, with a confidence, a short reason and a quote from each report.

Token Factory features used

Every call goes through one small wrapper, in the pipeline and in this app. These are the features it relies on.

Structured JSON output
A json_schema response format, enforced on Nano, Super and Ultra. Every concern, promise and verdict arrives as valid JSON, with no free-text parsing.
Reasoning toggle
chat_template_kwargs turns reasoning off for extraction and sets it to low effort for the judge (1.3m reasoning tokens from Super in all).
Live prices
The models endpoint with verbose pricing gives the price per token. Each call is costed from it, and so is the budget guard.

At request time: the live agent

The same models run on demand. Give the brief a new report and it works through these steps while you watch, showing what each step found, its tokens and its cost.

1. Read the report and find the organisation
Archive or judiciary.uk
2. Extract the concerns
Nemotron Nano

Does the judge earn its cost? On a blind test of recurrences that coroners themselves flagged, the hybrid with the Super judge found 52%, against 34% for TF-IDF alone, 41% for embeddings alone and 52% for Hybrid + report route + Nemotron Super judge alone, at the same number of shown links.

Evaluation Methodology

How Nemotron did the work

What surprised us

From archive to evidence

What Super decided

Right-sized models

Nemotron 3 Nano

Nemotron 3 Super

Token Factory features used

At request time: the live agent

Nemotron 3 Ultra

Qwen3 Embedding 8B

Where the $5.65 went