nebius-build-week/antex-runs

Antex colony runs

Real, recorded agent traces from Antex, an ant-colony research agent that stress-tests a product for a persona ("before you buy: find the reason it might be wrong for you, and show the evidence"). Built at Nebius Build Week × NVIDIA, Berlin, 18 Oct 2026.

Each run: a Queen (NVIDIA Nemotron on Nebius Token Factory) turns the persona into weighted criteria and dealbreakers, Explorers search the web with Tavily, Gatekeepers make one-token logprob decisions about each page, Foragers extract pages with Tavily and pull evidence crumbs with Nemotron, and Builders verify quotes and assemble the verdict.

There are three presets (one run each):

preset	product	persona
anker	Anker PowerCore 10000	a 12-year-old on a school trip (possibly flying)
magsafe	Apple MagSafe charger (USB-C)	a Samsung Galaxy S25 owner
stanley	Stanley Switchback travel mug	a parent buying one secondhand for a child

Provenance

All rows come from real live runs. Every URL, title, snippet and quote was retrieved from real public web pages by Tavily search / Tavily extract during the run; nothing is hand-written or synthetic. The rows are a direct export of the recorded event logs in fixtures/*.jsonl, produced by scripts/export_dataset.py. Web pages change, so a quote that was verified at run time may no longer be on the live page.

How quotes are verified

A forager crumb only becomes evidence if its quote is found verbatim in the page text that Tavily extract returned for that URL during the run (backend/antex/verify.py). Both strings are normalised first: Unicode NFKC, curly quotes/dashes/ellipsis mapped to ASCII, markdown characters (asterisks, underscores, backticks, hashes, >, pipes, square brackets) removed, whitespace collapsed, lower-cased; leading/trailing quote marks and periods are stripped from the quote, which must be at least 12 characters. Crumbs that fail this check are kept in events (quote_verified: false) but are excluded from evidence. A model writing a quote does not make it evidence; finding it in the fetched page does.

Configs

evidence (default): one row per verified evidence crumb

column	type	description
preset	string	run preset (anker, magsafe, stanley)
product	string	product as resolved by the Queen (brand + model + variant)
persona	string	persona as summarised by the Queen
criterion	string	criterion id from the Queen's plan (e.g. safety_battery)
stance	string	supports, contradicts or neutral (relative to the criterion's requirement)
quote	string	verbatim quote from the page (verified, see above)
url	string	source page URL
domain	string	source domain
region	string	source region/country code (INTL, DE, US, UK, ...)
criterion_name	string	human-readable criterion name
claim	string	the forager's one-line claim the quote supports
page_type	string	gatekeeper's page-type label (e.g. independent test/review)
title	string	page title
evidence_id	string	evidence id within the run (e1, e2, ...), as cited by the verdict
run_id	string	run id

events: the full agent trace, one row per event

column	type	description
preset	string	run preset
run_id	string	run id
seq	int	event order within the run
t	float	Unix timestamp (seconds)
type	string	event type: queen.plan, queen.decision, queen.replan, queen.stop, explorer.out, explorer.found, gate.decision, forager.out, forager.eat, forager.crumb, forager.empty, build.brick, build.crack, verdict
agent	string	emitting agent (queen, explorer-N, gatekeeper-N, forager-N, builder)
model	string/null	model id used for the step (null for non-LLM steps)
data	string	JSON-encoded, type-specific payload (decode with json.loads)

data is stored as a JSON string because its shape differs per event type. The per-type payloads are documented in the repo's backend contract. The verdict event holds the final verdict, decision record (rows), what_could_flip, unknowns and confidence.

Usage

from datasets import load_dataset
import json

evidence = load_dataset("nebius-build-week/antex-runs", "evidence", split="train")
events = load_dataset("nebius-build-week/antex-runs", "events", split="train")
verdicts = [json.loads(e["data"]) for e in events if e["type"] == "verdict"]

Limitations

• Small: three runs, one per preset. Model judgements (stance, gate decisions, verdicts) can be wrong; only the presence of the quote on the page is mechanically verified.
• Verdicts are evidence summaries for a persona, not professional, safety or purchasing advice.
• Quotes are short excerpts of third-party public pages; copyright in the quoted text stays with the original publishers. The CC BY 4.0 license covers the dataset's structure, annotations and agent outputs.