The CARLOS story, and the hard questions

Working document for the funding conversations, 2026-09-14. Drafted by Claude on James's instruction, from the platform repo, carlosframework.com, 11factor.org, elevenmessenger.com, and the two bake-off result sets. Every number below has a source named beside it. Things I could not verify are marked check.

1. The story in five beats

Get this order right and the objections mostly answer themselves.

1. Writing software just became cheap. Two people built Eleven, a messenger with native Mac and iPhone apps and end-to-end encryption from the first message, in about two weeks. The CARLOS platform itself is ~1,570 commits since 1 August, six weeks ago.
2. Running software did not. Keeping an app up for other people, keeping their data safe, paying for it, upgrading it, moving it, and handing it on when you lose interest: none of that got cheaper. This is where community software has always died, long before LLMs. LLMs make it worse by producing far more apps that need running.
3. The fix is a paved road for operating, not another tool for writing. Rails proved twenty years ago that strong conventions turn a fast tool into reliable software. CARLOS applies the same move one layer down: one binary, one bucket, one DNS zone is a whole deployment. One process and one SQLite file per group. Instances sleep when idle and wake in milliseconds. Every app declares what its host can see.
4. It exists and we run our own businesses on it. Fifteen apps in the family, nine live, including Tito (a real ticketing business, moved from a Rails monolith), Eleven, Keymail, Woodstar, and amadan (the git forge and CI that gates the platform's own code). Jelly, the shared inbox James runs, is rebuilt on it and about to deploy. Carloku, the hosted offering, runs on CARLOS. The console is a CARLOS app.
5. We measure the claims instead of asserting them, and publish what we find. A controlled bake-off built one spec on six stacks with blind adversarial audits. We are asking for money to turn that into a tested, published method that any community can pick up, and to find out where it breaks.

One sentence: Vibe-coding made writing software cheap. CARLOS is the part that makes owning and running it cheap, safe, and boring.

2. "How is this better than people vibe-coding their own apps right now?"

This is the most important objection because it is half right, and the answer has to concede the half that is right.

Concede: authoring is solved, and we are not competing with it. Our own bake-off says so. Across every stack tried, Rails, Next.js, plain Go, Rastrillo and two CARLOS chassis arms, the frontier model produced a correct multi-user app with no cross-tenant leak and no role escalation in any arm, in both runs. The "vibe-coded apps leak everything" fear did not materialise on a clear spec. If someone wants a throwaway tool for themselves on their laptop, vibe-coding is fine and we have nothing to add.

Then reframe: the question is where the vibe-coded app lands. Today the answer is Vercel plus Supabase plus Firebase, which means rented, per-seat, centralised, and the data lives with a vendor who can read it. The app "works" but nobody owns it, it costs a subscription per project, and it dies when the author's card expires. CARLOS is the other landing place: a single binary and a SQLite file that you can copy to any Linux box, or have hosted for pennies, and that the host provably cannot read if the app says so.

Then name what the platform makes structural that vibe-coding leaves to vigilance. These are the actual differentiators, each backed by something real:

• Isolation is the platform's job, not the app's. One instance per team or family or group, one process, one database file. The single most dangerous class of generated-app bug, cross-tenant leakage, is removed from the app entirely because there is no other tenant in the process. Run 2 of the bake-off was rewritten around exactly this.
• Trust is declared, not assumed. The six trust classes (S, A1, A2, B, C) force every app to say what its host can see. Eleven and Keymail are A2: the servers hold only ciphertext, so an operator copying a database file cannot expose a message. Vibe-coded apps have no vocabulary for this at all.
• Cost is structural. Hibernation means an idle app costs nothing and wakes in milliseconds. The platform page claims 30× cheaper for the typical app that is used a few hours a week. A community with two hundred small apps cannot pay a PaaS twenty dollars a month per app; it can pay pennies.
• The safe path is the short path. The bake-off's strongest result: handed a documented CARLOS chassis, a fresh model one-shot the second app at the lowest token cost in the field with the only perfect security score (run 1), and in run 2 the CARLOS-native arms took the top three places with two perfect security scores. The auditor's line is the one to quote: "the chassis buys the mechanism, not the discipline." We are building the mechanism so the discipline has less to cover.
• Boring outlives the vibe. Single-binary Go, SQLite, systemd, no containers, no pipeline. Factor X of the Eleven Factors: humans come first, the code stays readable and reworkable by a person with no AI. A vibe-coded Next.js app is unmaintainable the day its author stops prompting.

The line to land: Vibe-coding is the demand side. Every vibe-coded app is an app that someone now has to run. We are the supply side for running them in a way that respects the people using them.

Honest gap to admit if pushed: the authoring evidence is n=1 per arm, one model, one app. The deployment and cost claims rest on our own fleet, not third parties. Both are things the funding fixes.

3. "Why should I invest, what will my return be?"

Decide who is being asked before the meeting. A financial investor and an impact funder want different returns, and hedging between them is the commonest way a pitch like this dies. Both answers exist in the material. Pick one as primary.

If the ask is commercial

The return comes from the cost structure, and there are three routes to it that already exist in some form:

• Hosting margin (Carloku). "CARLOS run for you, hosted here, never locked here." Hibernation lets one fleet host orders of magnitude more idle apps than a container platform can. Eleven already prices at €11 per member per year for business and €0.11 per member per year for communities plus 11% of subscriptions. That pricing only works because the marginal instance costs almost nothing; it is the moat expressed as a price list.
• Migration of existing products (Tito, Jelly). Tito on CARLOS is described as "10× cheaper, 10× better," and Jelly is the second real business making the move, as a rebuild rather than a lift. Every SaaS with a Rails or Django monolith and a hosting bill is a customer for one of those two patterns, and the Rails-behind-the-edge proof of concept in the specs is the productised form of the first.
• The chassis and skill as a builder product. The measured result is that a good chassis plus a tight skill doc lets a frontier model one-shot a safe app. That is a sellable artefact to anyone building on the platform, and to teams who want the method without the hosting.

State the milestones the money buys, with dates. Suggested shape: replicate the bake-off at n greater than one across models and publish; onboard a named number of paying Carloku accounts; move one external customer's monolith; publish the chassis. Investors fund outcomes; "exploring" is not one.

If the ask is for a foundation, grant or impact fund

The return is public goods with measurement attached:

• Open, licensed artefacts. Eleven and amadan are MPL-2.0. The Eleven Factors are published. The bake-off protocol, spec, rubric and results are all written so that a stranger can re-run them. Factor VI: built by humanity, owned by humanity.
• Measured independence. The deliverable is a number of communities running their own software at a stated cost per month with a stated incident record, not a promise.
• A method others can adopt without us. Factor VII, self-hosting is a right: the point is that the funded work does not create a dependency on the two of us or on Carloku.

Eleven's about page already says the aim is a cooperative ownership model. If that is the honest destination, say it early; it is a feature for this audience and a red flag for the other one.

The risk list, stated before they find it

• Model priors drift. Run 1's Next.js arm cost 2.5× the cheapest arm because the model was a version behind. Our own bespoke framework paid a similar tax. Mitigation: lean on libraries the model already knows; the bake-off is the instrument that detects drift.
• Bus factor. The git history is honest: one person wrote roughly 98% of the platform's commits. That is the strongest argument for the ask, not against it. Turning a two-person proof into a team and a method is the job.
• Scale shape. SQLite-per-instance means an app that needs one huge shared database is not a fit. The answer is "many small things," and Eleven's planet-scale design is the worked example, but be plain that this is a design commitment, not a limitation we will engineer away.
• External security review. Trust classes make claims checkable but nobody outside has checked them yet. Budget for an audit and say so.

4. "Why are you the right people to investigate this?"

Three arguments, in order of strength.

You have run real software for real money for a long time. Paul has built and operated Tito, an event ticketing business with paying customers, for over a decade check dates. James has twenty years of Rails experience, wrote the Rails Engines plugin and co-founded Go Free Range check the last two, and runs Jelly, a shared team inbox with paying customers. Between you that is twenty-plus years each of building and, more importantly, operating web software. The problem we are solving is an operations problem. Most people pitching AI-built software have never carried a pager.

You have already demonstrated the method at speed, on yourselves. Eleven from nothing to native apps with end-to-end encryption in two weeks. A deployment platform with fifteen apps in six weeks, nine of them live, one of them the CI that gates the platform itself. Tito, an existing business, moved onto it, and Jelly, James's own product, is being rebuilt on it and is about to deploy. Both founders are putting the businesses that pay them onto the platform. This is not a deck; the funder can log in.

You are unusually honest about what you know, and that is the thing an "exploration" funder is actually buying. Every result document carries an "honest limits" section. The bake-off used blind adversarial auditors, and the results that embarrassed our own framework were published anyway. The platform's README says "three honest limits, none of them papered over" about its own domain feature. AI-written text is always marked. Someone funding an investigation needs to trust the negative results as much as the positive ones, and the paper trail shows you will report them.

One more, for the Rails lineage: James built on and around Rails through the last time building web software got ten times cheaper, and that wave's lesson was that conventions and a paved road are what turn a fast tool into reliable software. The bake-off just showed the same thing for LLMs. That is not a coincidence you are guessing at; it is a pattern you have already lived through once.

The weakness to pre-empt: "two people, no company, one of them wrote almost all the code." Answer it directly. That is exactly what the money converts.

5. Other objections you will get, with short answers

• "Isn't this just another PaaS, Heroku with extra steps?" No control plane server. The bucket is the control plane, and every box converges from it. Self-hosting is one sentence: one bucket, some Linux boxes with one binary, DNS. Heroku cannot be self-hosted at all.
• "SQLite doesn't scale." One instance is one group; scale is more instances on more boxes, not a bigger database. Eleven's design doc is literally called planet-scale. Concede the case it does not fit.
• "You built a bespoke Go framework nobody's model has seen. That's the trap you named yourselves." Yes, and we measured it: Rastrillo paid the tax in run 1. We pivoted to a thin layer over libraries the model already knows, and the measurement flipped. Show the table.
• "AI-written code is slop." The platform ran a blind audit that found zero cross-tenant defects in any arm. The code is single-binary Go with the tests in the repo. Judge it by reading it.
• "Won't the hyperscalers or the model labs do this?" They will do the hosted, host-sighted version. They will not build the host-blind, self-hostable, hibernate-to-your-own-bucket version, because it is the opposite of their business.
• "Communities don't want to run servers." They do not have to. Carloku is hosted and never locked. Self-hosting is the right that keeps the hosted price honest, not a requirement.
• "What's the moat if it's all open?" The cost structure, the operating knowledge encoded in the platform, and the family of apps already on it. Code is copyable; an operated fleet with a decade of ops judgement is not.
• "Who audits your security claims?" Nobody external yet. That is a line item in the ask. What exists: trust classes that make each claim falsifiable, E2EE apps where host compromise leaks nothing, and blind adversarial audits in the bake-off.

6. Numbers worth having to hand

Claim	Number	Source
Time to build Eleven	~2 weeks, 2 people	elevenmessenger.com/about
Platform commits since 2026-08-01	1,567	carlosframework/platform commits (private)
Apps in the CARLOS family / live	15 / 9	carlosframework.com/apps
Hibernation cost claim	30× cheaper	carlosframework.com/platform
Tito on CARLOS	"10× cheaper, 10× better"	carlosframework.com/apps
Eleven pricing	€11/member/yr business; €0.11/member/yr community + 11%	elevenmessenger.com
Bake-off run 2, top three	carrillo 49, rastrillo 49, plain-go 48 of 50, all CARLOS-native	RESULTS-run2.md (private)
Bake-off run 2, token band	160K to 216K across all five arms	RESULTS-run2.md (private)
Cross-tenant leaks found, both runs	0 in any arm	RESULTS.md, RESULTS-run2.md (private)
Run 1 cost spread	244K (carrillo) to ~610K (Next.js)	RESULTS.md (private)
The Eleven Factors	factors VI, VII, IX, X cited above	11factor.org
Trust classes	S, A1, A2, B, C	carlosframework.com/trust

The two GitHub repositories are private at the time of writing, so those links work only for people who have been granted access. Make the bake-off repo public, or share a rendered copy of the results, before sending this to anyone outside.