Halluminate's homepage sends visitors to "Our latest benchmark", the 88-problem Westworld Finance Diligence Bench. That post has charts and a trajectory explorer, but no leaderboard page, no dataset or sample tasks, no repository and no way to get in touch. Three months after publication its last line is still "We will release an accompanying paper soon." DealTrace, the private-equity benchmark, has none of these either. Yet the 2025 browser benchmark, Web Bench, had all of them: a leaderboard at webbench.ai, a Hugging Face dataset, a GitHub repo (still the company's most-starred, with 100 stars) and a contact line. That leaderboard's entries all date from May 2025. Rivals put finance leaderboards on their homepages: Mercor shows current top scores for APEX-Accounting and τ³-Banking, and Surge AI lists DAYJOB: Finance. Halluminate's research post exists to "Lead our public-facing benchmark and leaderboard efforts for frontier models", but the core finance results are still a one-time blog post.
First action
Put the Diligence Bench results table on its own page today, with a run date and a contact link.
How to measure it
Leaderboard citations plus inbound evaluation requests
- Now
- –
- Target
- 10 citations and 5 qualified requests
- By
- 90 days
Steps
Decide what is public
Which scores, sample tasks and trajectories can be shown without exposing customer work or deal data.
CEO 2 days
Build the hub
/leaderboards with one page per finance benchmark: table by model and harness, run dates, methods, citation block.
Member of Technical Staff, Platform Engineering 1 week
Release a sample
Ten to twenty anonymized tasks per benchmark on GitHub and Hugging Face, with a README and license.
Member of Technical Staff, Research / Post-Training 1 week
Add a contact path
"Evaluate your model or agent" form routed to the forward-deployed engineers.
Member of Technical Staff, Forward Deployed Engineer 1 day
Link it everywhere
Homepage, research page, every benchmark post and the GitHub org profile.
Member of Technical Staff, Platform Engineering 1 day
Set a rerun rule
Rerun within two weeks of any major model release and date-stamp each change.
Member of Technical Staff, Research / Post-Training ongoing
Decide on webbench.ai
Refresh it or mark it archived and point to the hub.
Member of Technical Staff, Research / Post-Training 1 day
Track citations
Monthly count of papers, model cards and posts citing the benchmarks.
Member of Technical Staff, Finance Researcher monthly
Why it works
Build one /leaderboards hub on halluminate.ai with a page each for Diligence Bench and DealTrace. Each page gets the current scores by model and harness, the date of each run, a methods summary, a small public sample of tasks (anonymized, as the posts already are), the explorer, a citation block, and a "Have a model or agent to evaluate?" contact path. Rerun scores when major models ship and date-stamp every update. Link the hub from the homepage, the research page and every benchmark post. Then either refresh webbench.ai or mark it archived and point it to the hub.
Labs and agent companies cite benchmarks that are kept up to date in their release notes and papers, and every citation is a free reminder of Halluminate to the few buyers that matter. A live leaderboard also gives the research team a reason to talk to each lab at every model launch. The same page is a credibility signal for finance experts deciding whether to apply, and for firms weighing a data partnership.
Ready to use
ChecklistFinance leaderboard page launch list
[ ] /leaderboards hub linked from homepage and /research [ ] Diligence Bench page: scores by model and harness, run dates [ ] DealTrace page: stage-by-stage scores [ ] Methods summary and citation block on each [ ] Public sample tasks with README and license [ ] "Evaluate your model or agent" contact form [ ] Rerun rule and "last updated" date [ ] webbench.ai refreshed or marked archived
Risks and how to handle them
A public sample lets labs train on the test set
Publish a small sample and keep the scored set private, as private-benchmark leaderboards do.
A stale leaderboard hurts credibility
A written rerun rule and a visible "last updated" date.
Evidence
- Your websiteWestworld Finance Diligence Bench (August 11, 2026), the homepage's "Our latest benchmark": 88 due-diligence problems written and reviewed by finance deal professionals, run in desktop environments over hundreds of steps; the best configuration's mean score is 0.61. The post links an interactive trajectory explorer, but no leaderboard page, dataset, GitHub repository or contact path, and ends: "We will release an accompanying paper soon."Source
- Your websiteDealTrace: "a staged review of real private-equity deals" over ten deal packages and five stages. "Forecast quality is the capability that decides the verdict": supplying the true forecast lifts the judge score "from 53 to 80", and a flagged discrepancy reaches the recommendation "only 7 to 29 percent of the time". The post has no leaderboard, dataset, repository link or call to action, and is dated May 15, 2026 on /research but June 17, 2026 on /blogs.Source
- Your websiteWeb Bench (May 30, 2025), released "in partnership with Skyvern": an official leaderboard at webbench.ai, a Hugging Face dataset, a GitHub repository and a direct contact for teams interested in environments. Every leaderboard entry is still dated 2025-05-30.Source
- HiringMember of Technical Staff, Research / Post-Training: "Lead our public-facing benchmark and leaderboard efforts for frontier models"; "Train open-source frontier models on our in-house environments to validate quality and build research insights"; "Publish findings in blogs and papers to share with the broader research community".Source
- Competitor researchMercor's homepage: "5M+ experts", a "$90/hr average" rate and "$4M+ daily payouts"; benchmark leaderboards with current top scores (APEX-Accounting, τ³-Banking, APEX-SWE and others); open roles shown with hourly pay and recently-hired counts (for example Retail Trader at about $400 an hour).Source
- Competitor researchSurge AI's homepage lists benchmarks including "DAYJOB: Finance" and an Expert Workforce offer; Scale AI's says "Our leaderboards run private benchmarks for the most ambitious AI companies" and that 90% of leading generative AI model builders use its platform.Source