Nitrode growth playbook

Publish the reproduction scripts at the exact URL the paper cites, or, if unrealbench is the public release, rename it or put a README at the cited address that points to it. Add a "Run GameEngineBench" block to /research: repo link, what is public (tasks, starter projects, verifiers) and what stays held out, hardware needs, a BibTeX citation and a "get your model scored" link to the contact form. Update the arXiv listing with a new version that carries the working link, and link the repo README back to nitrode.com/research and the paper. Then track repo traffic, clones and contact-form leads tagged "Research / GameEngineBench" each week.

Benchmarks spread through people running them. A working repo turns each paper reader into a possible user, citer or contributor, and those users are the evals and post-training teams at labs who buy datasets and RL environments. It also keeps the open-source promise, which builds trust with researchers.

[ ] github.com/Nitrode-Research/GameEngineBench loads
[ ] README: what it measures, public vs held-out split, hardware, Epic access, quick start
[ ] License and NOTICE
[ ] BibTeX citation
[ ] Repo links to nitrode.com/research and arXiv
[ ] /research links to the repo
[ ] "Get your model scored" link to /contact
[ ] arXiv version with the working link
[ ] Weekly repo and lead tracking

• Releasing all tasks lets models train on the test set

  Publish scripts and a public split, and keep a held-out split for official scores, as unrealbench already does with solutions.

• Engine licensing limits what can be shared

  Keep the NOTICE file and require users to bring their own Epic-linked access, as unrealbench does.

• Research paperGameEngineBench paper (arXiv, submitted July 3, 2026, revised July 15, 2026; eight authors): tasks from "nine open-source Unreal projects" such as ActionRoguelike, GASShooter and Bomber; Appendix D points readers to github.com/Nitrode-Research/GameEngineBench and says "The current repository already contains the key scripts required to reproduce benchmark runs." It also compares cost and speed across configurations.Source
• GitHubgithub.com/Nitrode-Research/GameEngineBench returns 404 (checked twice on November 9, 2026). The organization's only public repository is "unrealbench": 14 Unreal Engine coding tasks for agents on Harbor, Apache-2.0, with instructions, starter projects and verifiers public; 1 star, 0 forks, and no link to nitrode.com or the paper.Source
• Your websiteGameEngineBench page: "110 UE5 C++ tasks", "9 source projects", "12 model configurations evaluated", "55.5% best pass rate"; every other configuration scores under 30% and 31 tasks are unsolved by all; failure analysis by category (save/persistence, replication, actor lifecycle). The leaderboard rows are not in the page text, only the top score, and the only outbound links are the arXiv paper, X and LinkedIn: no code, dataset download or submission route.Source
• Launch YCLaunch post "Nitrode - Advancing AI for Game Development": Nitrode "will open-source a suite of benchmarks" over the coming months; Mint connects game development experts (engine programmers, technical artists) and has "hundreds of contributors"; the ask is companies that need custom datasets or evals, and introductions to creative or gaming companies; contact [email protected]; model rankings track spatial reasoning benchmarks more closely than traditional agentic coding benchmarks.Source

Render the leaderboard as a plain HTML table: configuration, effort setting, pass@1, compile rate, cost per task and the date scored. Add Dataset structured data, a page title and a meta description that name GameEngineBench and UE5 C++. Add a "last updated" date and an anchor link for each row so people can share a single result.

A leaderboard that can be quoted gets cited in model comparisons, newsletters and AI answers. Each citation sends researchers and lab engineers to the page.

• Others copy the table without credit

  Add a citation line and the paper link under the table.

• Your websiteGameEngineBench page: "110 UE5 C++ tasks", "9 source projects", "12 model configurations evaluated", "55.5% best pass rate"; every other configuration scores under 30% and 31 tasks are unsolved by all; failure analysis by category (save/persistence, replication, actor lifecycle). The leaderboard rows are not in the page text, only the top score, and the only outbound links are the arXiv paper, X and LinkedIn: no code, dataset download or submission route.Source
• PageSpeedGoogle PageSpeed (November 9, 2026): mobile 96, desktop 100; accessibility, best practices and SEO 100. Real visitors pass Core Web Vitals (75th percentile LCP 0.64s, INP 44ms, CLS 0); the largest lab saving is the hero background image (about 317 KB).Source

Add an "Experts" page and a nav link: who Nitrode needs (engine programmers, technical artists, Blender and Unreal specialists, AAA veterans), what the work looks like, how it is paid, and an apply button to the existing contributor platform. Add "Contribute as an expert" to the contact form's Interest list. Link it from the Game Developer job post, since applicants who don't fit the role may still be strong contributors.

Supply is the constraint in specialist data work. Game developers who search for paid AI work or click from the paper should be able to sign up without guessing where to go.

• A flood of unqualified applicants

  Ask for a portfolio and engine experience upfront.

• Your websiteContact page: a form only (Name, Email, Organization, Interest, Message, Website). Interest options are Research / GameEngineBench, Products for labs, Enterprise, Careers and Other, with none for experts who want to contribute; no email address, booking link or promised response time.Source
• Your websiteFor labs page: coding datasets and evaluations for game engines, C++, CUDA, Rust and debugging, plus traces of how experienced game developers work; tool-calling RL environments; computer-use trajectories for Blender, Unreal, Resolume, SAP and Jira; multimodal datasets; an expert network of "hundreds of game developers and creative software specialists". No pricing, process, sample data, named customers or case studies; the main CTA is "See GameEngineBench".Source
• Your websiteSitemap (checked November 9, 2026): six URLs (/, /research, /products, /enterprise, /careers, /contact), all last modified October 15, 2026; no blog, case studies, documentation, sample data or page for experts who want to contribute.Source

Update the Product Hunt product description and maker comment to say the engine was the origin and the company now builds GameEngineBench and expert data. Then do the same on every directory and profile that still describes the engine (X bio, LinkedIn About, YC profile, startup directories). Keep the history; change the present tense.

Buyers check a vendor's name before they reply. Consistent profiles make the pivot look deliberate, and the old launch becomes proof that the team has built games.

• Old engine users feel abandoned

  Say plainly what happened to the engine and where to go now.

• Product HuntProduct Hunt still lists Nitrode as an "AI game engine to prototype 3D games in a day", launched July 28, 2025 (#3 Day Rank, 318 points), linking to www.nitrode.com.Source
• Site historyChange tracking since December 4, 2025: the homepage headline changed from "The Ultimate AI Game Engine." to "Advancing game development through frontier AI research.", and the site dropped Next.js and React.Source

Choose one of three paths. Keep Familiar as a showcase of Nitrode's multimodal and transcription work, and say so on both sites. Move it to its own brand and domain. Or sunset it with notice to users. Whatever the choice, fix the /about link today (point it to the homepage or a real About page) and make sure Familiar's footer explains the relationship.

A clear portfolio keeps lab buyers' attention on the core offer. Fixing a dead link from a live product is free.

• Losing a small paying user base

  Spin out rather than shut down if usage is real.

• Your websitefamiliar.nitrode.com is a live Nitrode page for Familiar, a D&D assistant ("Your DnD Session's Second Brain") with Play Now, Sign In, Pricing and a Discord link; its "About Nitrode" link points to nitrode.com/about, which returns 404 (checked November 9, 2026).Source
• Your websiteHomepage: H1 "Advancing game development through frontier AI research."; meta description "Nitrode is building the evaluation, data, and post-training layer for specialized AI systems in game development."; sections for Research (GameEngineBench), For labs ("contractor-sourced UE5 data and RL environments") and For enterprises; badges "SOC 2 Type I" and "Backed by Y Combinator"; footer "© 2025 Inception Technologies"; no privacy policy or terms link.Source

Add a /trust page: SOC 2 Type I scope and how to request the report, plans for Type II, data handling for customer traces (storage, retention, access, deletion), contributor confidentiality and how experts are vetted. Link a privacy policy and terms in the footer, update the year and add "Nitrode is a product of Inception Technologies, Inc." (or the correct entity).

Security review is the slowest step in an enterprise deal. A trust page answers the first questions before they are asked, and a current footer signals an active company.

• Overstating security

  Describe only controls in the audited scope.

• Your websiteHomepage: H1 "Advancing game development through frontier AI research."; meta description "Nitrode is building the evaluation, data, and post-training layer for specialized AI systems in game development."; sections for Research (GameEngineBench), For labs ("contractor-sourced UE5 data and RL environments") and For enterprises; badges "SOC 2 Type I" and "Backed by Y Combinator"; footer "© 2025 Inception Technologies"; no privacy policy or terms link.Source
• Your websiteEnterprise page: custom evaluation frameworks for gaming and creative enterprises; routing harnesses that work with existing coding agents to cut costs; post-training that turns proprietary workflow traces into SFT datasets and RL environments for private, domain-specific models; CTA "If your team is evaluating or deploying foundation models for specialized workflows, let's talk." No customers, pricing, timelines or case studies.Source

After the Interest field, show a reply promise ("We reply within one business day") and, for Products for labs and Enterprise, a booking link for a 30-minute scoping call with a founder. Show [email protected], the address used in the YC launch, for researchers. Route each Interest option to the right owner and set an alert for anything not answered within a day.

Fast, specific replies win early B2B deals, and a booking link removes a round of email.

• Founders' calendars fill with poor-fit calls

  Show the booking link only after Organization is filled in.

• Your websiteContact page: a form only (Name, Email, Organization, Interest, Message, Website). Interest options are Research / GameEngineBench, Products for labs, Enterprise, Careers and Other, with none for experts who want to contribute; no email address, booking link or promised response time.Source
• Launch YCLaunch post "Nitrode - Advancing AI for Game Development": Nitrode "will open-source a suite of benchmarks" over the coming months; Mint connects game development experts (engine programmers, technical artists) and has "hundreds of contributors"; the ask is companies that need custom datasets or evals, and introductions to creative or gaming companies; contact [email protected]; model rankings track spatial reasoning benchmarks more closely than traditional agentic coding benchmarks.Source

Add a "Try a sample environment" section to /products: what unrealbench contains, how to run one task in an hour, and how a paid environment differs (scale, held-out tasks, expert traces, custom categories). Give the repo a README section called "Need more? Custom environments from Nitrode" that links to the labs page. Offer a guided run on a call for lab teams.

Lab data teams judge vendors by running their samples. A runnable sample shortens evaluation and shows the verifier quality that sets Nitrode apart.

• The sample sets the price expectation at zero

  Show clearly what only the paid work includes.

• #1 Make the GameEngineBench code link in the paper work, and give the benchmark a public home on nitrode.com

• GitHubgithub.com/Nitrode-Research/GameEngineBench returns 404 (checked twice on November 9, 2026). The organization's only public repository is "unrealbench": 14 Unreal Engine coding tasks for agents on Harbor, Apache-2.0, with instructions, starter projects and verifiers public; 1 star, 0 forks, and no link to nitrode.com or the paper.Source
• Your websiteFor labs page: coding datasets and evaluations for game engines, C++, CUDA, Rust and debugging, plus traces of how experienced game developers work; tool-calling RL environments; computer-use trajectories for Blender, Unreal, Resolume, SAP and Jira; multimodal datasets; an expert network of "hundreds of game developers and creative software specialists". No pricing, process, sample data, named customers or case studies; the main CTA is "See GameEngineBench".Source
• Competitor researchRivals' own homepages: Mercor, "organizing human intelligence to power the AI economy", with an expert network and benchmarks; Surge AI, "Human Intelligence for AGI"; Turing, "Accelerate superintelligence", which invites visitors to request dataset samples or join its expert network and lists RL environments.See the research Source

Publish a "Get scored" process: a model team submits an endpoint or weights under NDA, Nitrode runs the held-out split, and the team gets a private report of pass@1 by category, the hardest failed tasks and example traces. Public listing is optional. The report ends with matched offers: RL environments and expert traces for the categories where the model failed.

It turns the benchmark from a press asset into the first step of a sale, and the private report gives the team something specific to act on.

• Scoring looks like a conflict of interest when Nitrode also sells training data

  Keep a held-out split that no customer data ever touches, and publish the policy.

• #1 Make the GameEngineBench code link in the paper work, and give the benchmark a public home on nitrode.com

• Your websiteGameEngineBench page: "110 UE5 C++ tasks", "9 source projects", "12 model configurations evaluated", "55.5% best pass rate"; every other configuration scores under 30% and 31 tasks are unsolved by all; failure analysis by category (save/persistence, replication, actor lifecycle). The leaderboard rows are not in the page text, only the top score, and the only outbound links are the arXiv paper, X and LinkedIn: no code, dataset download or submission route.Source
• Your websiteFor labs page: coding datasets and evaluations for game engines, C++, CUDA, Rust and debugging, plus traces of how experienced game developers work; tool-calling RL environments; computer-use trajectories for Blender, Unreal, Resolume, SAP and Jira; multimodal datasets; an expert network of "hundreds of game developers and creative software specialists". No pricing, process, sample data, named customers or case studies; the main CTA is "See GameEngineBench".Source

Set a standing rule: when a major model or agent ships, run the public and held-out splits within seven days, update the leaderboard with a dated row, and post a short note on what changed by category. Keep a changelog on /research.

Release-week results get shared while attention is highest, and a living leaderboard gives people a reason to come back and to cite Nitrode by name.

• Rushed runs produce wrong numbers

  Mark scores "preliminary" until the judge audit is done.

• #2 Put the GameEngineBench leaderboard rows in the page text, so search engines and AI assistants can quote them

• Research paperGameEngineBench paper (arXiv, submitted July 3, 2026, revised July 15, 2026; eight authors): tasks from "nine open-source Unreal projects" such as ActionRoguelike, GASShooter and Bomber; Appendix D points readers to github.com/Nitrode-Research/GameEngineBench and says "The current repository already contains the key scripts required to reproduce benchmark runs." It also compares cost and speed across configurations.Source
• Your websiteGameEngineBench page: "110 UE5 C++ tasks", "9 source projects", "12 model configurations evaluated", "55.5% best pass rate"; every other configuration scores under 30% and 31 tasks are unsolved by all; failure analysis by category (save/persistence, replication, actor lifecycle). The leaderboard rows are not in the page text, only the top score, and the only outbound links are the arXiv paper, X and LinkedIn: no code, dataset download or submission route.Source

Start /research/notes with a monthly post on one failure category (replication, actor lifecycle, save/load, gameplay ability systems). Show the task, what the agents tried, why it broke and what an experienced Unreal developer does differently. Have the Game Developer hire write or review each one, and share it where Unreal developers gather.

Specific, technical failure stories travel in developer communities, attract expert contributors and show labs the depth of Nitrode's judgment.

• Leaking held-out tasks

  Write only about public tasks.

• Your websiteGameEngineBench page: "110 UE5 C++ tasks", "9 source projects", "12 model configurations evaluated", "55.5% best pass rate"; every other configuration scores under 30% and 31 tasks are unsolved by all; failure analysis by category (save/persistence, replication, actor lifecycle). The leaderboard rows are not in the page text, only the top score, and the only outbound links are the arXiv paper, X and LinkedIn: no code, dataset download or submission route.Source
• Your websiteSitemap (checked November 9, 2026): six URLs (/, /research, /products, /enterprise, /careers, /contact), all last modified October 15, 2026; no blog, case studies, documentation, sample data or page for experts who want to contribute.Source
• HiringCareers page (November 9, 2026), four open full-time roles on the Y Combinator job board: Game Developer (San Francisco or remote US, $95K-$145K, Unreal/Unity, "Triple-A game development" preferred), Full-Stack Engineer ($110K-$160K), Member of Technical Staff, Evals ($120K-$250K, 1-3% equity) and Member of Technical Staff, Post-Training ($150K-$350K, 1-3% equity). YC lists a team of 8, founded 2024, batch W25.Source

Add a chart to /enterprise and /research: cost per solved task and time per solved task for each configuration, by task category. Add a worked example: a studio running 1,000 agent tasks a month, with costs for the most accurate setting, the cheapest acceptable one and a routed mix. End with "Get this analysis on your own codebase".

Engineering leaders buy on cost per useful outcome. Showing the trade-off with Nitrode's own data makes routing concrete.

• Model prices change quickly

  Date the chart and show the price assumptions.

• Your websiteEnterprise page: custom evaluation frameworks for gaming and creative enterprises; routing harnesses that work with existing coding agents to cut costs; post-training that turns proprietary workflow traces into SFT datasets and RL environments for private, domain-specific models; CTA "If your team is evaluating or deploying foundation models for specialized workflows, let's talk." No customers, pricing, timelines or case studies.Source
• Research paperGameEngineBench paper (arXiv, submitted July 3, 2026, revised July 15, 2026; eight authors): tasks from "nine open-source Unreal projects" such as ActionRoguelike, GASShooter and Bomber; Appendix D points readers to github.com/Nitrode-Research/GameEngineBench and says "The current repository already contains the key scripts required to reproduce benchmark runs." It also compares cost and speed across configurations.Source

Offer "GameEngineBench on your codebase": in about four weeks, Nitrode's experts write 20-30 tasks from the studio's own engine code, score 4-6 model and agent configurations, and deliver a report with a routing recommendation and a cost estimate. Set one public starting price and credit it toward post-training work. Put the offer, its steps and a sample report outline on /enterprise.

A fixed scope lowers the risk of a first purchase for studios that are new to AI evaluation, and the report leads naturally to routing and post-training.

Week 1: kickoff, pick 3 workflows, sign data terms
Week 2: experts write 20-30 tasks with verifiers
Week 3: score 4-6 model/agent configurations
Week 4: report: pass@1 by category, failure examples, cost per solved task, routing recommendation, post-training estimate
Price: one fixed fee, credited toward post-training

• Studios won't share code

  Offer on-premises runs and tasks written from interface specs.

• Your websiteEnterprise page: custom evaluation frameworks for gaming and creative enterprises; routing harnesses that work with existing coding agents to cut costs; post-training that turns proprietary workflow traces into SFT datasets and RL environments for private, domain-specific models; CTA "If your team is evaluating or deploying foundation models for specialized workflows, let's talk." No customers, pricing, timelines or case studies.Source
• Launch YCLaunch post "Nitrode - Advancing AI for Game Development": Nitrode "will open-source a suite of benchmarks" over the coming months; Mint connects game development experts (engine programmers, technical artists) and has "hundreds of contributors"; the ask is companies that need custom datasets or evals, and introductions to creative or gaming companies; contact [email protected]; model rankings track spatial reasoning benchmarks more closely than traditional agentic coding benchmarks.Source

For each product line, publish a short data card: one redacted example record (a developer trace, a computer-use trajectory in Blender or Unreal, an RL task with its verifier), how experts are selected, how each item is reviewed, typical volume per month and the formats delivered. Add a "Request a sample pack" button that routes to the labs inbox.

Data cards answer the first technical questions before a call and show that Nitrode's quality process is real.

• Revealing client work

  Use only internally made or public-task examples.

• #8 Present unrealbench as the free sample of Nitrode's RL environments

• Your websiteFor labs page: coding datasets and evaluations for game engines, C++, CUDA, Rust and debugging, plus traces of how experienced game developers work; tool-calling RL environments; computer-use trajectories for Blender, Unreal, Resolume, SAP and Jira; multimodal datasets; an expert network of "hundreds of game developers and creative software specialists". No pricing, process, sample data, named customers or case studies; the main CTA is "See GameEngineBench".Source
• Competitor researchRivals' own homepages: Mercor, "organizing human intelligence to power the AI economy", with an expert network and benchmarks; Surge AI, "Human Intelligence for AGI"; Turing, "Accelerate superintelligence", which invites visitors to request dataset samples or join its expert network and lists RL environments.See the research Source

Add an anonymized breakdown of the network to /products and the experts page: counts by discipline (engine programmers, gameplay, technical art, animation, rendering, tools), by engine (Unreal, Unity, Godot, Blender), years of experience, the share with shipped AAA or indie titles, and the vetting steps. Refresh it every quarter.

Specific numbers make the specialist claim checkable, and they help experts see that they would be working alongside peers.

• Numbers look small next to generalist vendors

  Lead with depth (years, shipped titles) rather than headcount.

• #3 Add a "join the expert network" path to nitrode.com and the contact form

• Your websiteFor labs page: coding datasets and evaluations for game engines, C++, CUDA, Rust and debugging, plus traces of how experienced game developers work; tool-calling RL environments; computer-use trajectories for Blender, Unreal, Resolume, SAP and Jira; multimodal datasets; an expert network of "hundreds of game developers and creative software specialists". No pricing, process, sample data, named customers or case studies; the main CTA is "See GameEngineBench".Source
• Competitor researchRivals' own homepages: Mercor, "organizing human intelligence to power the AI economy", with an expert network and benchmarks; Surge AI, "Human Intelligence for AGI"; Turing, "Accelerate superintelligence", which invites visitors to request dataset samples or join its expert network and lists RL environments.See the research Source

Add a "Source projects" section to /research that credits each project with a link and license. Write to each maintainer with how their project is used and an offer: early access to results on their project, a "used in GameEngineBench" badge for their README and an invitation to join the expert network. Respect any maintainer who prefers not to be listed.

Crediting sources is good research practice, and maintainers who add a badge or mention send Unreal developers back to Nitrode.

• A maintainer objects to the use

  Check licenses first and honor opt-out requests.

• #1 Make the GameEngineBench code link in the paper work, and give the benchmark a public home on nitrode.com

• Research paperGameEngineBench paper (arXiv, submitted July 3, 2026, revised July 15, 2026; eight authors): tasks from "nine open-source Unreal projects" such as ActionRoguelike, GASShooter and Bomber; Appendix D points readers to github.com/Nitrode-Research/GameEngineBench and says "The current repository already contains the key scripts required to reproduce benchmark runs." It also compares cost and speed across configurations.Source
• Your websiteGameEngineBench page: "110 UE5 C++ tasks", "9 source projects", "12 model configurations evaluated", "55.5% best pass rate"; every other configuration scores under 30% and 31 tasks are unsolved by all; failure analysis by category (save/persistence, replication, actor lifecycle). The leaderboard rows are not in the page text, only the top score, and the only outbound links are the arXiv paper, X and LinkedIn: no code, dataset download or submission route.Source

Invite five to eight technical directors or engine leads from studios of different sizes to a quarterly advisory group. They propose task categories from real pain points, see results two weeks early and get a private readout. They don't pay, and Nitrode commits to nothing commercial. The group becomes a natural source of pilots.

Shaping the research gives studio leaders a reason to engage now, and Nitrode gets task ideas grounded in production.

• Members' legal teams limit participation

  Keep it advisory, with no data sharing required.

• Launch YCLaunch post "Nitrode - Advancing AI for Game Development": Nitrode "will open-source a suite of benchmarks" over the coming months; Mint connects game development experts (engine programmers, technical artists) and has "hundreds of contributors"; the ask is companies that need custom datasets or evals, and introductions to creative or gaming companies; contact [email protected]; model rankings track spatial reasoning benchmarks more closely than traditional agentic coding benchmarks.Source
• Your websiteEnterprise page: custom evaluation frameworks for gaming and creative enterprises; routing harnesses that work with existing coding agents to cut costs; post-training that turns proprietary workflow traces into SFT datasets and RL environments for private, domain-specific models; CTA "If your team is evaluating or deploying foundation models for specialized workflows, let's talk." No customers, pricing, timelines or case studies.Source

Publish a fair page on choosing a data vendor for game and creative-software work: which tasks need engine-level judgment, how verifiers run in real Unreal builds, why runtime tests beat reference diffs, and where a generalist network is fine. Use GameEngineBench failure examples as proof. Avoid claims about competitors that can't be verified.

Labs comparing vendors search for this, and a clear specialist case keeps Nitrode on shortlists where it would otherwise lose on brand.

• Reads as an attack on bigger vendors

  Frame it as a guide for choosing, not a takedown.

• Competitor researchRivals' own homepages: Mercor, "organizing human intelligence to power the AI economy", with an expert network and benchmarks; Surge AI, "Human Intelligence for AGI"; Turing, "Accelerate superintelligence", which invites visitors to request dataset samples or join its expert network and lists RL environments.See the research Source
• Your websiteGameEngineBench page: "110 UE5 C++ tasks", "9 source projects", "12 model configurations evaluated", "55.5% best pass rate"; every other configuration scores under 30% and 31 tasks are unsolved by all; failure analysis by category (save/persistence, replication, actor lifecycle). The leaderboard rows are not in the page text, only the top score, and the only outbound links are the arXiv paper, X and LinkedIn: no code, dataset download or submission route.Source
• Your websiteSitemap (checked November 9, 2026): six URLs (/, /research, /products, /enterprise, /careers, /contact), all last modified October 15, 2026; no blog, case studies, documentation, sample data or page for experts who want to contribute.Source

If the contributor platform doesn't already pay for referrals, add a referral bonus paid once the referred expert's first accepted task passes review. Set higher bonuses for the disciplines with the longest backlog. Show a "Most needed this month" list on the experts page and in contributor updates.

Referrals from vetted experts arrive pre-screened and match the skills the current projects need.

• Low-quality referrals for the bonus

  Pay only after the first task passes review.

• #3 Add a "join the expert network" path to nitrode.com and the contact form

• Launch YCLaunch post "Nitrode - Advancing AI for Game Development": Nitrode "will open-source a suite of benchmarks" over the coming months; Mint connects game development experts (engine programmers, technical artists) and has "hundreds of contributors"; the ask is companies that need custom datasets or evals, and introductions to creative or gaming companies; contact [email protected]; model rankings track spatial reasoning benchmarks more closely than traditional agentic coding benchmarks.Source
• Your websiteFor labs page: coding datasets and evaluations for game engines, C++, CUDA, Rust and debugging, plus traces of how experienced game developers work; tool-calling RL environments; computer-use trajectories for Blender, Unreal, Resolume, SAP and Jira; multimodal datasets; an expert network of "hundreds of game developers and creative software specialists". No pricing, process, sample data, named customers or case studies; the main CTA is "See GameEngineBench".Source

Build a benchmark of tasks in the Unreal and Blender editors (place and light a scene, rig and export an asset, fix a material or a broken level) with programmatic checks on the saved project. Keep a held-out split and launch it with the same paper, repo and leaderboard pattern. Paid computer-use trajectories then come from the same task design.

A second benchmark makes the "suite" real and extends Nitrode's authority into the computer-use data it already sells.

• Editor automation is flaky

  Score saved files, not screenshots.

• #1 Make the GameEngineBench code link in the paper work, and give the benchmark a public home on nitrode.com

• Launch YCLaunch post "Nitrode - Advancing AI for Game Development": Nitrode "will open-source a suite of benchmarks" over the coming months; Mint connects game development experts (engine programmers, technical artists) and has "hundreds of contributors"; the ask is companies that need custom datasets or evals, and introductions to creative or gaming companies; contact [email protected]; model rankings track spatial reasoning benchmarks more closely than traditional agentic coding benchmarks.Source
• Your websiteFor labs page: coding datasets and evaluations for game engines, C++, CUDA, Rust and debugging, plus traces of how experienced game developers work; tool-calling RL environments; computer-use trajectories for Blender, Unreal, Resolume, SAP and Jira; multimodal datasets; an expert network of "hundreds of game developers and creative software specialists". No pricing, process, sample data, named customers or case studies; the main CTA is "See GameEngineBench".Source

Publish a yearly report: leaderboard history, progress by category, the tasks still unsolved, cost trends, and a survey of the expert network on where they use AI today. Release it with a launch post, a webinar for studios and a short version for press. Gate the full data appendix behind an email for follow-up.

An annual report gets cited all year, starts sales conversations with both audiences and grows a list of studio and lab readers.

• Too few data points in year one

  Pair the leaderboard with the expert survey.

• #10 Score every major new model on GameEngineBench within a week of its release

• Your websiteGameEngineBench page: "110 UE5 C++ tasks", "9 source projects", "12 model configurations evaluated", "55.5% best pass rate"; every other configuration scores under 30% and 31 tasks are unsolved by all; failure analysis by category (save/persistence, replication, actor lifecycle). The leaderboard rows are not in the page text, only the top score, and the only outbound links are the arXiv paper, X and LinkedIn: no code, dataset download or submission route.Source
• Launch YCLaunch post "Nitrode - Advancing AI for Game Development": Nitrode "will open-source a suite of benchmarks" over the coming months; Mint connects game development experts (engine programmers, technical artists) and has "hundreds of contributors"; the ask is companies that need custom datasets or evals, and introductions to creative or gaming companies; contact [email protected]; model rankings track spatial reasoning benchmarks more closely than traditional agentic coding benchmarks.Source

After a pilot, offer a subscription: Nitrode keeps the studio's private task set, re-scores it within a week of each major release, updates the routing configuration and sends a short readout. Higher tiers add quarterly task refreshes by experts and continuous post-training on the studio's traces.

It turns one-off pilots into recurring revenue and makes Nitrode part of each studio's model decisions.

• Studios want results without paying for upkeep

  Price the first pilot to include one quarter of re-scoring.

• #13 Package the enterprise offer as a fixed-scope, fixed-price studio evaluation pilot

• Your websiteEnterprise page: custom evaluation frameworks for gaming and creative enterprises; routing harnesses that work with existing coding agents to cut costs; post-training that turns proprietary workflow traces into SFT datasets and RL environments for private, domain-specific models; CTA "If your team is evaluating or deploying foundation models for specialized workflows, let's talk." No customers, pricing, timelines or case studies.Source
• Research paperGameEngineBench paper (arXiv, submitted July 3, 2026, revised July 15, 2026; eight authors): tasks from "nine open-source Unreal projects" such as ActionRoguelike, GASShooter and Bomber; Appendix D points readers to github.com/Nitrode-Research/GameEngineBench and says "The current repository already contains the key scripts required to reproduce benchmark runs." It also compares cost and speed across configurations.Source

Send one honest update to the engine's sign-up list, the Product Hunt followers and the Familiar Discord: what Nitrode does now, and an invitation for experienced developers, technical artists and Blender or Unreal users to apply as paid contributors. Track applications from each source and accept only on the usual vetting.

Warm communities convert better than cold job posts, and the update also closes the loop with early users.

• Hobbyists outnumber professionals

  State experience requirements clearly.

• #3 Add a "join the expert network" path to nitrode.com and the contact form
• #4 Retire the "AI game engine" description on Product Hunt and other old profiles

• Product HuntProduct Hunt still lists Nitrode as an "AI game engine to prototype 3D games in a day", launched July 28, 2025 (#3 Day Rank, 318 points), linking to www.nitrode.com.Source
• Your websitefamiliar.nitrode.com is a live Nitrode page for Familiar, a D&D assistant ("Your DnD Session's Second Brain") with Play Now, Sign In, Pricing and a Discord link; its "About Nitrode" link points to nitrode.com/about, which returns 404 (checked November 9, 2026).Source

Run the same small set of agents on both benchmarks and publish a short side-by-side: where rankings agree, where they diverge, and what each measures (tutorial-scale tasks versus engine-scale C++). Invite the GameDevBench authors to co-write or comment. Link both ways.

A joint view positions GameEngineBench as the engine-scale counterpart in a recognised category, and attaches Nitrode to the academic attention around game-dev agents.

• The authors decline

  Publish a fair independent comparison with their public tasks, citing them properly.

• #1 Make the GameEngineBench code link in the paper work, and give the benchmark a public home on nitrode.com

• Competitor researchGameDevBench (arXiv, submitted February 11, 2026, revised June 30, 2026): 333 game development tasks derived from web and video tutorials; the best agent and method solves 53.8%, and visual feedback raises scores.See the research Source
• Your websiteGameEngineBench page: "110 UE5 C++ tasks", "9 source projects", "12 model configurations evaluated", "55.5% best pass rate"; every other configuration scores under 30% and 31 tasks are unsolved by all; failure analysis by category (save/persistence, replication, actor lifecycle). The leaderboard rows are not in the page text, only the top score, and the only outbound links are the arXiv paper, X and LinkedIn: no code, dataset download or submission route.Source