Theorem growth playbook

Build a /proofs/ ledger on theorem.dev from the public repository. Give each library a card showing what was proved (for example, the entry point never crashes on any input up to 64 MiB, under ASan and UBSan builds), how many entry points are covered, the Rocq version that checked it, and a link to the proof files. Add a homepage strip with the headline figures (32 libraries proved, 123 in the research collaboration) and a few recognisable names, linking to the ledger. End the ledger with two calls to action: 'Want this for your code? Talk to us' for companies and 'Enroll your project' for maintainers. List only what the repository already publishes, and nothing a project's publish setting holds back. Then each new proof becomes a page, a post and a sales email.

In a category where buyers are told to trust claims about AI-generated proofs, a public list of proved, widely used libraries is the proof itself. It turns the strongest asset Theorem has into something a security lead can forward to an auditor, and it gives search engines and AI assistants named pages for queries like 'formal verification openssl'. Newer rivals have no equivalent to show.

[ ] Library name and one-line description
[ ] What is proved, in plain English
[ ] Entry points covered (n of m)
[ ] Sanitizer builds (ASan / UBSan)
[ ] Input limit (64 MiB) and model caveats
[ ] Checked with Rocq version
[ ] Link to proof files
[ ] Date last re-proved
[ ] CTA: Talk to us / Enroll your project

• Readers overread what a proof covers

  Link every card to 'What a Proof Covers' and state the input limit and the model's limits.

• Maintainers object to being featured

  Show only projects whose publish setting allows it and offer an opt-out.

• Your GitHuboss-verify-proofs (checked November 9, 2026): machine-checked Rocq proofs about OSS-Fuzz targets compiled as RISC-V binaries; 90 entry points in 32 libraries proved, including openssl, zlib, zstd, mbedtls, libpcap, simdjson, bzip2, varnish and boost-json; of 20,421 theorem statements, 526 are proved and 19,600 open; 2,148 of 5,290 fuzz targets have statements; 204 stars, the most of any Theorem repository; MIT licensed.Source
• Your websiteHomepage (checked November 9, 2026): H1 "Verify all software." over "Verification is the only defense that holds while adversaries scale."; the hero "Talk to us →" button links to /product/, while the header "Talk to us" goes to the form at /talktous/; the page names no open-source library, gives no count of verified projects and does not link to the proofs repository; no customer logos, testimonials, case studies or newsletter signup.Source
• Your websiteProduct page: "Testing vs. verification." compares input coverage, rewrites, optimizations and code review; "What you get" lists "Clear requirements", "Proof you can check" (validated by a proof assistant "independently of the AI that produced it"), "Bugs uncovered during verification" and "Evidence you can share" for "customers, auditors, and procurement teams". No pricing, supported languages, customer logos, numbers or form; the only action is "Talk to us".Source
• News coverageResearch post with Anthropic (October 22, 2026): proofs across 123 open-source libraries and 239 entry points, about 180 million lines of Rocq, a typical proof about 600,000 lines, and "hundreds of bugs" found while proving; Theorem supplied the RISC-V and Linux process semantics and "is also offering to run the same pipeline for other open-source projects". theorem.dev shows it only as an outbound link on the homepage and blog.Source

Send the hero 'Talk to us' straight to /talktous/ and add a secondary 'See how it works' link to /product/. In the sandbox post, link to /talktous/ with the project box prefilled with the topic (?topic=sandbox). Add a short form, or a clear button to the form, at the bottom of /product/.

A button labelled 'Talk to us' that opens a comparison page breaks the visitor's expectation, and a sales-led company has no other conversion step. Fixing it costs minutes.

• More unqualified contacts

  Route by role and company size (idea 4).

• Your websiteHomepage (checked November 9, 2026): H1 "Verify all software." over "Verification is the only defense that holds while adversaries scale."; the hero "Talk to us →" button links to /product/, while the header "Talk to us" goes to the form at /talktous/; the page names no open-source library, gives no count of verified projects and does not link to the proofs repository; no customer logos, testimonials, case studies or newsletter signup.Source
• Your website"Proving file integrity for an xv6 sandbox" (November 3, 2026): a proof, over compiled RISC-V code on eight cores and including crashes and power loss, that a guest program cannot change host files, for "any program, including deliberately malicious code"; a research prototype; teams working on preventing agent escapes are invited to get in touch via /product/, a page with no form.Source
• Your websiteProduct page: "Testing vs. verification." compares input coverage, rewrites, optimizations and code review; "What you get" lists "Clear requirements", "Proof you can check" (validated by a proof assistant "independently of the AI that produced it"), "Bugs uncovered during verification" and "Evidence you can share" for "customers, auditors, and procurement teams". No pricing, supported languages, customer logos, numbers or form; the only action is "Talk to us".Source

Add a two-line box at the end of every post: 'Want this property proved for your code? Tell us what it must never do' (to /talktous/ with the topic prefilled), plus 'Maintain an open-source library? Enroll it free' (to /oss-verify/). Keep the hiring line beneath it.

Readers who finish a deep technical post are the highest-intent visitors a research-led company gets. One line converts some of them without diluting the research tone.

• Feels salesy to a research audience

  Keep it to one factual line with no hype.

• Your websiteResearch posts: "lf-lean" (February 5, 2026) verified all 1,276 statements of Logical Foundations, 97% autonomously, with an estimated "350× speedup on verification effort alone"; it and "Catching bugs with fractional proofs" (October 7, 2025) both end with "we're hiring" and a careers link, not a product link. On Hacker News the fractional-proofs post drew 80 points and 20 comments and lf-lean 47 points. The blog shows no newsletter signup or visible RSS link, though a feed exists at /blog/index.xml.Source
• Your website"Bootstrapping the Verified Software Stack" (September 29, 2026): a nixpkgs snapshot of 123,471 binaries in 28,650 packages; 9 MB of shared libraries is the full dependency set for about two-thirds of nixpkgs machine code; OpenSSL's proof would cost about $220K, "about $14 per package that reuses it"; verification speed on their eval doubling about every 1.9 months; it asks software builders to say which end-to-end properties they want proved.Source

Add three lines above the button: 'We reply within one business day. A first call covers the property you need proved and a sample specification. No cost.' After submission, show a calendar link for companies with 50+ staff, and send academics and maintainers to the right resources (OSS Verify, the proofs repository). Replace the open text box prompt with 'What must your code never do?'

Telling people what comes next lowers the hesitation to submit, and routing by role puts the scarce founder time on the contacts most likely to buy.

• Promised reply time slips

  Promise one business day only if a named person covers the inbox.

• Your websiteContact page "Get in touch": name, work email, company, role (founder, executive or engineering leader; security professional; platform or infrastructure engineer; academic or researcher; government or public-sector professional; other), a free-text project box and company size (under 50 to over 5,000); open-source maintainers are pointed to OSS Verify; the page does not say what happens after sending, and there is no calendar booking.Source
• Your websiteProduct page: "Testing vs. verification." compares input coverage, rewrites, optimizations and code review; "What you get" lists "Clear requirements", "Proof you can check" (validated by a proof assistant "independently of the AI that produced it"), "Bugs uncovered during verification" and "Evidence you can share" for "customers, auditors, and procurement teams". No pricing, supported languages, customer logos, numbers or form; the only action is "Talk to us".Source

Write a one-page case study with the customer's permission, anonymised if needed ('a networking hardware company'): the problem (huge specification, slow legacy code), what was specified, what was proved, the bugs found, the 100× throughput gain, and the 'no manual review' outcome. Put a summary card on /product/ under 'Checking a rewrite' and 'Checking an optimization', which describe exactly this.

The product page argues in the abstract. One concrete result, with numbers an engineering leader can repeat, shows the claim working in production.

• Customer declines

  Publish it fully anonymised, with figures they approve.

• News coverageVentureBeat (January 27, 2026): $6M seed led by Khosla Ventures with Y Combinator, e14, SAIF and Halcyon; four employees at the time; customers include AI research labs, electronic design automation firms and GPU-accelerated computing companies; one customer with a 1,500-page specification and a legacy implementation got a parser that raised throughput from 10 Mbps to 1 Gbps, with 16,000 generated lines deployed without manual review; plans to expand into robotics, renewable energy, cryptocurrency and drug synthesis.Source
• Your websiteSitemap (14 URLs, checked November 9, 2026): /teams/ (last modified June 7, 2024), /research/ and /waitlist/ are one-line "Redirecting to" stubs (to Careers, the blog and the contact form) but are listed as pages; four research posts plus the blog index; /oss-verify/ and its terms; no use-case, case-study, pricing or results pages.Source
• Your websiteProduct page: "Testing vs. verification." compares input coverage, rewrites, optimizations and code review; "What you get" lists "Clear requirements", "Proof you can check" (validated by a proof assistant "independently of the AI that produced it"), "Bugs uncovered during verification" and "Evidence you can share" for "customers, auditors, and procurement teams". No pricing, supported languages, customer logos, numbers or form; the only action is "Talk to us".Source

Add a one-field 'Get new proofs and research by email' form to the blog, the ledger (idea 1) and OSS Verify, plus a visible RSS link. Send one short email per release: what was proved, why it matters and one call to action.

Formal verification sales cycles are long. A list of engineers and security leads who opted in keeps Theorem in front of them until they have a project.

• Low send frequency

  Include proof-ledger updates between posts.

• Your websiteResearch posts: "lf-lean" (February 5, 2026) verified all 1,276 statements of Logical Foundations, 97% autonomously, with an estimated "350× speedup on verification effort alone"; it and "Catching bugs with fractional proofs" (October 7, 2025) both end with "we're hiring" and a careers link, not a product link. On Hacker News the fractional-proofs post drew 80 points and 20 comments and lf-lean 47 points. The blog shows no newsletter signup or visible RSS link, though a feed exists at /blog/index.xml.Source
• Your websiteSocial links: the site links only to X (@theoremlabs) and GitHub; no LinkedIn, YouTube or newsletter is linked from any page checked.Source

Replace the three stubs with real 301 redirects at Cloudflare and remove them from the sitemap. Fix the low-contrast text flagged by PageSpeed. Don't spend time on speed, which already passes for real visitors.

On a 14-page site, three thin pages are a fifth of what search engines see. Cleaning them up concentrates signals on the pages that sell.

• Old inbound links break

  301s keep them working.

• Your websiteSitemap (14 URLs, checked November 9, 2026): /teams/ (last modified June 7, 2024), /research/ and /waitlist/ are one-line "Redirecting to" stubs (to Careers, the blog and the contact form) but are listed as pages; four research posts plus the blog index; /oss-verify/ and its terms; no use-case, case-study, pricing or results pages.Source
• PageSpeedGoogle PageSpeed (November 10, 2026): mobile performance 99, desktop 99; real visitors pass Core Web Vitals (field LCP 650 ms, INP 66 ms, CLS 0); 143 KB page weight; one accessibility issue, insufficient colour contrast; analytics is Plausible.Source

Create or claim a LinkedIn company page, link it in the footer, and post each release in buyer language: what was proved, for which library, and what it means for companies that ship it. Ask the team to reshare. Use X for the research audience and LinkedIn for buyers.

Theorem's results are newsworthy to security leaders, but they currently reach mostly researchers. One more channel puts the same content in front of buyers at no extra cost.

• Another channel to maintain

  Repost only releases; no daily posting.

• Your websiteSocial links: the site links only to X (@theoremlabs) and GitHub; no LinkedIn, YouTube or newsletter is linked from any page checked.Source
• Your websiteContact page "Get in touch": name, work email, company, role (founder, executive or engineering leader; security professional; platform or infrastructure engineer; academic or researcher; government or public-sector professional; other), a free-text project box and company size (under 50 to over 5,000); open-source maintainers are pointed to OSS Verify; the page does not say what happens after sending, and there is no calendar booking.Source
• HiringCareers page (checked November 9, 2026): Theorem calls itself "a programming languages and AI lab". Open roles: OSS Ecosystem Lead ($180K–370K; "Work with open source maintainers, foundations, and security teams to coordinate responsible disclosure at unprecedented scale"), Systems Engineer ($200K–500K), ML Researcher ($200K–500K), Recruiter ($150K–220K) and a 3–6 month Research Fellowship ($125K/yr, prorated). No locations or full descriptions; every role applies through one shared form.Source

Publish /careers/<role>/ pages with location (or remote policy), what the person will do in the first 90 days, requirements, salary range and a link to a relevant post (for example, the OSS Verify page for the Ecosystem Lead). Add JobPosting structured data so the roles appear in job search, and give each role its own form or tagged form link.

Specific role pages attract better-matched candidates and appear in job search results. Hiring speed directly limits how many verification projects Theorem can take on.

• Revealing strategy

  Describe outcomes, not internal plans.

• HiringCareers page (checked November 9, 2026): Theorem calls itself "a programming languages and AI lab". Open roles: OSS Ecosystem Lead ($180K–370K; "Work with open source maintainers, foundations, and security teams to coordinate responsible disclosure at unprecedented scale"), Systems Engineer ($200K–500K), ML Researcher ($200K–500K), Recruiter ($150K–220K) and a 3–6 month Research Fellowship ($125K/yr, prorated). No locations or full descriptions; every role applies through one shared form.Source

Publish /specs/ with six to eight short sample specifications in plain English, each with the formal statement beneath: never crashes on any input, output equals the reference implementation, a rewrite matches the old behaviour, a sandbox never alters host files, and similar. Add a 'Works on compiled binaries, so any language that compiles to supported targets' line to /product/.

Buyers can't ask for a property they can't imagine. Concrete examples let a security lead match one to their own system and come to the first call ready.

• Specs look too academic

  Lead with the plain-English guarantee and put the formal statement second.

• Your websiteProduct page: "Testing vs. verification." compares input coverage, rewrites, optimizations and code review; "What you get" lists "Clear requirements", "Proof you can check" (validated by a proof assistant "independently of the AI that produced it"), "Bugs uncovered during verification" and "Evidence you can share" for "customers, auditors, and procurement teams". No pricing, supported languages, customer logos, numbers or form; the only action is "Talk to us".Source
• Your websiteOSS Verify: "Free formal verification for open source." Rocq proofs about the compiled binary, such as "This entry point never crashes, on any input", for inputs up to 64 MiB; eligibility follows OSS-Fuzz lines (established, widely used, OSI license, on a security boundary); maintainers enroll by pull request with project.yaml, a Dockerfile and build.sh; static riscv64 builds are required; bugs are reported privately; the FAQ says findings and patches are AI-generated and a human review before sending "is not promised"; secondary call to action "Not open source? Talk to us".Source

Offer maintainers an opt-in badge ('Proved: no crash on any input · Theorem') that links to the project's ledger card (idea 1), showing the proof date and coverage. Generate it automatically when a proof is published, and open a polite pull request to add it only where the maintainer has agreed.

Badges are how open source signals quality: CI, coverage and the OpenSSF scorecard all spread this way. Each badge sends readers, including engineers at companies that ship the library, to theorem.dev, and it gives other maintainers a reason to enroll.

• Maintainers see it as advertising

  Opt-in only; the badge describes the guarantee, not the company.

• #1 Put the 32 libraries Theorem has already proved (OpenSSL, zlib, zstd, mbedtls and more) on theorem.dev as a public proof ledger

• Your websiteOSS Verify: "Free formal verification for open source." Rocq proofs about the compiled binary, such as "This entry point never crashes, on any input", for inputs up to 64 MiB; eligibility follows OSS-Fuzz lines (established, widely used, OSI license, on a security boundary); maintainers enroll by pull request with project.yaml, a Dockerfile and build.sh; static riscv64 builds are required; bugs are reported privately; the FAQ says findings and patches are AI-generated and a human review before sending "is not promised"; secondary call to action "Not open source? Talk to us".Source
• Your GitHuboss-verify-proofs (checked November 9, 2026): machine-checked Rocq proofs about OSS-Fuzz targets compiled as RISC-V binaries; 90 entry points in 32 libraries proved, including openssl, zlib, zstd, mbedtls, libpcap, simdjson, bzip2, varnish and boost-json; of 20,421 theorem statements, 526 are proved and 19,600 open; 2,148 of 5,290 fuzz targets have statements; 204 stars, the most of any Theorem repository; MIT licensed.Source

Rank seeded and candidate projects by how many packages depend on them (from the nixpkgs analysis already done), then have the OSS Ecosystem Lead write personally to core maintainers. Each note explains what Theorem has already stated about their code, what a proof would give them, and that a coding agent guided by AGENTS.md can do the riscv64 port. Offer a 30-minute enrollment call.

The libraries the most software depends on also carry the most weight on the proof ledger and are the strongest pitch to companies. A personal invitation from the hired ecosystem lead converts far better than a README.

• Unwanted outreach to busy maintainers

  One personal note, one follow-up, and an easy no.

• Your GitHuboss-verify: 106 stars, 18 forks, 219 commits and 26 pull requests; enrollment by fork, templates and validation tools, with an AGENTS.md for coding agents; the README mentions "seeded projects" a maintainer can adopt; proofs are published when a project's publish setting allows. The theorem-labs organisation has 144 followers and 34 repositories.Source
• Your website"Bootstrapping the Verified Software Stack" (September 29, 2026): a nixpkgs snapshot of 123,471 binaries in 28,650 packages; 9 MB of shared libraries is the full dependency set for about two-thirds of nixpkgs machine code; OpenSSL's proof would cost about $220K, "about $14 per package that reuses it"; verification speed on their eval doubling about every 1.9 months; it asks software builders to say which end-to-end properties they want proved.Source
• HiringCareers page (checked November 9, 2026): Theorem calls itself "a programming languages and AI lab". Open roles: OSS Ecosystem Lead ($180K–370K; "Work with open source maintainers, foundations, and security teams to coordinate responsible disclosure at unprecedented scale"), Systems Engineer ($200K–500K), ML Researcher ($200K–500K), Recruiter ($150K–220K) and a 3–6 month Research Fellowship ($125K/yr, prorated). No locations or full descriptions; every role applies through one shared form.Source

Build /check/: paste a lockfile, SBOM or package list and get a report showing which dependencies are proved (and what was proved), which are stated but open, and which are not covered. Processing happens in the browser, so nothing is uploaded. The report ends with 'Prove the rest' (to /talktous/ with the list attached on request) and 'Are you a maintainer? Enroll'.

It turns a public research asset into a personal, shareable result, the same pattern as free security scanners. Security teams send it to colleagues, and every report is a qualified lead that has already shown what it depends on.

• Reports look thin while coverage is small

  Show what is stated and in progress, and ask visitors which dependency they want proved next.

• Package names map wrongly

  Mark uncertain matches.

• #1 Put the 32 libraries Theorem has already proved (OpenSSL, zlib, zstd, mbedtls and more) on theorem.dev as a public proof ledger

• Your GitHuboss-verify-proofs (checked November 9, 2026): machine-checked Rocq proofs about OSS-Fuzz targets compiled as RISC-V binaries; 90 entry points in 32 libraries proved, including openssl, zlib, zstd, mbedtls, libpcap, simdjson, bzip2, varnish and boost-json; of 20,421 theorem statements, 526 are proved and 19,600 open; 2,148 of 5,290 fuzz targets have statements; 204 stars, the most of any Theorem repository; MIT licensed.Source
• Your website"Bootstrapping the Verified Software Stack" (September 29, 2026): a nixpkgs snapshot of 123,471 binaries in 28,650 packages; 9 MB of shared libraries is the full dependency set for about two-thirds of nixpkgs machine code; OpenSSL's proof would cost about $220K, "about $14 per package that reuses it"; verification speed on their eval doubling about every 1.9 months; it asks software builders to say which end-to-end properties they want proved.Source
• Your websiteContact page "Get in touch": name, work email, company, role (founder, executive or engineering leader; security professional; platform or infrastructure engineer; academic or researcher; government or public-sector professional; other), a free-text project box and company size (under 50 to over 5,000); open-source maintainers are pointed to OSS Verify; the page does not say what happens after sending, and there is no calendar booking.Source

Design a standard evidence pack: a two-page summary (system, specification in plain English, what is and isn't covered, trusted base, assumptions), a proof certificate with hashes of the binary and the proof, and a one-command check anyone can rerun. Publish a sample pack for a public library from the ledger, and offer the pack as a named deliverable in every engagement.

Formal verification gets bought when it satisfies someone other than the engineers. A pack procurement teams can file shortens the buyer's internal sale, and nothing similar appears on rivals' homepages.

• Auditors want a recognised standard

  Map the pack to the controls buyers already report against, after checking them with buyers.

• Your websiteProduct page: "Testing vs. verification." compares input coverage, rewrites, optimizations and code review; "What you get" lists "Clear requirements", "Proof you can check" (validated by a proof assistant "independently of the AI that produced it"), "Bugs uncovered during verification" and "Evidence you can share" for "customers, auditors, and procurement teams". No pricing, supported languages, customer logos, numbers or form; the only action is "Talk to us".Source
• Your websiteContact page "Get in touch": name, work email, company, role (founder, executive or engineering leader; security professional; platform or infrastructure engineer; academic or researcher; government or public-sector professional; other), a free-text project box and company size (under 50 to over 5,000); open-source maintainers are pointed to OSS Verify; the page does not say what happens after sending, and there is no calendar booking.Source
• Competitor researchRivals' own homepages: Imandra, "AI reasoning where correctness matters most", sells CodeLogician with published plans at $25, $129 and $799 a month and lists clients such as DARPA and Goldman Sachs; &Truth, "Provable Trust for the AI Era", shows no product, benchmarks or results and collects sign-ups for demos; Axiom, "The starting point for reasoning", publishes mathematics formalizations.See the research Source

Write /use-cases/agent-sandboxes/: the threat (agents running untrusted code), what Theorem proved in the research prototype, what a production engagement would prove (file integrity, then isolation properties), the honest limits (hardware semantics, confidentiality not yet covered), and a 'Tell us about your sandbox' form. Link it from the post, the homepage and the ledger.

Agent sandboxes are a fast-growing, security-critical category with a crisp property to prove. A dedicated page catches that search intent and qualifies the conversation before the first call.

• Overselling a research prototype

  State plainly what is proved today and what a project would add.

• #2 Point the homepage's main button and the sandbox post at the contact form, not at a page with no form

• Your website"Proving file integrity for an xv6 sandbox" (November 3, 2026): a proof, over compiled RISC-V code on eight cores and including crashes and power loss, that a guest program cannot change host files, for "any program, including deliberately malicious code"; a research prototype; teams working on preventing agent escapes are invited to get in touch via /product/, a page with no form.Source
• Your websiteSitemap (14 URLs, checked November 9, 2026): /teams/ (last modified June 7, 2024), /research/ and /waitlist/ are one-line "Redirecting to" stubs (to Careers, the blog and the contact form) but are listed as pages; four research posts plus the blog index; /oss-verify/ and its terms; no use-case, case-study, pricing or results pages.Source

Define a 'Proof Pilot': one component up to a stated binary size, one or two properties chosen from the spec gallery (idea 10), delivered in a set number of weeks with the evidence pack (idea 14) and a bug report, at a published 'from' price. Put it on /product/ with the scope, the deliverables and what happens after.

A bounded, priced first step turns a research conversation into a purchase decision a single engineering leader can approve, and it filters out unqualified contacts.

• Underpricing complex code

  Cap binary size and properties; scope review before signing.

• #10 A "what can be proved" gallery of sample specifications, so buyers can picture their own
• #14 An auditor-ready evidence pack, so security and public-sector buyers can take the proof to procurement

• Competitor researchRivals' own homepages: Imandra, "AI reasoning where correctness matters most", sells CodeLogician with published plans at $25, $129 and $799 a month and lists clients such as DARPA and Goldman Sachs; &Truth, "Provable Trust for the AI Era", shows no product, benchmarks or results and collects sign-ups for demos; Axiom, "The starting point for reasoning", publishes mathematics formalizations.See the research Source
• Your website"Bootstrapping the Verified Software Stack" (September 29, 2026): a nixpkgs snapshot of 123,471 binaries in 28,650 packages; 9 MB of shared libraries is the full dependency set for about two-thirds of nixpkgs machine code; OpenSSL's proof would cost about $220K, "about $14 per package that reuses it"; verification speed on their eval doubling about every 1.9 months; it asks software builders to say which end-to-end properties they want proved.Source
• Your websiteProduct page: "Testing vs. verification." compares input coverage, rewrites, optimizations and code review; "What you get" lists "Clear requirements", "Proof you can check" (validated by a proof assistant "independently of the AI that produced it"), "Bugs uncovered during verification" and "Evidence you can share" for "customers, auditors, and procurement teams". No pricing, supported languages, customer logos, numbers or form; the only action is "Talk to us".Source

Publish /compare/ as a factual table: what is proved (binary vs. source model), who checks the proof (the proof-assistant kernel), the languages covered, public results, and the first engagement on offer. Include testing and fuzzing as the status-quo column. Date the page and link every claim about rivals to their own site.

New categories are compared before they are bought. Theorem has the public results to win a fair comparison, and the page also tells AI assistants how to describe the category.

• Claims about rivals go stale

  Date the page and cite rivals' own pages.

• Competitor researchRivals' own homepages: Imandra, "AI reasoning where correctness matters most", sells CodeLogician with published plans at $25, $129 and $799 a month and lists clients such as DARPA and Goldman Sachs; &Truth, "Provable Trust for the AI Era", shows no product, benchmarks or results and collects sign-ups for demos; Axiom, "The starting point for reasoning", publishes mathematics formalizations.See the research Source
• Your websiteOSS Verify: "Free formal verification for open source." Rocq proofs about the compiled binary, such as "This entry point never crashes, on any input", for inputs up to 64 MiB; eligibility follows OSS-Fuzz lines (established, widely used, OSI license, on a security boundary); maintainers enroll by pull request with project.yaml, a Dockerfile and build.sh; static riscv64 builds are required; bugs are reported privately; the FAQ says findings and patches are AI-generated and a human review before sending "is not promised"; secondary call to action "Not open source? Talk to us".Source
• Your GitHuboss-verify-proofs (checked November 9, 2026): machine-checked Rocq proofs about OSS-Fuzz targets compiled as RISC-V binaries; 90 entry points in 32 libraries proved, including openssl, zlib, zstd, mbedtls, libpcap, simdjson, bzip2, varnish and boost-json; of 20,421 theorem statements, 526 are proved and 19,600 open; 2,148 of 5,290 fuzz targets have statements; 204 stars, the most of any Theorem repository; MIT licensed.Source

Publish /oss-verify/disclosure/: how findings are triaged, the promise that a person checks every report before it is sent, how dependencies are notified, and aggregate statistics only (reports sent, confirmed and fixed, median time to fix), with no details of unpatched issues. Change the FAQ line from 'not promised' to the new commitment once the Ecosystem Lead is in place.

Maintainers enroll in programmes they trust not to bury them in noise. A human check and transparent statistics turn the bug-finding volume into a reason to join rather than a reason to avoid.

• Human checks slow throughput

  Check the first reports per project in full, then sample.

• Your websiteOSS Verify: "Free formal verification for open source." Rocq proofs about the compiled binary, such as "This entry point never crashes, on any input", for inputs up to 64 MiB; eligibility follows OSS-Fuzz lines (established, widely used, OSI license, on a security boundary); maintainers enroll by pull request with project.yaml, a Dockerfile and build.sh; static riscv64 builds are required; bugs are reported privately; the FAQ says findings and patches are AI-generated and a human review before sending "is not promised"; secondary call to action "Not open source? Talk to us".Source
• HiringCareers page (checked November 9, 2026): Theorem calls itself "a programming languages and AI lab". Open roles: OSS Ecosystem Lead ($180K–370K; "Work with open source maintainers, foundations, and security teams to coordinate responsible disclosure at unprecedented scale"), Systems Engineer ($200K–500K), ML Researcher ($200K–500K), Recruiter ($150K–220K) and a 3–6 month Research Fellowship ($125K/yr, prorated). No locations or full descriptions; every role applies through one shared form.Source
• News coverageResearch post with Anthropic (October 22, 2026): proofs across 123 open-source libraries and 239 entry points, about 180 million lines of Rocq, a typical proof about 600,000 lines, and "hundreds of bugs" found while proving; Theorem supplied the RISC-V and Linux process semantics and "is also offering to run the same pipeline for other open-source projects". theorem.dev shows it only as an outbound link on the homepage and blog.Source

Present the plan where those communities meet: a talk or discussion thread for nixpkgs maintainers, and briefings for open-source security foundations and distribution security teams. Propose a shared, public priority list of libraries to prove next, chosen with them, and invite them to co-sign or co-fund it.

Foundations and distributions are trusted by the maintainers Theorem needs to enroll. A roadmap they help shape becomes their project too, and it creates natural funding partners (see idea 21).

• Slow foundation processes

  Start with community discussion, which needs no formal approval.

• Your website"Bootstrapping the Verified Software Stack" (September 29, 2026): a nixpkgs snapshot of 123,471 binaries in 28,650 packages; 9 MB of shared libraries is the full dependency set for about two-thirds of nixpkgs machine code; OpenSSL's proof would cost about $220K, "about $14 per package that reuses it"; verification speed on their eval doubling about every 1.9 months; it asks software builders to say which end-to-end properties they want proved.Source
• Your websiteResearch posts: "lf-lean" (February 5, 2026) verified all 1,276 statements of Logical Foundations, 97% autonomously, with an estimated "350× speedup on verification effort alone"; it and "Catching bugs with fractional proofs" (October 7, 2025) both end with "we're hiring" and a careers link, not a product link. On Hacker News the fractional-proofs post drew 80 points and 20 comments and lf-lean 47 points. The blog shows no newsletter signup or visible RSS link, though a feed exists at /blog/index.xml.Source
• HiringCareers page (checked November 9, 2026): Theorem calls itself "a programming languages and AI lab". Open roles: OSS Ecosystem Lead ($180K–370K; "Work with open source maintainers, foundations, and security teams to coordinate responsible disclosure at unprecedented scale"), Systems Engineer ($200K–500K), ML Researcher ($200K–500K), Recruiter ($150K–220K) and a 3–6 month Research Fellowship ($125K/yr, prorated). No locations or full descriptions; every role applies through one shared form.Source

Offer companies a sponsorship: fund the proof of a library (or a basket of libraries) they ship. They get their name on the ledger card, early access to the evidence pack for that library, and private briefings on fixes. Maintainers still get everything free. Use the dependency checker (idea 13) to show each prospect which of their dependencies are unproved, and price from the published cost-per-KB.

It turns the free programme into revenue without charging maintainers. Companies already fund open-source security, and here they get a concrete, checkable deliverable. Each sponsored proof also grows the ledger.

• Maintainers see conflicts of interest

  Publish the rule that sponsors get no say over findings, disclosure or patches.

• Free-rider companies

  Pool sponsors per library to lower each one's share.

• #13 A free dependency checker: paste a lockfile or SBOM and see which of your dependencies are already proved

• Your website"Bootstrapping the Verified Software Stack" (September 29, 2026): a nixpkgs snapshot of 123,471 binaries in 28,650 packages; 9 MB of shared libraries is the full dependency set for about two-thirds of nixpkgs machine code; OpenSSL's proof would cost about $220K, "about $14 per package that reuses it"; verification speed on their eval doubling about every 1.9 months; it asks software builders to say which end-to-end properties they want proved.Source
• Your websiteOSS Verify: "Free formal verification for open source." Rocq proofs about the compiled binary, such as "This entry point never crashes, on any input", for inputs up to 64 MiB; eligibility follows OSS-Fuzz lines (established, widely used, OSI license, on a security boundary); maintainers enroll by pull request with project.yaml, a Dockerfile and build.sh; static riscv64 builds are required; bugs are reported privately; the FAQ says findings and patches are AI-generated and a human review before sending "is not promised"; secondary call to action "Not open source? Talk to us".Source
• Your GitHuboss-verify-proofs (checked November 9, 2026): machine-checked Rocq proofs about OSS-Fuzz targets compiled as RISC-V binaries; 90 entry points in 32 libraries proved, including openssl, zlib, zstd, mbedtls, libpcap, simdjson, bzip2, varnish and boost-json; of 20,421 theorem statements, 526 are proved and 19,600 open; 2,148 of 5,290 fuzz targets have statements; 204 stars, the most of any Theorem repository; MIT licensed.Source

Package a named offer: bring legacy or slow code and its specification, and get back a faster or modernised implementation with a proof that it matches the requirements. That covers language migrations, performance rewrites and replacing a hand-written parser. Lead with the parser case study, price per component, and include the evidence pack. Build a landing page and a short technical webinar.

Rewrites and migrations already have budgets and deadlines, and their biggest risk is behaviour changing. A proof removes that risk, making this an easier first purchase than abstract assurance.

• Engagements grow beyond scope

  Fixed components and a scope review at the start.

• #5 Publish the 10 Mbps to 1 Gbps parser as an anonymised case study on the product page
• #16 A fixed-scope, fixed-price proof pilot: one component, one property, a set number of weeks

• Your websiteProduct page: "Testing vs. verification." compares input coverage, rewrites, optimizations and code review; "What you get" lists "Clear requirements", "Proof you can check" (validated by a proof assistant "independently of the AI that produced it"), "Bugs uncovered during verification" and "Evidence you can share" for "customers, auditors, and procurement teams". No pricing, supported languages, customer logos, numbers or form; the only action is "Talk to us".Source
• News coverageVentureBeat (January 27, 2026): $6M seed led by Khosla Ventures with Y Combinator, e14, SAIF and Halcyon; four employees at the time; customers include AI research labs, electronic design automation firms and GPU-accelerated computing companies; one customer with a 1,500-page specification and a legacy implementation got a parser that raised throughput from 10 Mbps to 1 Gbps, with 16,000 generated lines deployed without manual review; plans to expand into robotics, renewable energy, cryptocurrency and drug synthesis.Source
• Your websiteResearch posts: "lf-lean" (February 5, 2026) verified all 1,276 statements of Logical Foundations, 97% autonomously, with an estimated "350× speedup on verification effort alone"; it and "Catching bugs with fractional proofs" (October 7, 2025) both end with "we're hiring" and a careers link, not a product link. On Hacker News the fractional-proofs post drew 80 points and 20 comments and lf-lean 47 points. The blog shows no newsletter signup or visible RSS link, though a feed exists at /blog/index.xml.Source

Pick the industry with the clearest property to prove and the shortest path to a binary-level proof (for example, control software for robots or energy equipment, where 'never enters an unsafe state' is the specification). Recruit one design partner on pilot terms (idea 16), publish the result with their permission, then decide whether to build a vertical page and offer.

A single published result in a new industry is worth more than four vertical pages, and it keeps engineering focused while testing the stated strategy.

• Distraction from the core market

  One industry, one partner, with a fixed review date.

• #16 A fixed-scope, fixed-price proof pilot: one component, one property, a set number of weeks

• News coverageVentureBeat (January 27, 2026): $6M seed led by Khosla Ventures with Y Combinator, e14, SAIF and Halcyon; four employees at the time; customers include AI research labs, electronic design automation firms and GPU-accelerated computing companies; one customer with a 1,500-page specification and a legacy implementation got a parser that raised throughput from 10 Mbps to 1 Gbps, with 16,000 generated lines deployed without manual review; plans to expand into robotics, renewable energy, cryptocurrency and drug synthesis.Source
• Your websiteSitemap (14 URLs, checked November 9, 2026): /teams/ (last modified June 7, 2024), /research/ and /waitlist/ are one-line "Redirecting to" stubs (to Careers, the blog and the contact form) but are listed as pages; four research posts plus the blog index; /oss-verify/ and its terms; no use-case, case-study, pricing or results pages.Source

Publish the index as a small page with a chart, updated monthly: KB verified per hour and cost per KB on a fixed benchmark, with the method stated. Add a line on what it means for buyers ('a component that cost $X to prove in May costs $Y today'). Announce each update on X, LinkedIn and the email list.

A recurring, citable number earns repeat press and social attention and positions Theorem as the measure of progress in the category. It also answers the cost objection with a trend rather than a quote.

• Progress plateaus

  Publish honestly; a plateau is also news, and the method stays credible.

• Your website"Bootstrapping the Verified Software Stack" (September 29, 2026): a nixpkgs snapshot of 123,471 binaries in 28,650 packages; 9 MB of shared libraries is the full dependency set for about two-thirds of nixpkgs machine code; OpenSSL's proof would cost about $220K, "about $14 per package that reuses it"; verification speed on their eval doubling about every 1.9 months; it asks software builders to say which end-to-end properties they want proved.Source