Pagefifty

The model and where it runs

Model
nvidia/nemotron-3-super-120b-a12b
Provider
Nebius Token Factory
Region
us-central1
Data retention
Zero Data Retention on
Web search
Tavily

How the AI is used

When the snapshot is captured, one model reads the published selection criteria of the tender lots the capture run sends it. The criteria of a lot without a reading are shown as published and not compared with your firm. The site shows those readings as recorded and calls no model while you read.

Only the public criteria text of each lot is sent to the model. No firm profile, contact or personal data is sent to the model or the search provider.

The region is the one the provider's model list names for this model. For a public model endpoint, Nebius says the processing location is decided dynamically and is not tied to any single data center.

Source: Nebius Token Factory, Legal Quick Guide, 5.1

Zero Data Retention is switched on for the account that runs the read. With it, Nebius says, inputs and outputs are not stored on its systems after each request is processed, and are not used to train, fine-tune or improve any model.

Source: Nebius Token Factory, Legal Quick Guide, 4.1

Only when the buyer search runs, the buyer's name and fixed search words are sent to Tavily. Tavily's privacy policy says it is a US based entity, may use query data to improve its responses to future queries unless a contract says otherwise, and may share query data with third-party search index providers.

Source: Tavily Privacy Policy

Since 20 October 2026, the setting "Allow use of query data" is switched off for the Tavily account that runs the search. With it off, Tavily says, query data is not stored or used for improvements. One test search, made on 15 October 2026 before the setting was switched off, is not covered. The buyer searches this site shows ran on 22 October 2026, with the setting off.

Source: Tavily Help Center, the Allow use of query data setting

What code decides

The model points at the clause that states each requirement, quotes it, and writes a plain English reading of it. Each reading on a tender page carries an AI reading label.

Code checks every quote against the notice's own text and rejects any quote it cannot find there. Code reads every figure from the checked quote, compares it with the firm profile in your browser, and decides each criterion's state and each lot's verdict. The model decides no state and no verdict.

What a verdict does not cover

A Go covers only the selection criteria read in the notice, never the exclusion grounds, the procurement documents or the buyer's shortlisting.

Where a buyer names a category of references, staff or business lines, only the firm can judge whether its own match, so that criterion reads Ask first.

The recorded run

This run read the capture of 2026-10-22. The example run shows its readings on each tender page.

Model read of this capture

Model
nvidia/nemotron-3-super-120b-a12b
Provider
Nebius Token Factory
Region
us-central1
Run date
2026-10-22
Input tokens
293,177
Output tokens
208,761
Reasoning tokens
0
Cost in USD
0.275838
Quotes verified
1,789
Quotes rejected
76