The provable-AI ladder

Own the AI you're renting.

A private model, with a receipt that could fail, that anyone can re-run.

Everyone can install a local model — that lane is crowded, and ownership is table stakes. The moat is the proof: a mutation-gated receipt that proves the test could fail, re-run by a party who doesn't hold your key. This is the strategy that turns that one defensible asset into revenue — sequenced, priced, and mapped to what's already live.

The first move → See the ladder The live mint ↗
Sell the proof, not the code MIT-clean · charge for friction, liability & minting First buyer: lean AI-native SMEs Parallel: £1m–£3m UK gov track
The wedge

One position nobody else occupies

Two lanes are already crowded — and neither is this. Local-inference tooling (Ollama, NIM, vLLM) runs a model but has no proof. AI-governance platforms (Credo, Holistic, Fiddler, IBM) do cloud paperwork but never touch the actual inference. Sovereign local execution plus a verifiable per-inference receipt is unoccupied.

Not this

Owned / local models

Crowded lane. Ownership is table stakes — it doesn't win the deal on its own.

Not this

Signed receipts alone

A software-key signature is Level 0 — the holder can fabricate it. Never argue it as the moat.

The moat

Proof that could fail + re-run

The witness gate proves the test could fail; a CI run re-executes it by someone who doesn't hold your key. "TEEs prove where it ran. We prove the test could fail — and anyone can re-run it."

Do this now — the sharpest first-revenue move

Sell one done-for-you mint. Get one invoice paid.

The product is already live and gated — the only work left is sales. Seventeen hundred repos and £0 because the number that never moved is humans who paid, not things built.

£2,500 – £6,000 / model  ·  cash in weeks
  • Qualify with one question: "What task do you feed an LLM most, and how many times a month?"
  • Deliver a private specialist model on their task, proven to beat the base model on their own held-out data, with a signed scorecard and a one-file installer.
  • The guarantee only you can make: "If the receipt doesn't say BEATS, you don't pay." Safe — the gate is objective and returns LOSES.
  • Funnel: fall-os (the game) → fallforgemint (mint · prove · own, live) → the calculator that honestly says rent-wins below their threshold. The tool that tells them not to buy is what makes them trust the one that says buy.
  • Cash path: flip fallforge-store live (Stripe + host) for self-serve checkout, or just invoice the first three by hand. Invoice #1 doesn't need Stripe.
The ladder

Four tiers — each one sells the proof

MIT-clean throughout: never charge for the code. Charge for the deployment friction, the enterprise liability, and the custom model minting.

Tier 1 · the revenue engine

Mint

Self-serve funnel + paid done-for-you

Paste a task and examples, walk out owning a private specialist. Free to try, paid to have it done and weight-tuned.

Self-serve DIYFree
Done-for-you · per model£2.5k–£6k
Multi-domain pack (3–5)£9k–£20k
Keep-it-sharp add-on£300–£900/mo
Live now: fallforgemint (witness 192/193) · fallforge-store (Stripe+host to go live) · mint pitch page.
Tier 2 · high-ticket pilot

Foundry Build

"Proof-installed" sovereign stack

Not "we install Ollama." A private, mutation-gated, receipt-emitting AI stack on their metal — every organ proves it works and could fail, and their people can re-run the proof.

Foundry Lite · software, their hardware£6k–£12k
Foundry Standard · turnkey box + stack£15k–£30k
Foundry Rack · department / multi-user£35k–£75k+
Assets: fallbrain · minted models · agent-proof ledger · witness · the "how a company runs on FallForge" explainer.
Tier 3 · the real business

Sovereign Care

Proof as a subscription · the MRR

Red Hat model: never charge for the MIT code — charge to keep it alive, proven and accountable. Every month, a signed receipt that their AI still works and hasn't drifted.

Care Basic · monthly drift receipt£750–£1.5k/mo
Care Standard · + governance pack£2k–£4k/mo
Care Sovereign · + dedicated ops agent£5k–£9k/mo
Assets: witness (re-run gate) · agent-proof (ledger) · sididy mesh (ops, free on local) · kar-foldcycle (run-cheap: 95% prefill / 67% embed, measured).
Tier 4 · the platform · later

Kard Marketplace

Boot keys · curated organs · rev-share

Tamper-evident boot keys (fallkard), curated skins, a marketplace of pre-minted proven organs. The Series-A-shaped platform — it needs a backend, accounts and the money rail.

Boot key / kardPriced later
MarketplaceRev-share
Sequence: ships after Tiers 1–3 have paying customers. Opening it first is how the last 1,700 repos happened. Park it, don't kill it.
The finance director's yes

The foundry hardware is a Year-1 write-off

The box and setup are qualifying plant & machinery. Under UK Full Expensing, a company deducts 100% in year one on new main-rate plant & machinery — and that explicitly includes IT equipment and software, uncapped, now permanent. (The £1m Annual Investment Allowance does the same for companies and sole traders; a 40% first-year allowance covers anything outside full expensing from Jan 2026.) Depreciated over five years it's a slow cost; written off this year it's an easy yes.

Foundry Standard£20,000
Year-1 relief @ 25% CT−£5,000
~£15,000
effective net cost, written off in year one
The big parallel track
£1m – £3m

UK Sovereign AI procurement — via agent-proof

agent-proof is NCSC Challenge 3, "Enabling safe AI agent adoption," almost word-for-word — MIT, no-LLM-judge, tamper-evident audit ledger, witness-clean, CI-green. This is procurement, not a grant: the government is the customer, you keep all the IP (they get a public-sector licence), contracts £250k–£10m.

The move nowFile the free EOI / Approved Supplier List on sovereignai.gov.uk — it unlocks the Dec 2026 and Feb 2027 batches. EOI ~2 weeks before → ~17 Nov 2026.
Who runs itCounsel leads, you sign — that's how gov procurement works. The product is already built; the gap is a named delivery partner + one specific gov use-case.
Why it's cleanYou keep all IP, no equity, no dilution. Biggest single number on the table, nearest hard deadline. Don't build for it — file the EOI.
The sequence

What to do, in order

One paid mint invoice — now. Drive the qualifying question, close done-for-you on one SME, "no BEATS no pay." £2.5k–£6k, cash in weeks.
File the Sovereign AI EOI — now, in parallel. Free, counsel-led. Unlocks the £1m–£3m pipeline (~17 Nov).
Land the first Foundry pilot — weeks. Lite or Standard, £6k–£30k, off the first mint customer or a warm intro. Lead with the write-off.
Attach Care to every foundry. £750–£1.5k/mo recurring — the MRR flywheel.
The marketplace — later. Once foundries are paying.
LinePriceCash timing
Done-for-you mint£2,500–£6,000 / modelweeks
Multi-domain pack£9,000–£20,000weeks
Foundry Lite / Standard / Rack£6k / £15k–£30k / £35k–£75k+1–2 months
Sovereign Care£750–£9,000 / morecurring
UK Sovereign AI contract£1m–£3m6–12 months

Five foundries on Care ≈ £10k–£20k/mo recurring within a couple of quarters, on top of mint one-offs — while the £1m–£3m gov track runs in parallel on a product that's already done.

The free unlock

Land witness in one outside team's CI

Get one external FAIL in a repo you don't own. Free, and it converts "private garden" → "product with a user." It's the proof the entire ladder rests on — do it before Care and the audit trail go out the door.

Positioning

Pitch the threshold, not a multiple

The own-vs-rent calculator gives this buyer's real number and honestly returns rent-wins below ~500k–2M tokens/day. That honesty is the conversion asset.

Positioning

Sit in the seam

Local-inference tooling has no proof; AI-governance has no runtime receipt. Sovereign local + a verifiable per-inference receipt is the unoccupied ground.

Positioning

Sell friction, not code

The code stays MIT. Charge for the deployment friction, the enterprise liability, and the minting. Konomi architecture credited on every surface.

Growth rails

Four channels — all selling the same one thing

Affiliate, white-label, "NFT" and crypto — reframed through what the estate already owns. The estate is web3's substance without its baggage: an ownable capability wallet, a signed royalty cascade, transferable credentials, Merkle anchoring, a p2p mesh — sovereign, MIT, and provable.

White-label · scale channel

Every agency becomes a reseller

Skin £500–£1.5k/mo + rev-share · OEM £2.5k–£6k/mo

The estate is already re-skinnable per brand — a kard carries a skinRef, konomi-skins is the theming layer. An agency drops their brand on a working proof-and-mint stack and resells it (the GoHighLevel / Vendasta model, applied to provable AI). One good white-label partner is a sales team you don't pay a salary.

Native asset: kard skinRef + konomi-design-kit — the white-label kernel already exists.

Affiliate · reach channel

A signed royalty cascade, not an honour system

30% lifetime recurring on Care · flat bounty per mint

Conventional program on FirstPromoter / PartnerStack now — recurring-on-the-retainer is what makes it sticky. The viral artifact is the signed scorecard: a proof that can say you LOST is more shareable than a brag; put a referral handle in its metadata.

Native asset: the-wallet's lineageSplit — downstream value mints upstream credit, signed and capped so provenance appreciates but payout can't run away. Affiliate enforced by the kernel, no platform in the middle.

"NFT" · ownable proof

Ownable proof, not collectibles

The Kard Marketplace (ladder Tier 4)

Skip NFT-as-JPEG entirely. The real primitive — a signed, ownable, verifiable credential — is what the estate already mints: a kard is an ownable capability token, a scorecard is a verifiable credential, and provenance bundles carry real conformance levels. Anchor a hash on-chain for public provenance when wanted — never the content.

Native assets: kard + scorecard + kcc-mint (1,410 live) + kar-anchor (Merkle). Resale royalties flow via lineageSplit — the one legitimate NFT-royalty mechanic, capped.

Crypto · rail, not a coin

Payment + anchor + mesh

Stablecoin in · hashes anchored · p2p distribution

Accept stablecoin alongside Stripe (global, instant, on-thesis for the sovereign buyer). Anchor proof hashes for timestamped provenance. Let the mesh (p2p, identity-as-address, kards as bearer credentials) be the distribution. No speculative token — a coin launch turns "provable AI" into "another crypto project" and invites regulators before the first enterprise reference; keep the internal credit closed ("no money out").

Native assets: the mesh (r7 + webrtc + fallroom), kcc (closed credit), kar-anchor.

The one insight that ties it together

FallForge is web3's substance without web3's baggage. The ownable / verifiable / royalty-bearing / decentralized primitives that NFTs and tokens promised and mostly failed to deliver — here they're sovereign, MIT, and provable. White-label is the scale channel (do it first), affiliate is the reach channel (with the signed royalty cascade underneath), "NFT" means ownable proof, and crypto is a payment-and-distribution rail — every one of them selling the same thing the ladder sells: proof you own and anyone can re-run.