ProgrammableMarketing/mlcommons.org (opens in a new tab)
PDF

Views

  • Plan
  • Roadmap
  • Research

On this page

  • The brief
  • Start here
  • Idea portfolio
  • Effort vs impact
  • This week · 7
  • This quarter · 9
  • Big bets · 3
  • Wildcards · 3
  • How we can help
Generated Nov 4, 2026
  1. Programmable
  2. Playbooks
  3. MLCommons
MLCommons logo

growth playbook · Nov 4, 2026

mlcommons.org (opens in a new tab)

MLCommons growth playbook

Diagnosis

MLCommons has the credibility most AI testing vendors lack: 125+ members including the largest AI companies, 89.7k+ MLPerf results and consensus-built AILuminate benchmarks. Its hiring shows the turn to selling: a Marketing Director for its first paid products through direct sales and channel partners, and a General Manager to own MLPerf's P&L.

The website has not caught up: Candela can only be bought by email, its price table is still labelled a preview, membership is requested by email under a 2025 dues notice, and the homepage offers buyers no path.

This playbook fixes the buying path first, then builds buyer pages, partner channels and member amplification, and bets on insurers and procurement making the score a requirement.

Growth Score

6/10

Biggest Opportunity

Turn the consortium's neutral, consensus-built scores into a product buyers can order in minutes, and get the score written into procurement and insurance so demand comes to MLCommons.

Biggest Risk

A small, engineering-minded team moving into commercial sales can stall on manual onboarding while commercial AI testing vendors sell faster, and selling must not weaken the neutrality that makes the benchmarks trusted.

01

Start here

The three moves to make first. Each one is written up in full below.

  1. 1

    Replace Candela's "email us to get started" with a booking form and a $200 practice-test checkout

    Add a five-field request form above the email address on the getting-started page.

    High impact First result in 2 weeks
  2. 2

    Give buyers a path on the homepage, not just "Get Involved"

    Draft the two new hero buttons and their landing pages.

    High impact First result in 2 weeks
  3. 3

    Remove the June 2025 dues notice and add a membership enquiry form

    Delete the dated notice today.

    Medium impact First result in 2 weeks

02

Your idea portfolio

22 ideas, grouped by when to run them. Open any idea for how to run it: the steps, who to approach, copy you can use, and how to measure it.

  • 6 ideas are marked Unexpected: plays you are unlikely to have tried.

Every idea placed by the effort it takes against the impact it could have. Quick wins sit top left; select a number to open that idea. The filters below narrow the list by status, timing, category and what Programmable can build.

Effort vs impactEach number is an idea. Select one to open it.
Impact →
Quick wins · 3Big bets · 9Fill-ins · 7Reconsider · 3
Effort →

Showing 22 of 22 ideas

This week 7

Quick wins with a clear first action.

  • 01

    The Marketing Director post says MLCommons is selling paid products "for the first time" and the role owns everything "from brand to launch to pipeline". Candela is that product, and its pricing is already public, but the only way to buy is an email address: "Automated signup and billing are coming soon." The table is still headed "Candela pricing table preview", and the Candela page's "See the assessments" button opens an unpublished preview link instead of a published page. A developer ready to pay $200 for a practice test has no button to press.

    High impactLow effort First result in 2 weeks Marketing Director

    First action

    Add a five-field request form above the email address on the getting-started page.

    How to measure it

    Candela requests

    Now
    –
    Target
    40 a month
    By
    60 days

    Steps

    1. 1

      Fix the labels

      Rename "pricing table preview" and remove the stray "- " in the Candela title.

      Web and content manager 1 hour

    2. 2

      Fix the button

      Point "See the assessments" at the published results page.

      Web and content manager 1 hour

    3. 3

      Build the form

      Path, system, member status, tier and timing.

      Web and content manager 2 days

    4. 4

      Add a payment link

      For $200 and $300 non-member practice tests.

      Finance and operations 2 days

    5. 5

      Route in HubSpot

      Tag each request by path and tier; alert the Candela team.

      Marketing Director 1 day

    6. 6

      Add the button everywhere

      Candela page, launch post, results pages and newsletter.

      Marketing Director 1 day

    7. 7

      Report weekly

      Requests, paid practice tests and upgrades to official results.

      Marketing Director weekly

    First result
    2 weeks
    Cost
    Staff time and a payment link
    Owner
    Marketing Director

    Smallest test

    Put the form live and count requests for two weeks against the email inbox.

    Why it works

    Until automated billing ships, put a short request form on the getting-started page (developer, deployer or assurance vendor; model or system to test; member or not; test tier) that books a setup call and, for non-member practice tests, sends a payment link straight away. Rename the table heading and page title, point "See the assessments" to the published results page, and add one clear "Book a test" button to the Candela page and the launch post. Log every request in HubSpot by path and tier.

    The cheapest test is the first step to a $6,000-$13,000 official result. A form converts better than an email address, tells sales which buyer and tier each lead is, and gives the new marketing lead a pipeline number from week one.

    Ready to use

    ChecklistCandela getting-started fixes

    [ ] Heading and title: "Candela pricing" (not "preview")
    [ ] Candela page title without the leading "- "
    [ ] "See the assessments" goes to the published results page
    [ ] Request form: path, system, member?, tier, timing
    [ ] Payment link for non-member practice tests
    [ ] HubSpot source and tier tags
    [ ] "Book a test" on Candela, launch post and results pages

    Risks and how to handle them

    • Requests outpace the team running setups

      Offer set setup slots each week and say so on the form.

    • Members try to pay

      Ask for member status first and skip payment for members.

    Evidence

    • HiringMarketing Director (2026 posting, $175-200k, remote): "For the first time, we are offering paid products through direct sales and channel partners." The role manages "both a member-driven organization and an enterprise SaaS company in one, from brand to launch to pipeline", shifting "from a primarily standards-oriented, community-centric communications model to one that incorporates B2B product marketing"; it must unlock members' "own marketing and communications" and drive "membership growth"; "we do not have a consumer-facing brand nor a strong PR practice"; "we have some budget to build a team". The post says MLCommons has "over 150 members".Source
    • Your websiteGetting Started with Candela: "We charge per run of the test. MLCommons members get free and discounted tests." "Automated signup and billing are coming soon. In the meantime, please email us to get started." The price table is headed "Candela pricing table preview" (also the page title): Safety practice test $200 for non-members (free for members), annotated $4,800, official $6,000; Jailbreak official $8,000; Safety + Jailbreak bundle official $13,000. Users get started via three paths: "a developer, a deployer. or a third-party assurance vendor".Source
    • Your websiteCandela page: "Benchmarking-as-a-Service for AI Risk and Reliability", "Evaluations on demand for developers and enterprise deployers", "official results for marketing and decision-making"; developers "commission an official result for a score they can put in a system card or spec sheet". Its "See the assessments" button opens an unpublished page-preview address rather than a published page, and the page title begins with a stray "- ".Source
    • ChannelsYouTube @mlcommons: 430 subscribers, 54 videos, 12,908 total views. Also a Discord server, GitHub, X and LinkedIn; HubSpot is verified on the domain and the footer has a newsletter signup.Source
  • 02

    The homepage now promises to "enable informed decisions", yet its only button is "Get Involved", which leads to membership. Candela and the MLPerf Endpoints buyer's guide, the two things a buyer would act on, appear only as news items.

    High impactLow effort First result in 2 weeks Marketing Director

    First action

    Draft the two new hero buttons and their landing pages.

    How to measure it

    Homepage-to-product clicks

    Now
    –
    Target
    3x
    By
    30 days

    Steps

    1. 1

      Agree the paths

      Members, testers and buyers.

      Marketing Director 1 day

    2. 2

      Write the copy

      Two buttons and a buyer band.

      Marketing Director 1 day

    3. 3

      Build

      Hero and band in the theme.

      Web and content manager 2 days

    4. 4

      Tag clicks

      Events in Google Analytics.

      Web and content manager 1 day

    5. 5

      Review

      Click share per path after two weeks.

      Marketing Director 2 weeks

    6. 6

      Iterate

      Swap copy on the weaker path.

      Marketing Director 1 week

    First result
    2 weeks
    Cost
    $0
    Owner
    Marketing Director

    Smallest test

    Add one "Test your AI system" button for two weeks.

    Why it works

    Add a second hero path: "Test your AI system" (Candela) and "Compare inference endpoints" (MLPerf Endpoints results), with the member path kept beside them. Add a short "For buyers" band under the numbers with the three buyer types the Endpoints guide names.

    Visitors arrive with different jobs; showing the paying path on the first screen is the simplest way to grow product traffic without new content.

    Risks and how to handle them

    • Members feel sidelined

      Keep "Get Involved" as the first button.

    Evidence

    • Your websiteHomepage: H1 "Standardizing AI measurement and enabling informed decisions" (a year ago it read "Better AI for Everyone"); the only hero button is "Get Involved"; "125+ MLCommons Members and Affiliates", "10 Benchmark Suites", "89.7k+ MLPerf Performance Results to-date", "700k Datasets using the Croissant metadata vocabulary". Candela appears only in the Latest Insights list.Source
    • Your websiteGetting Started with Candela: "We charge per run of the test. MLCommons members get free and discounted tests." "Automated signup and billing are coming soon. In the meantime, please email us to get started." The price table is headed "Candela pricing table preview" (also the page title): Safety practice test $200 for non-members (free for members), annotated $4,800, official $6,000; Jailbreak official $8,000; Safety + Jailbreak bundle official $13,000. Users get started via three paths: "a developer, a deployer. or a third-party assurance vendor".Source
    • Your websiteMLPerf Endpoints buyer's guide (October 2026): "Endpoints runs on rolling submissions with publication cohorts every few weeks"; results carry a status: "verified", "provisional" or "unverified"; written for "Enterprise leaders", "AI infrastructure decision-makers" and "Model decision-makers".Source
  • 03

    The Get Involved page still warns that dues "will be increasing from June 1st, 2025", 17 months after the date, and both membership and affiliate status are requested by email. The Marketing Director is asked to drive "membership growth".

    Medium impactLow effort First result in 2 weeks Membership and partnerships l…

    First action

    Delete the dated notice today.

    How to measure it

    Membership enquiries

    Now
    –
    Target
    +50%
    By
    90 days

    Steps

    1. 1

      Remove the notice

      Replace with current wording.

      Web and content manager 1 hour

    2. 2

      Write the levels

      Who each is for and what it includes.

      Membership and partnerships lead 1 day

    3. 3

      Build the form

      Size, interest, working groups.

      Web and content manager 1 day

    4. 4

      Route

      HubSpot owner and reply template.

      Membership and partnerships lead 1 day

    5. 5

      Track

      Enquiries by level.

      Membership and partnerships lead monthly

    6. 6

      Follow up

      Reply within two working days.

      Membership and partnerships lead ongoing

    First result
    2 weeks
    Cost
    $0
    Owner
    Membership and partnerships lead

    Smallest test

    Swap the notice and add the form; compare enquiries with last quarter.

    Why it works

    Replace the notice with current levels and what each includes, and add an enquiry form (organization size, level of interest, working groups wanted) that routes to the membership lead in HubSpot. Show the free path for academics and small startups first so they self-select.

    A dated notice makes prospects wonder whether the page is maintained, and a form gives the membership lead a measurable pipeline.

    Risks and how to handle them

    • Pricing questions by email continue

      Answer the top three in an FAQ beside the form.

    Evidence

    • Your websiteGet Involved: membership and affiliate status are requested by email; "We welcome academics, individuals and small startups as non-paying members"; paid members get voting rights, chair roles, "co-marketing support for MLPerf results" and logo placement; the membership section still says "membership dues and initiation fees will be increasing from June 1st, 2025".Source
    • HiringMarketing Director (2026 posting, $175-200k, remote): "For the first time, we are offering paid products through direct sales and channel partners." The role manages "both a member-driven organization and an enterprise SaaS company in one, from brand to launch to pipeline", shifting "from a primarily standards-oriented, community-centric communications model to one that incorporates B2B product marketing"; it must unlock members' "own marketing and communications" and drive "membership growth"; "we do not have a consumer-facing brand nor a strong PR practice"; "we have some budget to build a team". The post says MLCommons has "over 150 members".Source
    • ChannelsYouTube @mlcommons: 430 subscribers, 54 videos, 12,908 total views. Also a Discord server, GitHub, X and LinkedIn; HubSpot is verified on the domain and the footer has a newsletter signup.Source
  • 04

    The homepage, Get Involved page, Our Members page and Candela launch post say 125+, while the Marketing Director post says "over 150 members". A neutral measurement body is judged on precise numbers.

    Medium impactLow effort First result in 1 week Marketing Director

    First action

    Confirm today's member and affiliate counts.

    How to measure it

    Pages updated

    Now
    –
    Target
    100%
    By
    1 week

    Steps

    1. 1

      Count

      Members and affiliates separately.

      Membership and partnerships lead 1 day

    2. 2

      Define

      What the headline number includes.

      Marketing Director 1 hour

    3. 3

      Find mentions

      Search pages and posts for the number.

      Web and content manager 2 hours

    4. 4

      Update

      Every page in one pass.

      Web and content manager 1 day

    5. 5

      Centralize

      One setting the theme reads.

      Web and content manager 1 day

    6. 6

      Review

      Every quarter.

      Membership and partnerships lead quarterly

    First result
    1 week
    Cost
    $0
    Owner
    Marketing Director

    Smallest test

    Fix the homepage number first.

    Why it works

    Agree the current count and how it is defined (members, affiliates, or both), update every page in one pass, and keep the number in one place the theme reads from so it never drifts again.

    Consistent numbers support credibility with press and buyers, and a higher, correct number is free proof.

    Risks and how to handle them

    • Old posts keep old numbers

      Leave dated news posts; update evergreen pages.

    Evidence

    • Your websiteHomepage: H1 "Standardizing AI measurement and enabling informed decisions" (a year ago it read "Better AI for Everyone"); the only hero button is "Get Involved"; "125+ MLCommons Members and Affiliates", "10 Benchmark Suites", "89.7k+ MLPerf Performance Results to-date", "700k Datasets using the Croissant metadata vocabulary". Candela appears only in the Latest Insights list.Source
    • HiringMarketing Director (2026 posting, $175-200k, remote): "For the first time, we are offering paid products through direct sales and channel partners." The role manages "both a member-driven organization and an enterprise SaaS company in one, from brand to launch to pipeline", shifting "from a primarily standards-oriented, community-centric communications model to one that incorporates B2B product marketing"; it must unlock members' "own marketing and communications" and drive "membership growth"; "we do not have a consumer-facing brand nor a strong PR practice"; "we have some budget to build a team". The post says MLCommons has "over 150 members".Source
    • Your websiteCandela launch post (October 29, 2026): an organization "of over 125 companies (including Microsoft, Google, Qualcomm, NVIDIA, OpenAI, Anthropic)"; Safety 1.2.1 results "for [about 30] named open-weight models are public now, with Jailbreak results to follow"; coming benchmarks for "agentic security, privacy, and psychosocial harms" and "agent reliability profiles for industries like financial services"; scoring by a private classifier, not an LLM judge, using item response theory "the statistics behind the SAT and GRE"; "Businesses already know how to handle risk you can measure - that's the whole premise of the insurance industry."Source
    • Your websiteGet Involved: membership and affiliate status are requested by email; "We welcome academics, individuals and small startups as non-paying members"; paid members get voting rights, chair roles, "co-marketing support for MLPerf results" and logo placement; the membership section still says "membership dues and initiation fees will be increasing from June 1st, 2025".Source
  • 05

    Search engines are being asked to index form-confirmation pages, a "link-expired" page, profile-settings pages and old event RSVP pages, and one whitepaper's address contains a full email-campaign tracking string. These dilute a 50-page sitemap and can show up in search for a brand that sells trust.

    Medium impactLow effort First result in 1 month Web and content manager

    First action

    List every utility page in the page sitemap.

    How to measure it

    Indexed utility pages

    Now
    –
    Target
    0
    By
    30 days

    Steps

    1. 1

      List

      Utility and RSVP pages.

      Web and content manager 1 hour

    2. 2

      Noindex

      Yoast setting per page.

      Web and content manager 1 hour

    3. 3

      Fix the whitepaper address

      Clean slug and redirect.

      Web and content manager 1 hour

    4. 4

      Fix the template

      Campaign tags in the query string only.

      Marketing Director 1 hour

    5. 5

      Request removal

      Search Console removals for old URLs.

      Web and content manager 1 day

    6. 6

      Check monthly

      Sitemap and index report.

      Web and content manager monthly

    First result
    1 month
    Cost
    $0
    Owner
    Web and content manager

    Smallest test

    Noindex the five most obvious utility pages today.

    Why it works

    Set noindex on utility and RSVP pages and drop them from the sitemap, give the whitepaper a clean address with a redirect from the old one, and add a rule to the HubSpot email template so campaign parameters stay in the query string, not in page slugs.

    A tidy index focuses crawl and ranking on benchmark, results and Candela pages, and removes an embarrassing result from search.

    Risks and how to handle them

    • Breaking event links

      Redirect RSVP pages to the event's news post.

    Evidence

    • Your websiteSitemap (checked November 4, 2026): 171 posts and 50 pages, including form-confirmation, "link-expired", profile-settings and event RSVP pages; one whitepaper page's address contains a full email-campaign tracking string; no contact page; 12 benchmark pages and 23 working-group pages.Source
  • 06

    Real visitors pass Core Web Vitals, but in Google's phone test the hero image appears at 19.2 seconds and the homepage weighs 4.2 MB, with about 1.8 MB of image savings available. New visitors from conferences and press links on slower connections see the lab case, not the average.

    Medium impactLow effort First result in 1 week Web and content manager

    First action

    Re-export the hero image as WebP at three widths.

    How to measure it

    Mobile LCP (lab)

    Now
    –
    Target
    under 4s
    By
    30 days

    Steps

    1. 1

      Export

      Hero in WebP/AVIF at three widths.

      Web and content manager 2 hours

    2. 2

      Add srcset

      In the hero block template.

      Web and content manager 2 hours

    3. 3

      Prioritize

      High fetch priority for the hero.

      Web and content manager 1 hour

    4. 4

      Resize others

      Member logos and insight images.

      Web and content manager 1 day

    5. 5

      Retest

      PageSpeed mobile.

      Web and content manager 1 hour

    6. 6

      Watch field data

      Chrome UX Report monthly.

      Web and content manager monthly

    First result
    1 week
    Cost
    $0
    Owner
    Web and content manager

    Smallest test

    Swap the hero file and retest.

    Why it works

    Serve the hero in modern formats at phone sizes, set it to load with high priority, and resize other homepage images. Keep the design unchanged.

    Faster first screens help visitors from press and events reach the product paths, and it is a day of work.

    Risks and how to handle them

    • Image quality

      Compare side by side before release.

    Evidence

    • Google PageSpeedMobile performance 66 (desktop 76). In the phone lab test the hero image (largest content) appears at 19.2s and the homepage weighs 4.2 MB, with about 1.8 MB of image savings available. Real visitors pass Core Web Vitals (field LCP 2.3s, INP 96ms, CLS 0).Source
  • 07

    MLCommons is moving from community communications to "brand to launch to pipeline". HubSpot is already on the domain, but signups, Candela requests, membership enquiries and newsletter subscribers are not tied together, so nobody can yet say which channel produces paying buyers.

    High impactLow effort First result in 1 month Marketing Director

    First action

    Create the Candela deal stages in HubSpot.

    How to measure it

    Deals with source

    Now
    –
    Target
    90%
    By
    60 days

    Steps

    1. 1

      Define stages

      Candela and membership.

      Marketing Director 1 day

    2. 2

      UTM rules

      One sheet for every channel.

      Marketing Director 1 day

    3. 3

      Connect forms

      Candela, membership and newsletter.

      Web and content manager 2 days

    4. 4

      Build dashboard

      Requests to official results.

      Marketing Director 2 days

    5. 5

      Share monthly

      With the Executive Director and GMs.

      Marketing Director monthly

    6. 6

      Tune

      Drop channels that never convert.

      Marketing Director quarterly

    First result
    1 month
    Cost
    Staff time only
    Owner
    Marketing Director

    Smallest test

    Tag one month of Candela emails by source by hand.

    Why it works

    Set up one pipeline for Candela (request, setup, practice, annotated, official) and one for membership, with UTM rules for posts, newsletter, YouTube, Discord and partner links. Build a one-page dashboard the Executive Director sees each month.

    A new product line needs proof of traction quickly, and attribution set up now saves months of guessing.

    Risks and how to handle them

    • Too few deals to read trends

      Report counts and stories, not rates, for the first quarter.

    Evidence

    • HiringMarketing Director (2026 posting, $175-200k, remote): "For the first time, we are offering paid products through direct sales and channel partners." The role manages "both a member-driven organization and an enterprise SaaS company in one, from brand to launch to pipeline", shifting "from a primarily standards-oriented, community-centric communications model to one that incorporates B2B product marketing"; it must unlock members' "own marketing and communications" and drive "membership growth"; "we do not have a consumer-facing brand nor a strong PR practice"; "we have some budget to build a team". The post says MLCommons has "over 150 members".Source
    • ChannelsYouTube @mlcommons: 430 subscribers, 54 videos, 12,908 total views. Also a Discord server, GitHub, X and LinkedIn; HubSpot is verified on the domain and the footer has a newsletter signup.Source
    • Your websiteGetting Started with Candela: "We charge per run of the test. MLCommons members get free and discounted tests." "Automated signup and billing are coming soon. In the meantime, please email us to get started." The price table is headed "Candela pricing table preview" (also the page title): Safety practice test $200 for non-members (free for members), annotated $4,800, official $6,000; Jailbreak official $8,000; Safety + Jailbreak bundle official $13,000. Users get started via three paths: "a developer, a deployer. or a third-party assurance vendor".Source

This quarter 9

The core plays for the next three months.

  • 08

    Candela's getting-started page names three buyer paths ("a developer, a deployer. or a third-party assurance vendor") but sends them all to one email address and one generic page. Each has a different reason to pay: a score for a system card, a check of a configured deployment, or a standard to resell.

    High impactMedium effort First result in 1 quarter Marketing Director

    First action

    Write the deployer page first, from the composability post.

    How to measure it

    Candela requests from landing pages

    Now
    –
    Target
    50% of requests
    By
    1 quarter

    Steps

    1. 1

      Interview

      Two early buyers per path.

      Marketing Director 2 weeks

    2. 2

      Write

      Three pages with sample reports.

      Marketing Director 3 weeks

    3. 3

      Build

      Templates in WordPress.

      Web and content manager 1 week

    4. 4

      Link

      From Candela, results and posts.

      Web and content manager 1 day

    5. 5

      Promote

      Newsletter and LinkedIn.

      Marketing Director 2 weeks

    6. 6

      Measure

      Requests per path.

      Marketing Director monthly

    First result
    1 quarter
    Cost
    Staff time only
    Owner
    Marketing Director

    Smallest test

    Publish the deployer page and compare its requests with the generic page.

    Why it works

    Write a page for each with its job to be done, the right test tier, a sample report, and a matching form path. Deployers get the composability argument from the launch post ("safety scoring is not composable"); developers get the official-result-for-a-system-card story.

    Specific pages convert better and rank for the different searches each buyer makes.

    Risks and how to handle them

    • Sample reports reveal test content

      Use the illustrative examples already on the Candela page.

    Evidence

    • Your websiteGetting Started with Candela: "We charge per run of the test. MLCommons members get free and discounted tests." "Automated signup and billing are coming soon. In the meantime, please email us to get started." The price table is headed "Candela pricing table preview" (also the page title): Safety practice test $200 for non-members (free for members), annotated $4,800, official $6,000; Jailbreak official $8,000; Safety + Jailbreak bundle official $13,000. Users get started via three paths: "a developer, a deployer. or a third-party assurance vendor".Source
    • Your websiteCandela page: "Benchmarking-as-a-Service for AI Risk and Reliability", "Evaluations on demand for developers and enterprise deployers", "official results for marketing and decision-making"; developers "commission an official result for a score they can put in a system card or spec sheet". Its "See the assessments" button opens an unpublished page-preview address rather than a published page, and the page title begins with a stray "- ".Source
    • Your websiteCandela launch post (October 29, 2026): an organization "of over 125 companies (including Microsoft, Google, Qualcomm, NVIDIA, OpenAI, Anthropic)"; Safety 1.2.1 results "for [about 30] named open-weight models are public now, with Jailbreak results to follow"; coming benchmarks for "agentic security, privacy, and psychosocial harms" and "agent reliability profiles for industries like financial services"; scoring by a private classifier, not an LLM judge, using item response theory "the statistics behind the SAT and GRE"; "Businesses already know how to handle risk you can measure - that's the whole premise of the insurance industry."Source
  • 09

    Safety 1.2.1 results for about 30 named open-weight models are public, and Jailbreak results are coming. Teams who look up a model they run are the warmest Candela buyers, because the launch post says behavior shifts once a model gets a system prompt, retrieval or tools.

    High impactMedium effort First result in 1 quarter AI Risk & Reliability program…

    First action

    Add the button to the Safety results page.

    How to measure it

    Results-page requests

    Now
    –
    Target
    15 a month
    By
    1 quarter

    Steps

    1. 1

      Write the line

      Why deployments differ.

      AI Risk & Reliability program lead 1 day

    2. 2

      Add buttons

      Results page and tables.

      Web and content manager 2 days

    3. 3

      Per-model pages

      One page per tested model.

      MLPerf Software Engineer 1 month

    4. 4

      Structured data

      Dataset markup for results.

      Web and content manager 1 week

    5. 5

      Jailbreak next

      Same pattern when results ship.

      AI Risk & Reliability program lead 1 month

    6. 6

      Measure

      Requests by model.

      Marketing Director monthly

    First result
    1 quarter
    Cost
    Engineering time
    Owner
    AI Risk & Reliability program lead

    Smallest test

    Add one button and count requests for a month.

    Why it works

    On each model's result, add "Your deployment of this model may score differently" with a button to book a practice test for that configuration. Add per-model pages so each result can rank for "<model> safety benchmark".

    It turns free results into the top of the paid funnel at the moment of interest.

    Risks and how to handle them

    • Seen as selling against model makers

      Keep the wording neutral and about configuration, not the model.

    Evidence

    • Your websiteCandela launch post (October 29, 2026): an organization "of over 125 companies (including Microsoft, Google, Qualcomm, NVIDIA, OpenAI, Anthropic)"; Safety 1.2.1 results "for [about 30] named open-weight models are public now, with Jailbreak results to follow"; coming benchmarks for "agentic security, privacy, and psychosocial harms" and "agent reliability profiles for industries like financial services"; scoring by a private classifier, not an LLM judge, using item response theory "the statistics behind the SAT and GRE"; "Businesses already know how to handle risk you can measure - that's the whole premise of the insurance industry."Source
    • Your websiteGetting Started with Candela: "We charge per run of the test. MLCommons members get free and discounted tests." "Automated signup and billing are coming soon. In the meantime, please email us to get started." The price table is headed "Candela pricing table preview" (also the page title): Safety practice test $200 for non-members (free for members), annotated $4,800, official $6,000; Jailbreak official $8,000; Safety + Jailbreak bundle official $13,000. Users get started via three paths: "a developer, a deployer. or a third-party assurance vendor".Source
    • HiringMLPerf Senior Software Engineer (12-month contract, start "Immediately"): builds "the automation behind the benchmarks and the critical processes around submission, review, publication, and reproducibility, as well as the visualization of benchmarking results"; nice-to-have "results visualization experience".Source
  • 10

    The Marketing Director post names "channel partners" as a sales route, the getting-started page names third-party assurance vendors as users, and AIRR working groups already include Accenture, Deloitte and RAND. Firms that audit AI for clients need a standard score to cite.

    High impactMedium effort First result in 2 quarters Marketing Director

    First action

    Ask two working-group firms what they would need to resell Candela.

    How to measure it

    Partner tests

    Now
    –
    Target
    10 a quarter
    By
    2 quarters

    Steps

    1. 1

      Interview

      Three firms.

      Marketing Director 2 weeks

    2. 2

      Set terms

      Pricing and referral fee.

      Executive Director 2 weeks

    3. 3

      Train

      Run-book and certification.

      AI Risk & Reliability program lead 1 month

    4. 4

      Make the kit

      Report covers, slides, page listing.

      Marketing Director 1 month

    5. 5

      Launch

      Partner page and joint post.

      Marketing Director 2 weeks

    6. 6

      Review

      Tests and revenue per partner.

      Marketing Director quarterly

    First result
    2 quarters
    Cost
    Partner discounts
    Owner
    Marketing Director

    Smallest test

    Sign one partner and run one client test together.

    Why it works

    Create a partner tier: training on running Candela, partner pricing, co-branded report covers, a listing on a partner page and a referral code. Start with three assurance or consulting firms already in the working groups.

    Partners bring enterprise deployers that a small team cannot reach directly, and they sell the standard as part of their own work.

    Risks and how to handle them

    • Partners compete with each other

      Keep the program open on equal terms, as membership is.

    Evidence

    • HiringMarketing Director (2026 posting, $175-200k, remote): "For the first time, we are offering paid products through direct sales and channel partners." The role manages "both a member-driven organization and an enterprise SaaS company in one, from brand to launch to pipeline", shifting "from a primarily standards-oriented, community-centric communications model to one that incorporates B2B product marketing"; it must unlock members' "own marketing and communications" and drive "membership growth"; "we do not have a consumer-facing brand nor a strong PR practice"; "we have some budget to build a team". The post says MLCommons has "over 150 members".Source
    • Your websiteGetting Started with Candela: "We charge per run of the test. MLCommons members get free and discounted tests." "Automated signup and billing are coming soon. In the meantime, please email us to get started." The price table is headed "Candela pricing table preview" (also the page title): Safety practice test $200 for non-members (free for members), annotated $4,800, official $6,000; Jailbreak official $8,000; Safety + Jailbreak bundle official $13,000. Users get started via three paths: "a developer, a deployer. or a third-party assurance vendor".Source
    • Your websiteCandela launch post (October 29, 2026): an organization "of over 125 companies (including Microsoft, Google, Qualcomm, NVIDIA, OpenAI, Anthropic)"; Safety 1.2.1 results "for [about 30] named open-weight models are public now, with Jailbreak results to follow"; coming benchmarks for "agentic security, privacy, and psychosocial harms" and "agent reliability profiles for industries like financial services"; scoring by a private classifier, not an LLM judge, using item response theory "the statistics behind the SAT and GRE"; "Businesses already know how to handle risk you can measure - that's the whole premise of the insurance industry."Source
  • 11

    The MLPerf GM post links AMD and NVIDIA announcing their MLPerf results, and the Marketing Director is asked to unlock members' "own marketing and communications". Paid members are promised "co-marketing support for MLPerf results". Most member posts link to their own pages, not to MLCommons.

    High impactMedium effort First result in 1 quarter General Manager

    First action

    Build the kit for the next MLPerf release.

    How to measure it

    Member backlinks per release

    Now
    –
    Target
    20
    By
    2 releases

    Steps

    1. 1

      Gather claim rules

      From the messaging guidelines.

      General Manager, MLPerf 1 week

    2. 2

      Design assets

      Charts and cards.

      Marketing Director 2 weeks

    3. 3

      Send early

      Embargoed kit to submitters.

      General Manager, MLPerf 1 week

    4. 4

      Track posts

      Links and mentions.

      Marketing Director per release

    5. 5

      Thank members

      Feature top posts in the newsletter.

      Marketing Director per release

    6. 6

      Improve

      Ask submitters what they used.

      General Manager, MLPerf per release

    First result
    1 quarter
    Cost
    Design time
    Owner
    General Manager, MLPerf

    Smallest test

    Send a kit for one benchmark round.

    Why it works

    Before each results release, send submitters a kit: approved claim wording, a results chart with an MLCommons link, social cards, a quote line and a press-release paragraph about the benchmark. Track how many member posts link back.

    Members already publish on results day; giving them assets makes every release spread MLCommons' name and links with no extra budget.

    Risks and how to handle them

    • Claims overstate results

      Use the existing messaging guidelines in the kit.

    Evidence

    • HiringGeneral Manager, MLPerf: will "Own the P&L and overall MLPerf strategy", "drive revenue generation by cultivating the community", "champion a customer-success mindset", "Evolve the 'MLPerf benchmark factory'" for velocity and rigor, "Communicate the business impact of our benchmarks to advocate for membership value" and build "a product-focused team of product managers and relationship leads". The post links member announcements such as AMD and NVIDIA MLPerf results releases.Source
    • HiringMarketing Director (2026 posting, $175-200k, remote): "For the first time, we are offering paid products through direct sales and channel partners." The role manages "both a member-driven organization and an enterprise SaaS company in one, from brand to launch to pipeline", shifting "from a primarily standards-oriented, community-centric communications model to one that incorporates B2B product marketing"; it must unlock members' "own marketing and communications" and drive "membership growth"; "we do not have a consumer-facing brand nor a strong PR practice"; "we have some budget to build a team". The post says MLCommons has "over 150 members".Source
    • Your websiteGet Involved: membership and affiliate status are requested by email; "We welcome academics, individuals and small startups as non-paying members"; paid members get voting rights, chair roles, "co-marketing support for MLPerf results" and logo placement; the membership section still says "membership dues and initiation fees will be increasing from June 1st, 2025".Source
  • 12

    MLPerf Endpoints publishes on rolling cohorts "every few weeks", aimed at enterprise and infrastructure buyers. Without a digest, buyers have to check back to see new verified results.

    Medium impactMedium effort First result in 1 quarter General Manager

    First action

    Add an Endpoints subscription option to the HubSpot form.

    How to measure it

    Digest subscribers

    Now
    –
    Target
    2,000
    By
    2 quarters

    Steps

    1. 1

      Set up the list

      HubSpot subscription type.

      Marketing Director 1 day

    2. 2

      Template

      Three-part digest.

      Marketing Director 1 week

    3. 3

      Automate data

      Pull new results from the pipeline.

      MLPerf Software Engineer 1 month

    4. 4

      Promote

      Endpoints page and LinkedIn.

      Marketing Director ongoing

    5. 5

      Send

      Each cohort.

      General Manager, MLPerf every few weeks

    6. 6

      Measure

      Opens and clicks to results.

      Marketing Director monthly

    First result
    1 quarter
    Cost
    Staff time only
    Owner
    General Manager, MLPerf

    Smallest test

    Send one digest by hand for the next cohort.

    Why it works

    Send a short "New Endpoints results" email per cohort: what was added, what changed on the frontier, which results moved from provisional to verified. Offer it as a separate subscription on the Endpoints page.

    A regular reason to hear from MLCommons builds a buyer audience the GM can sell membership and sponsorship to.

    Risks and how to handle them

    • Too technical for buyers

      Lead with one plain-English takeaway.

    Evidence

    • Your websiteMLPerf Endpoints buyer's guide (October 2026): "Endpoints runs on rolling submissions with publication cohorts every few weeks"; results carry a status: "verified", "provisional" or "unverified"; written for "Enterprise leaders", "AI infrastructure decision-makers" and "Model decision-makers".Source
    • ChannelsYouTube @mlcommons: 430 subscribers, 54 videos, 12,908 total views. Also a Discord server, GitHub, X and LinkedIn; HubSpot is verified on the domain and the footer has a newsletter signup.Source
    • HiringGeneral Manager, MLPerf: will "Own the P&L and overall MLPerf strategy", "drive revenue generation by cultivating the community", "champion a customer-success mindset", "Evolve the 'MLPerf benchmark factory'" for velocity and rigor, "Communicate the business impact of our benchmarks to advocate for membership value" and build "a product-focused team of product managers and relationship leads". The post links member announcements such as AMD and NVIDIA MLPerf results releases.Source
  • 13

    Candela sells official results "for a score they can put in a system card or spec sheet". Each published score is also an advertisement, if it links back.

    High impactMedium effort First result in 1 quarter AI Risk & Reliability program…

    First action

    Design the badge and its usage rules.

    How to measure it

    Linked badges

    Now
    –
    Target
    25
    By
    2 quarters

    Steps

    1. 1

      Design

      Badge and results URL.

      Marketing Director 2 weeks

    2. 2

      Rules

      Version, date and configuration shown.

      AI Risk & Reliability program lead 1 week

    3. 3

      Build

      Results record per official test.

      MLPerf Software Engineer 1 month

    4. 4

      Include

      With every official result.

      AI Risk & Reliability program lead ongoing

    5. 5

      Track

      Referrer traffic from badges.

      Marketing Director monthly

    6. 6

      Enforce

      Report misuse.

      AI Risk & Reliability program lead ongoing

    First result
    1 quarter
    Cost
    Engineering time
    Owner
    AI Risk & Reliability program lead

    Smallest test

    Offer a badge to the next official result.

    Why it works

    Give each official result a verifiable badge and a results page URL, with usage rules (version, date, configuration). Ask buyers to link the badge in system cards, model pages and sales decks.

    Every badge is a link and a quiet referral to other developers, and verifiability protects the brand from misused scores.

    Risks and how to handle them

    • Old badges on new versions

      Show the benchmark version and date on the badge.

    Evidence

    • Your websiteCandela page: "Benchmarking-as-a-Service for AI Risk and Reliability", "Evaluations on demand for developers and enterprise deployers", "official results for marketing and decision-making"; developers "commission an official result for a score they can put in a system card or spec sheet". Its "See the assessments" button opens an unpublished page-preview address rather than a published page, and the page title begins with a stray "- ".Source
    • Your websiteGetting Started with Candela: "We charge per run of the test. MLCommons members get free and discounted tests." "Automated signup and billing are coming soon. In the meantime, please email us to get started." The price table is headed "Candela pricing table preview" (also the page title): Safety practice test $200 for non-members (free for members), annotated $4,800, official $6,000; Jailbreak official $8,000; Safety + Jailbreak bundle official $13,000. Users get started via three paths: "a developer, a deployer. or a third-party assurance vendor".Source
    • Your websiteCandela launch post (October 29, 2026): an organization "of over 125 companies (including Microsoft, Google, Qualcomm, NVIDIA, OpenAI, Anthropic)"; Safety 1.2.1 results "for [about 30] named open-weight models are public now, with Jailbreak results to follow"; coming benchmarks for "agentic security, privacy, and psychosocial harms" and "agent reliability profiles for industries like financial services"; scoring by a private classifier, not an LLM judge, using item response theory "the statistics behind the SAT and GRE"; "Businesses already know how to handle risk you can measure - that's the whole premise of the insurance industry."Source
  • 14

    The launch post says safety is not composable once a model gets a system prompt, retrieval or tools, and a financial-services working group is building agent reliability profiles. A published pilot gives deployers proof that testing their own system matters.

    High impactHigh effort First result in 2 quarters AI Risk & Reliability program…

    First action

    Ask the financial-services working group for five volunteers.

    How to measure it

    Pilot participants

    Now
    –
    Target
    10
    By
    2 quarters

    Steps

    1. 1

      Design

      Scope and anonymity rules.

      AI Risk & Reliability program lead 2 weeks

    2. 2

      Recruit

      Members first.

      Membership and partnerships lead 1 month

    3. 3

      Run tests

      Configured systems.

      AI Risk & Reliability program lead 2 months

    4. 4

      Analyze

      Score changes from base models.

      AI Risk & Reliability program lead 1 month

    5. 5

      Publish

      Report and webinar.

      Marketing Director 1 month

    6. 6

      Convert

      Offer official tests to participants.

      Marketing Director ongoing

    First result
    2 quarters
    Cost
    Discounts and staff time
    Owner
    AI Risk & Reliability program lead

    Smallest test

    Run two deployments and write a short case note.

    Why it works

    Recruit ten financial-services deployers (members first) for discounted annotated tests of their configured systems, and publish anonymized findings on how scores moved from the base model. Use the report to launch the industry profile.

    Regulated buyers need peer evidence before they budget, and the pilot turns a technical point into a sales story.

    Risks and how to handle them

    • Participants fear exposure

      Publish only aggregated, anonymized results.

    Evidence

    • Your websiteCandela launch post (October 29, 2026): an organization "of over 125 companies (including Microsoft, Google, Qualcomm, NVIDIA, OpenAI, Anthropic)"; Safety 1.2.1 results "for [about 30] named open-weight models are public now, with Jailbreak results to follow"; coming benchmarks for "agentic security, privacy, and psychosocial harms" and "agent reliability profiles for industries like financial services"; scoring by a private classifier, not an LLM judge, using item response theory "the statistics behind the SAT and GRE"; "Businesses already know how to handle risk you can measure - that's the whole premise of the insurance industry."Source
    • Your websiteGetting Started with Candela: "We charge per run of the test. MLCommons members get free and discounted tests." "Automated signup and billing are coming soon. In the meantime, please email us to get started." The price table is headed "Candela pricing table preview" (also the page title): Safety practice test $200 for non-members (free for members), annotated $4,800, official $6,000; Jailbreak official $8,000; Safety + Jailbreak bundle official $13,000. Users get started via three paths: "a developer, a deployer. or a third-party assurance vendor".Source
  • 15

    The YouTube channel has 54 videos and about 12,900 views in total. The job post says MLCommons needs its "perspective be sought and heard" without a strong PR practice. Short, regular video is the cheapest way to build that voice.

    Medium impactMedium effort First result in 1 quarter Marketing Director

    First action

    Book the first chair for next month.

    How to measure it

    Monthly views

    Now
    –
    Target
    5,000
    By
    2 quarters

    Steps

    1. 1

      Plan topics

      Six months of releases.

      Marketing Director 1 week

    2. 2

      Book chairs

      One per month.

      Marketing Director monthly

    3. 3

      Record

      Remote, 15 minutes.

      Marketing Director monthly

    4. 4

      Cut clips

      Three per episode.

      Marketing Director monthly

    5. 5

      Link

      Results and Candela pages.

      Web and content manager monthly

    6. 6

      Review

      Views and clicks.

      Marketing Director quarterly

    First result
    1 quarter
    Cost
    Editing time
    Owner
    Marketing Director

    Smallest test

    Record one episode and post three clips.

    Why it works

    Record a 15-minute monthly briefing with a working-group chair: what changed in a benchmark, what results show, what buyers should take from it. Cut three clips for LinkedIn and X and link each to the relevant results page.

    Regular expert voices make MLCommons the source journalists and buyers quote, and clips reach the LinkedIn audience where enterprise buyers are.

    Risks and how to handle them

    • Chairs speak for their companies

      Agree a neutral brief before recording.

    Evidence

    • ChannelsYouTube @mlcommons: 430 subscribers, 54 videos, 12,908 total views. Also a Discord server, GitHub, X and LinkedIn; HubSpot is verified on the domain and the footer has a newsletter signup.Source
    • HiringMarketing Director (2026 posting, $175-200k, remote): "For the first time, we are offering paid products through direct sales and channel partners." The role manages "both a member-driven organization and an enterprise SaaS company in one, from brand to launch to pipeline", shifting "from a primarily standards-oriented, community-centric communications model to one that incorporates B2B product marketing"; it must unlock members' "own marketing and communications" and drive "membership growth"; "we do not have a consumer-facing brand nor a strong PR practice"; "we have some budget to build a team". The post says MLCommons has "over 150 members".Source
    • Your websiteMLPerf Endpoints buyer's guide (October 2026): "Endpoints runs on rolling submissions with publication cohorts every few weeks"; results carry a status: "verified", "provisional" or "unverified"; written for "Enterprise leaders", "AI infrastructure decision-makers" and "Model decision-makers".Source
  • 16

    MLCommons was selected as the benchmarking team for an NSF-backed AI efficiency challenge to build MLPerf Agents, and it welcomes academics as non-paying members. Challenge teams are future submitters and working-group contributors.

    Medium impactMedium effort First result in 2 quarters General Manager

    First action

    Draft the welcome email for challenge teams.

    How to measure it

    Teams joining

    Now
    –
    Target
    30%
    By
    2 quarters

    Steps

    1. 1

      Welcome pack

      Account and working-group links.

      General Manager, MLPerf 2 weeks

    2. 2

      Leaderboard

      Challenge results page.

      MLPerf Software Engineer 1 month

    3. 3

      Invite

      Top teams to present.

      General Manager, MLPerf per round

    4. 4

      Feature

      Posts on standout entries.

      Marketing Director per round

    5. 5

      Track

      Joins and contributions.

      Membership and partnerships lead monthly

    6. 6

      Report

      To NSF and the board.

      General Manager, MLPerf quarterly

    First result
    2 quarters
    Cost
    Staff time only
    Owner
    General Manager, MLPerf

    Smallest test

    Invite one round's teams to an open working-group call.

    Why it works

    Give every challenge team an MLCommons account, a path to join working groups, and a public leaderboard for the challenge on mlcommons.org. Invite strong teams to present at a working-group meeting.

    The challenge brings in exactly the researchers who later build and use the benchmarks, at no acquisition cost.

    Risks and how to handle them

    • Challenge rules limit outreach

      Check the program terms before inviting teams.

    Evidence

    • Your websiteNews (October 15, 2026): "MLCommons Selected as Benchmarking Team for U.S. National Science Foundation-Backed AI Efficiency Challenge to Build MLPerf Agents".Source
    • Your websiteGet Involved: membership and affiliate status are requested by email; "We welcome academics, individuals and small startups as non-paying members"; paid members get voting rights, chair roles, "co-marketing support for MLPerf results" and logo placement; the membership section still says "membership dues and initiation fees will be increasing from June 1st, 2025".Source
    • HiringGeneral Manager, MLPerf: will "Own the P&L and overall MLPerf strategy", "drive revenue generation by cultivating the community", "champion a customer-success mindset", "Evolve the 'MLPerf benchmark factory'" for velocity and rigor, "Communicate the business impact of our benchmarks to advocate for membership value" and build "a product-focused team of product managers and relationship leads". The post links member announcements such as AMD and NVIDIA MLPerf results releases.Source

Big bets 3

Larger plays that can change the trajectory.

  • 17

    Automated signup and billing are "coming soon", and the MLPerf engineer post covers submission, publication and results visualization. Commercial testing vendors such as Giskard and Lakera let teams start on their own.

    High impactHigh effort First result in 3 quarters MLPerf Software Engineer

    First action

    Map the current manual onboarding steps.

    How to measure it

    Self-serve tests

    Now
    –
    Target
    70%
    By
    3 quarters

    Steps

    1. 1

      Map

      Manual steps today.

      MLPerf Software Engineer 2 weeks

    2. 2

      Accounts and billing

      Signup, payment, member check.

      MLPerf Software Engineer 2 months

    3. 3

      Client download

      Keys and setup in the dashboard.

      MLPerf Software Engineer 1 month

    4. 4

      Results view

      Scores and upgrade paths.

      MLPerf Software Engineer 1 month

    5. 5

      Beta

      Twenty users.

      AI Risk & Reliability program lead 1 month

    6. 6

      Launch

      With a post and webinar.

      Marketing Director 1 month

    First result
    3 quarters
    Cost
    Engineering time
    Owner
    MLPerf Software Engineer

    Smallest test

    Let five users sign themselves up with a manual billing step.

    Why it works

    Build account signup, card payment and member-discount checks, a dashboard to download the client and watch runs, and a results view with upgrade buttons from practice to annotated and official. Keep sales for volume and custom work.

    Self-serve removes the bottleneck of a small team answering emails and lets demand scale with the brand.

    Risks and how to handle them

    • Misuse of practice tests

      Rate limits and terms per account.

    Evidence

    • Your websiteGetting Started with Candela: "We charge per run of the test. MLCommons members get free and discounted tests." "Automated signup and billing are coming soon. In the meantime, please email us to get started." The price table is headed "Candela pricing table preview" (also the page title): Safety practice test $200 for non-members (free for members), annotated $4,800, official $6,000; Jailbreak official $8,000; Safety + Jailbreak bundle official $13,000. Users get started via three paths: "a developer, a deployer. or a third-party assurance vendor".Source
    • HiringMLPerf Senior Software Engineer (12-month contract, start "Immediately"): builds "the automation behind the benchmarks and the critical processes around submission, review, publication, and reproducibility, as well as the visualization of benchmarking results"; nice-to-have "results visualization experience".Source
    • Competitor researchCommercial AI testing vendors' own homepages: Giskard, "AI Red Teaming & LLM Security Platform"; Lakera, "The AI-Native Security Platform to Accelerate GenAI"; Patronus AI sells AI evaluation. They sell self-serve or sales-led testing without a consensus standard behind the score.Source
  • 18

    The launch post frames AI risk as "normal risk" and says measurable risk is "the whole premise of the insurance industry". Insurers writing AI cover need a standard, repeatable measure.

    High impactHigh effort First result in 4 quarters Executive Director

    First action

    List five insurers or brokers active in AI cover.

    How to measure it

    Insurer pilots

    Now
    –
    Target
    2
    By
    4 quarters

    Steps

    1. 1

      List

      Insurers and brokers.

      Executive Director 1 month

    2. 2

      Brief

      Method and limits of the score.

      AI Risk & Reliability program lead 1 quarter

    3. 3

      Invite

      Working-group participation.

      Membership and partnerships lead 1 quarter

    4. 4

      Pilot

      Use in underwriting.

      Executive Director 2 quarters

    5. 5

      Publish

      Joint findings.

      Marketing Director 1 month

    6. 6

      Expand

      More insurers.

      Executive Director ongoing

    First result
    4 quarters
    Cost
    Staff time and travel
    Owner
    Executive Director

    Smallest test

    Hold one briefing with one broker.

    Why it works

    Brief insurers and brokers that write AI or technology liability cover, offer them a working-group seat, and pilot Candela official results as one input to pricing or eligibility. Publish the outcome.

    If insurers ask for the score, every insured deployer becomes a buyer; that is demand no marketing budget can buy.

    Risks and how to handle them

    • Seen as endorsing products

      Keep MLCommons as the measurement provider only.

    Evidence

    • Your websiteCandela launch post (October 29, 2026): an organization "of over 125 companies (including Microsoft, Google, Qualcomm, NVIDIA, OpenAI, Anthropic)"; Safety 1.2.1 results "for [about 30] named open-weight models are public now, with Jailbreak results to follow"; coming benchmarks for "agentic security, privacy, and psychosocial harms" and "agent reliability profiles for industries like financial services"; scoring by a private classifier, not an LLM judge, using item response theory "the statistics behind the SAT and GRE"; "Businesses already know how to handle risk you can measure - that's the whole premise of the insurance industry."Source
    • Your websiteGetting Started with Candela: "We charge per run of the test. MLCommons members get free and discounted tests." "Automated signup and billing are coming soon. In the meantime, please email us to get started." The price table is headed "Candela pricing table preview" (also the page title): Safety practice test $200 for non-members (free for members), annotated $4,800, official $6,000; Jailbreak official $8,000; Safety + Jailbreak bundle official $13,000. Users get started via three paths: "a developer, a deployer. or a third-party assurance vendor".Source
  • 19

    The Endpoints guide is written for enterprise and infrastructure buyers, and Candela official results are designed for "marketing and decision-making". Procurement teams copy requirements from templates.

    High impactHigh effort First result in 3 quarters Marketing Director

    First action

    Draft the clause for one inference and one safety requirement.

    How to measure it

    Known RFPs citing MLCommons

    Now
    –
    Target
    20
    By
    4 quarters

    Steps

    1. 1

      Draft

      Inference and safety clauses.

      Marketing Director 1 month

    2. 2

      Review

      With member procurement teams.

      General Manager, MLPerf 1 month

    3. 3

      Publish

      Page and PDF.

      Web and content manager 2 weeks

    4. 4

      Promote

      Procurement groups and LinkedIn.

      Marketing Director 1 quarter

    5. 5

      Collect

      Ask buyers to report use.

      Membership and partnerships lead ongoing

    6. 6

      Update

      With each benchmark version.

      General Manager, MLPerf per version

    First result
    3 quarters
    Cost
    Staff time only
    Owner
    Marketing Director

    Smallest test

    Share the draft with three member buying teams.

    Why it works

    Publish model RFP language: request MLPerf Endpoints results at stated concurrency, and Candela official results for the configured system, with how to read each. Share it with procurement associations and working-group members' buying teams.

    Once a requirement appears in RFPs, vendors must test, and the standard becomes the default.

    Risks and how to handle them

    • Members object to being required

      Make it about evidence, not specific vendors.

    Evidence

    • Your websiteMLPerf Endpoints buyer's guide (October 2026): "Endpoints runs on rolling submissions with publication cohorts every few weeks"; results carry a status: "verified", "provisional" or "unverified"; written for "Enterprise leaders", "AI infrastructure decision-makers" and "Model decision-makers".Source
    • Your websiteCandela page: "Benchmarking-as-a-Service for AI Risk and Reliability", "Evaluations on demand for developers and enterprise deployers", "official results for marketing and decision-making"; developers "commission an official result for a score they can put in a system card or spec sheet". Its "See the assessments" button opens an unpublished page-preview address rather than a published page, and the page title begins with a stray "- ".Source
    • HiringMarketing Director (2026 posting, $175-200k, remote): "For the first time, we are offering paid products through direct sales and channel partners." The role manages "both a member-driven organization and an enterprise SaaS company in one, from brand to launch to pipeline", shifting "from a primarily standards-oriented, community-centric communications model to one that incorporates B2B product marketing"; it must unlock members' "own marketing and communications" and drive "membership growth"; "we do not have a consumer-facing brand nor a strong PR practice"; "we have some budget to build a team". The post says MLCommons has "over 150 members".Source

Wildcards 3

Unexpected experiments, each with a cheap way to test it.

  • 20

    The homepage says 700k datasets use the Croissant metadata vocabulary, and Croissant is MLCommons' most discussed project on Hacker News. Most of those dataset pages never link back to MLCommons.

    Medium impactLow effort First result in 1 quarter Croissant working group lead

    First action

    Build a validator page with the badge snippet.

    How to measure it

    Badged datasets and hosts

    Now
    –
    Target
    1,000
    By
    2 quarters

    Steps

    1. 1

      Validator

      Paste a URL, get a pass.

      MLPerf Software Engineer 1 month

    2. 2

      Badge

      Snippet and image.

      Marketing Director 1 week

    3. 3

      Landing page

      With working-group link.

      Web and content manager 1 week

    4. 4

      Hosts

      Ask the largest hosts.

      Croissant working group lead 1 quarter

    5. 5

      Track

      Referrals.

      Marketing Director monthly

    6. 6

      Feature

      Notable datasets.

      Marketing Director monthly

    First result
    1 quarter
    Cost
    Engineering time
    Owner
    Croissant working group lead

    Smallest test

    Offer the badge to ten dataset authors.

    Why it works

    Offer a validator and a small "Croissant-ready" badge that links to a Croissant page with a join-the-working-group path. Ask the largest dataset hosts to show it.

    Turns broad adoption into visible attribution and community leads.

    Risks and how to handle them

    • Hosts decline badges

      Ask for a text link instead.

    Evidence

    • Your websiteHomepage: H1 "Standardizing AI measurement and enabling informed decisions" (a year ago it read "Better AI for Everyone"); the only hero button is "Get Involved"; "125+ MLCommons Members and Affiliates", "10 Benchmark Suites", "89.7k+ MLPerf Performance Results to-date", "700k Datasets using the Croissant metadata vocabulary". Candela appears only in the Latest Insights list.Source
    • Open-source footprintDocker Hub org "mlcommons": 113 images and about 121,000 pulls; 21 Hacker News mentions, top story on Croissant, the metadata format for ML-ready datasets.Source
  • 21

    MLCommons has 113 Docker images with about 121,000 pulls. The engineers pulling them are future submitters and members, but the images do not invite them in.

    Medium impactLow effort First result in 1 month MLPerf Software Engineer

    First action

    Update the five most-pulled image READMEs.

    How to measure it

    Tracked visits from tools

    Now
    –
    Target
    500 a month
    By
    1 quarter

    Steps

    1. 1

      Rank images

      By pulls.

      MLPerf Software Engineer 1 day

    2. 2

      Write the line

      Results, join, Candela.

      Marketing Director 1 day

    3. 3

      Update READMEs

      Top images.

      MLPerf Software Engineer 1 week

    4. 4

      Add to output

      Benchmark tool help text.

      MLPerf Software Engineer 2 weeks

    5. 5

      Track

      UTM links.

      Marketing Director monthly

    6. 6

      Extend

      GitHub repos.

      MLPerf Software Engineer 1 month

    First result
    1 month
    Cost
    $0
    Owner
    MLPerf Software Engineer

    Smallest test

    Update one README and watch visits.

    Why it works

    Add a short line to each image README and to tool output: where to see results, how to join a working group, and how to test an AI system with Candela, with tracked links.

    The people already using the tools are the warmest community and member prospects.

    Risks and how to handle them

    • Engineers dislike promotion in tools

      One plain line, no banners.

    Evidence

    • Open-source footprintDocker Hub org "mlcommons": 113 images and about 121,000 pulls; 21 Hacker News mentions, top story on Croissant, the metadata format for ML-ready datasets.Source
    • Your websiteGet Involved: membership and affiliate status are requested by email; "We welcome academics, individuals and small startups as non-paying members"; paid members get voting rights, chair roles, "co-marketing support for MLPerf results" and logo placement; the membership section still says "membership dues and initiation fees will be increasing from June 1st, 2025".Source
    • Your websiteGetting Started with Candela: "We charge per run of the test. MLCommons members get free and discounted tests." "Automated signup and billing are coming soon. In the meantime, please email us to get started." The price table is headed "Candela pricing table preview" (also the page title): Safety practice test $200 for non-members (free for members), annotated $4,800, official $6,000; Jailbreak official $8,000; Safety + Jailbreak bundle official $13,000. Users get started via three paths: "a developer, a deployer. or a third-party assurance vendor".Source
  • 22

    Commercial vendors such as Giskard, Lakera and Patronus AI sell AI testing, while Candela's difference is a consensus standard, closed prompts, a custom classifier and percentile scoring. Buyers searching for red-teaming tools do not see that difference.

    Medium impactLow effort First result in 1 quarter Marketing Director

    First action

    List the five questions buyers ask about the difference.

    How to measure it

    Comparison-page requests

    Now
    –
    Target
    10 a month
    By
    2 quarters

    Steps

    1. 1

      Collect questions

      From Candela requests.

      Marketing Director 2 weeks

    2. 2

      Write

      Factual page.

      Marketing Director 2 weeks

    3. 3

      Review

      Neutrality check with AIRR leads.

      AI Risk & Reliability program lead 1 week

    4. 4

      Publish

      Linked from Candela.

      Web and content manager 1 day

    5. 5

      Promote

      Search and LinkedIn.

      Marketing Director ongoing

    6. 6

      Refresh

      Twice a year.

      Marketing Director 6 months

    First result
    1 quarter
    Cost
    Staff time only
    Owner
    Marketing Director

    Smallest test

    Publish a short FAQ answer first.

    Why it works

    Publish a factual "How Candela differs" page: who sets the standard, how prompts are protected, how scoring works, and when to use a commercial tool alongside it. Avoid naming claims you cannot verify; describe categories.

    Buyers compare before they buy; explaining the neutral-standard advantage wins the ones who need a citable result.

    Risks and how to handle them

    • Members include tool vendors

      Frame tools as complementary, not rivals.

    Evidence

    • Competitor researchCommercial AI testing vendors' own homepages: Giskard, "AI Red Teaming & LLM Security Platform"; Lakera, "The AI-Native Security Platform to Accelerate GenAI"; Patronus AI sells AI evaluation. They sell self-serve or sales-led testing without a consensus standard behind the score.Source
    • Your websiteCandela launch post (October 29, 2026): an organization "of over 125 companies (including Microsoft, Google, Qualcomm, NVIDIA, OpenAI, Anthropic)"; Safety 1.2.1 results "for [about 30] named open-weight models are public now, with Jailbreak results to follow"; coming benchmarks for "agentic security, privacy, and psychosocial harms" and "agent reliability profiles for industries like financial services"; scoring by a private classifier, not an LLM judge, using item response theory "the statistics behind the SAT and GRE"; "Businesses already know how to handle risk you can measure - that's the whole premise of the insurance industry."Source
    • Your websiteCandela page: "Benchmarking-as-a-Service for AI Risk and Reliability", "Evaluations on demand for developers and enterprise deployers", "official results for marketing and decision-making"; developers "commission an official result for a score they can put in a system card or spec sheet". Its "See the assessments" button opens an unpublished page-preview address rather than a published page, and the page title begins with a stray "- ".Source

03

How Programmable can help

Where an idea in this playbook is something Programmable builds or runs, it's listed here with what we'd do. Everything else is yours to run.

ProgrammableInfo

Turn the research into a list you can act on

These ideas start with finding the right companies. Info finds them and the right person at each, researches every one and drafts the outreach, then asks before it sends.

  • Build this list in Info for idea #19 (opens programmable.info in a new tab)

New accounts: 100,000 free credits with code BETA100K

Find the companies (opens programmable.info in a new tab)
ProgrammableDigital

Show it, don't just describe it

These ideas need something to show: a demo, a video, an animation, graphics or a deck. Digital turns screenshots or the site into all of them, in the company's own brand.

Make a demo (opens programmable.digital in a new tab)
More example playbooks and reviews

More growth playbooks

Other companies we have researched, each read from what they are hiring for.

  • FEATURE growth strategyA sneaker and fashion boutique started in Las Vegas in 2010, selling limited-edition footwear, designer apparel and lifestyle goods from…21 growth ideas
  • Lead Assassin growth strategyAn outbound and GTM engineering agency that builds pipeline systems (Clay enrichment, cold email and LinkedIn sequencing, CRM and RevOps…23 growth ideas
  • IHS | Industrial Heat Sources growth strategyAn authorized distributor of industrial hot air tools, heaters, blowers and plastic welding equipment from 19 brands, the maker of its own…24 growth ideas
  • GIVA growth strategyA Bengaluru jewellery brand selling hallmarked 925 silver, lab-grown diamond (Heer by GIVA) and gold jewellery through giva.co, its app…21 growth ideas
  • BloomChic growth strategyA digital-first plus-size fashion brand selling dresses, tops, swimwear and more in sizes 10–30 through bloomchic.com and its iOS and…21 growth ideas
  • Wonderschool growth strategyA child care marketplace and operating system: parents find and enroll in care, providers start and run programs with software and…22 growth ideas
  • Glia growth strategyA banking AI platform for voice and digital service: AI agents (Glia Banker, CoPilot, Analyst, Outreach), contact center and digital…21 growth ideas
  • Spara growth strategyAn AI agent platform for go-to-market teams: agents that qualify inbound leads across chat, email, voice and text, run AI product demos…21 growth ideas

See every playbook and website review

Independent research. Programmable wrote this playbook from public sources, as they stood on November 4, 2026, and cites them throughout. MLCommons did not commission, review or endorse it, and Programmable is not affiliated with MLCommons. Company and product names are trademarks of their owners, used only to identify them. The ideas are suggestions and opinion, not professional, legal or financial advice, and not statements about how MLCommons runs its business.

Something wrong, or do you represent MLCommons? Email [email protected] and we will correct it or take it down.

30/60/90 roadmap

What to focus on in each month, and the ideas that carry it.

  1. Days 1 to 307 ideas

    Candela booking form and fixes, buyer paths on the homepage, membership form, one member number, sitemap clean-up, hero image and HubSpot pipelines.

  2. Days 31 to 6011 ideas

    Candela buyer pages, results-page buttons, partner kit, results-day kit, Endpoints digest, badges, the financial-services pilot, monthly briefings and the NSF challenge; tool READMEs and the comparison page.

  3. Days 61 to 904 ideas

    Commit to self-serve Candela, the insurer pilot, the RFP clause and the Croissant badge.

The research behind the plan

3 research areas, each summarised in a few lines. Open any of them for the full detail.

Customers

What customers say

Customers value neutral, consensus-built benchmarks. The most cited problem: no common way to measure AI risk in a configured system.

  • Comparable, reproducible results
  • Benchmarks that models can learn
  • In their words: “industry standard”

Market

Competitive landscape

3 competitors compared: Giskard, Lakera and Patronus AI.

  • Gap to exploit: A citable, neutral AI risk score for configured systems
  • Giskard: No consensus standard behind the score
  • Lakera: Vendor-defined measures

Hiring signals

MLCommons is turning a standards consortium into a business.

  • Marketing Director: First paid products via direct sales and channel partners. Ideas 1, 2, 8, 10.
  • General Manager, MLPerf: Owns the MLPerf P&L and member value. Ideas 11, 12, 16.
  • MLPerf Senior Software Engineer: Submission, publication and results visualization. Ideas 9, 17.

From the team that wrote this playbook

All products
  • ProgrammableAgencyWorking AI systems in days, not quarters (opens programmable.agency in a new tab)
  • ProgrammableInfoAn agent that finds, researches and asks firstNew accounts get 100,000 free enrichment credits with code BETA100K (opens programmable.info in a new tab)
  • ProgrammableDigitalScreenshots in, an animated product demo out (opens programmable.digital in a new tab)
  • ProgrammableWatchKnow the day a supplier falls out of compliance (opens programmable.watch in a new tab)
  • ProgrammableTalkGet registrants for your events, and know who’s coming (opens programmable.talk in a new tab)
  • ProgrammableMediaAn AI note taker, interpreter and webinar host for your calls (opens programmable.media in a new tab)
  • ProgrammablePageText what you need, your customer gets a page that works (opens programmable.page in a new tab)
  • ProgrammableCommunityForums with labelled AI members, never empty (opens programmable.community in a new tab)
ProgrammableMarketing
ExamplesLeaderboardsProductsTermsPrivacyCookiesDPALegalContact

© 2026 AI-Powered Growth Intelligence