RECON: A benchmark for researching individuals

We're excited to announce RECON, our benchmark built for people research, on HUD. While most deep research benchmarks evaluate agents across a broad range of web research tasks, RECON is the first benchmark dedicated to the people research needed for due diligence, fraud investigations, and data enrichment. Given just a name and a few identifying details, agents must uncover hard-to-find facts about a person's life and work, following evidence across their online history while distinguishing them from others with the same name.

The dataset contains 514 verified facts about 140 people across occupations, countries, and levels of online visibility, with roughly a quarter sharing a name with a decoy. Questions range from a first employer or an old username to a high-school GPA or a company's original tagline, with answers drawn from profiles, local publications, organizational records, and archived websites.

To run RECON, HUD used our dataset and public benchmark harness. External models were run through the repository's evaluation scripts with access to web search and code execution in their native harnesses when supported. Each answer is graded as correct, wrong, or missing by gpt-4.1-mini, which agreed with human review on 98.5% of a 200-field sample. Hosting RECON on HUD allows us to inspect the research steps behind each result as well as compare scores.

People and fields

We compiled 140 individuals across industries, roles, and digital footprint levels, avoiding public figures whose information is too widely indexed to stress-test the benchmark. The dataset covers 514 verified facts that reflect what comes up most in real people research tasks. Every field was verified against a primary source with an unambiguous identity link to the person. Fields relying on inference were rejected.

The people span more than 60 occupational and geographic strata across more than a dozen countries, with digital footprints ranging from heavily indexed to near-invisible. Roughly a quarter are paired with a same-named decoy. The fields cover professional history, education, contact and identity records, and online accounts. A name or username match alone was not sufficient to establish a reference answer.

A sample task

Each entry pairs a person with a set of fields to find. person_info gives the agent its ground-truth anchor, and struct defines what it needs to return. This example shows one field from Zack Kline's task, formatted for readability:

{
  "person_info": "Zack Kline, founder of A.I.R. Lawn Care in Rockville, Maryland",
  "struct": {
    "company_original_tagline": "The company's original tagline"
  }
}

company_original_tagline requires following a historical lead. In a selected Sixtyfour run, local coverage identified the company's former domain. A 2012 archive of its homepage contained “Join us, and take the A.I.R. Dare ©.” The reference answer is shown separately below and is not supplied to the agent:

{
  "company_original_tagline": "Take the A.I.R. Dare ©"
}

Grading

Each field is graded by a GPT-4.1-mini judge as correct, incorrect, or missing, ignoring format differences. A 200-field audit found 98.5% agreement with human review. Weighted accuracy is computed as (correct − incorrect) / total evaluated fields; standard accuracy is simply correct / total evaluated fields. Precision is correct divided by nonblank answers, and recall is correct divided by all fields. Missing answers remain in the denominator.

Sixtyfour XHigh ranked first among the systems evaluated, with +70.6% weighted accuracy. High scored +54.3%, Scout +48.4%, Medium +44.7%, and Low +28.8%.

Weighted accuracy is the number of correct answers minus the number of wrong answers, divided by all requested fields. A correct answer earns one point, a wrong answer loses one, and a missing answer earns zero. Sixtyfour XHigh returned 434 correct answers, 71 wrong answers, and 9 missing answers across the 514 fields.

Precision measures the share of supplied answers that are correct. Recall measures the share of all requested facts answered correctly. The distinction matters here. GPT-6 Astra achieved the highest precision, 90.6%, while recovering 28.0% of the facts. Sixtyfour XHigh recovered 84.4% at 85.9% precision, High recovered 67.7% at 83.5%, and Medium recovered 56.0% at 83.2%.

Following an agent's research steps helps explain where its answers succeed or fall short. These selected cases compare Sixtyfour XHigh with GPT-6 Astra on reading inside a source, tracing a republished paper, and following a record series. A fourth shows an XHigh error.

Names, locations, and identifying links have been omitted from these examples.

Reading inside a source

A question asked for the 1985 book in which an architect's illustrations appeared. GPT-6 Astra issued 55 search queries and opened no page. Its final queries paired the book's title with her maiden name, yet it returned no answer. Sixtyfour XHigh retrieved an archived copy of her firm's publications page and searched the full text of the scanned book, where the illustration captions credit her under her maiden name.

Tracing a republished paper

A question asked for the title of a paper an executive presented at a 2008 industry congress and for the host city. Both systems identified the congress. GPT-6 Astra opened its online list of talks, which omits his session, and returned no answer. Sixtyfour XHigh found the paper in a trade magazine, published under his byline the following year with an endnote naming the congress.

Following a record series

A question asked for a violinist's job title at a church and the months he started and left. GPT-6 Astra opened the church newsletter announcing his hire and left the field blank. Sixtyfour XHigh read twelve weekly issues of the newsletter, from the announcement to the staff notice that ended the role, and returned the title and both dates.

A failure case

A question asked for a business owner's full German job title at her former engineering firm. Sixtyfour XHigh found both the short heading and the full title, but returned the shorter one. Sixtyfour High, Parallel Ultra 2x, and Claude Opus 5.5 made the same error; GPT-6 Astra returned the full title. XHigh had found the answer and erred in choosing which string to report.

Claude Opus 5.5 returned an API refusal for 7 of the 140 people and, for 9 more, replied without searching that the requested details were private. GPT-6 Astra left 19 people blank without searching. Refusals affect the observed results, but these runs alone do not establish their cause or the models' research ability on requests they accept.

RECON evaluates retrieval of specific facts about people. Broader tasks, such as writing a complete research report, require separate evaluation. Scores reflect each system's tools and configuration as well as its underlying model. Source availability and run-to-run variation can also affect results.

The benchmark harness is public. The dataset is available by request from Sixtyfour because it contains personal information.

Each score comes from a single run during the September–October 2026 evaluation window. Cost and latency are outside the scope of this benchmark.