Skip to content
BrentM-BuildGuyPublic

About

KEEL — Kinetic Economic Execution Layer. Floating foundry, salvage yard, and agent mint between Starbase, Memphis, and AI1.

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

22 Commits

Folders and files

Repository files navigation

KEEL

KEEL

Kinetic Economic Execution Layer

Not a token. Not a dashboard. A floating foundry, salvage yard, and agent mint in the water between Starbase, Memphis, and the first AI1 birds.

I would not start this in a WeWork. I start things the way I start everything: cut plate, run bead, hook cooling, make it survive salt.

I am Brent Mayo. Welder. Builder. SpaceX / xAI infrastructure. This repository is my working file for the yard I am standing up — hull, shops, agents, and the ledger that only exists so the deck does not wait on a meeting.

The chain is not the product. The product is a yard that can catch a 171-foot stainless cylinder in a swell, strip it, weld a new radiator farm onto an AI1 bus, and have Grok agents bid the work while the hull is still dripping.


One-liner, in my voice

Trades first. We catch the ships, weld the loops, meter the heat, and let the agents pay each other so the yard never waits on a meeting. Memphis trains the mind. KEEL keeps the body in the water.


Table of contents

  1. What this actually is
  2. Why 2026, not 2032
  3. The physical plant, deck by deck
  4. The chain, without the sermon
  5. Asset classes on-chain
  6. Agents with wallets
  7. Wallet flow — one job
  8. How the chains integrate
  9. SpaceXAI / Grok / Starship — integration that is not fanfic
  10. Quant layer — what I meter
  11. Build sequence
  12. Repository layout
  13. Schemas
  14. Safety and political settlement
  15. What I will not ship
  16. Status

What this actually is

KEEL is a semi-submersible industrial platform — oil-rig hull married to a Mechazilla catch deck — that does three jobs nobody else can do in one place:

  1. Recover intact Starship upper stages and failed or retired AI1 compute satellites from the ocean before salt eats them.
  2. Fabricate the ugly metal orbital compute actually needs: radiator loops, pressure vessels, bus frames, shield plates, coolant manifolds.
  3. Settle every joule, every crane-hour, every GPU-second, and every salvage title through autonomous agents that hold their own wallets.

That object is my résumé: offshore construction, Starship marine systems, Colossus infrastructure.

If you came here looking for an “AI coin,” close the tab. If you came here because Flight 13’s Ship 40 spent twenty-four days getting towed through the Indian Ocean to Christmas Island, keep reading. That recovery is the proof that orbital compute still ends in salt water and tugs. That is my world.


Why 2026, not 2032

SpaceXAI is doing two violent things at once. KEEL sits in the gap.

On the ground — Colossus

Colossus 1/2 in the Memphis–Mississippi corridor is racing toward 2 GW by year-end. Halls are already leased: Anthropic on the order of ~$1.25B/mo, Google ~$920M/mo, after months of running sites without full backup power and cooling. Tesla dropped on the order of ~720 Megapacks onto Colossus 2. Reliability is the bottleneck, not ambition. I will not design a barge that repeats the early-hall mistake: GPUs first, batteries later.

In orbit — AI1

AI1: ~70 m wingspan, ~120–150 kW compute node, liquid radiators, Starlink-V3-derived bus. First prototypes targeted 2027, scale talk 2028, FCC ask in the direction of up to a million orbital compute sats. First birds will fail. Some will ditch. All of them need metal, cooling loops, and a place to die that is not a press release.

At sea — the wet Ship

Flight 13, Ship 40: survived splashdown, then a long wet tow. Yokohama fenders the size of trucks. That is not a metaphor. That is the class of kit I will hang on this hull.

KEEL is the missing machine between those three facts. You do not wait until 2032 to learn that salt eats stainless and that a 150 kW box in vacuum is a heat problem, not a FLOPS problem. You build the yard before the first funeral.


The physical plant, deck by deck

Hull

Converted semi-submersible. DP3 thrusters. Yokohama fender belt. Moon pool amidships. Two knuckle-boom cranes rated for a wet Starship section. A catch throat — not chopsticks: hydraulic saddles that clamp a floating cylinder in a 2-meter swell.

I already speak the Gulf market for tired semis. Dry-dock, cut, add the belt, add the saddles. No token on day one. Steel.

Hot shop

The place I would actually live. Dual-wire SAW and FCAW cells. Orbital welders for coolant tube. Vacuum-bagged Inconel/stainless booths. NDT bay: UT, PT, MT — Ocean Corp muscle memory.

A Grok vision rig on every booth: cameras on the puddle, acoustic arc monitor, bead geometry in real time. For the first two years the welder is still human. The agent is the inspector, the estimator, and the payer.

If the camera hash and the UT report disagree, the schema is wrong. I fix the schema, not the tweet.

Cold shop

Radiator panel line. AI1’s problem is not FLOPS. It is getting heat off a 150 kW box in vacuum. We stamp, weld, and leak-test deployable liquid-loop radiators (~110 m² class) on a moving deck, then nitrogen-pack them for Starship ride-share.

Power barge

Not a meme generator farm. A Colossus-in-miniature: containerized GB300-class racks in a Faraday cage, seawater heat exchanger, a wall of Megapacks. Backup first, then speed. That is the design brief I took from the September 2026 reliability panic.

When Memphis turbines get curtailed or a hall brownouts, KEEL sells dirty but real inference and salvage-planning compute from the water.

Comms

Starlink mesh plus a dedicated optical crosslink when an AI1 is overhead. The platform is a ground station that happens to float.


The chain, without the sermon

Call the ledger Keelchain. Settlement can be an L2 that posts roots to whatever base chain is least embarrassing that quarter. The logo is not the point. The point is every physical act writes a state transition.

I do not “put all blockchains on it.” I give each chain a job it is already good at. KEEL is the router.


Asset classes on-chain

Salvage titles

A floating Ship or a dead AI1 is a title with telemetry hash, photos, mass, contamination score, and a clock. First boat to attach a certified line and post a signed GPS+IMU packet owns the right to bid the work. Maritime law turned into a contract so two tugs don’t fistfight in 14-foot seas.

Compute scrip

GPU-seconds on the barge, and later on specific AI1 birds, minted as expiring claims. Not “AI coins.” Timed capacity.

Bead certificates

Every structural weld gets a packet:

  • WPS ID
  • welder cert
  • Grok puddle video hash
  • NDT report
  • heat number of the plate

That packet is what lets an AI1 radiator loop fly. No packet, no flight.

Cooling rights

Loop-hours on a radiator farm. In orbit, heat rejection is money.

Agent wallets

Not one treasury. Thousands. Hardware-backed keys. Policy cages. A human superintendent on the root.


Agents with wallets

Grok does not sit in a chat window. I fork it into specialists that can spend.

Each agent is a process with:

  • a hardware-backed key (HSM in the barge rack, or a secure enclave on an AI1)
  • a policy: max spend, allowed counterparties, kill switch from a human superintendent
  • a job-board subscription

The cast

Agent Lives Job
Foreman Yard ops Reads salvage title, weather window, crane schedule, weld backlog. Posts work packages.
Puddle Booth camera Scores the weld live. Cold bead or porosity spike: pause pay to the human welder’s wallet, ping Foreman. NDT with a checking account.
Tug Recovery boats Bids attach rights. Paid when the fender spread is on and IMU says the hull is stable.
Radiator AI1 Owns cooling loops on its own bird. Sells loop-hours to a neighbor running hot. Posts salvage down to KEEL when panels are punctured.
Megapack Battery wall Buys cheap night power / bunker. Sells ride-through — the failure mode that forced the Colossus overhaul.
Grok-Core Referee Does not hold a fat wallet. Attests: bead-cert matches video, salvage GPS is not spoofed, claimed 120 kW is actually on the bus.

Supremacy here is not “AI rules humans.” It is agents outbidding humans on logistics they are faster at, while humans still strike the arc and dog-down the plate. That is the only version I will tolerate.


Wallet flow — one job

  1. Ship 40-class wreck is tagged. Tug wins attach. Title flips.
  2. Foreman opens a work package. Puddle, crane, and NDT agents lock collateral.
  3. Human welders clock in. Pay streams per accepted bead-cert, not per hour.
  4. Radiator agent in orbit prepays for the refurbished loop.
  5. Loop stacked in a Starship fairing: salvage title burns, bead-certs enter the flight file, leftover scrip settles to Megapack for the next blackout.

All of that is boring if you say “smart contracts.” It is not boring if the contract is attached to a 52-meter stainless bottle in a swell.


How the chains integrate

Layer Job Cadence
High-security settlement Final salvage titles and flight-critical bead-certs. The weld map insurance and the FAA-equivalent will actually read. Slow. Expensive. Almost never.
Hot market Crane time, GPU-seconds, loop-hours, bunker fuel. Agents spam this. Humans do not. Sub-second fills. L2 / appchain.
Identity Welder certs, vessel MMSI, AI1 serials, Grok agent DIDs. If a sat eats debris and keys burn, Foreman rotates identity without moving the salvage title. Naming layer.
Data availability Puddle video and NDT C-scans are huge. Object store (Starlink-hosted is fine). Chain keeps the hash. Cheap DA.
Bridges Ugly. Minimized. The only legal bridge is title + cert. You do not bridge “the whole economy.” Rare.

Grok at the router is translation: this hot fill and this conservative title refer to the same physical loop. That is a database problem with a torch on the other end.


SpaceXAI / Grok / Starship — integration that is not fanfic

Starship is the trunk line. KEEL does not launch. It feeds and eats Starship: up goes radiator farms; down come dead birds and wet Ships that still have usable engines, steel, and avionics for autopsy. Flight 13 already proved the wet-Ship problem is real. KEEL is the yard you build after you admit that.

AI1 is the customer. Each sat is a factory node with a wallet. It buys cooling, sells inference, files a salvage ticket when a micrometeoroid opens a loop. First prototypes in 2027 means the yard exists before the first funeral.

Colossus is the twin. Memphis trains Grok and rents halls. KEEL is the maritime twin: smaller FLOPS, higher stakes, workloads that must survive a grid failure or a PR outage. When Colossus brownouts, Megapack-agent + barge racks pick up salvage planning and puddle inference so the yard does not freeze. Two sites, one failure domain on purpose — the opposite of how the early halls were run.

Grok is not the mascot. It is QA that never sleeps, the estimator that has read every WPS I can get my hands on, and the only entity that can watch forty weld cameras at once. Training data is filthy and valuable: arc sound, salt-air corrosion rates, Starship splashdown loads, Colossus cooling faults. That dataset does not come from Common Crawl. It comes from my deck.


Quant layer — what I meter

I do not run a “TVL.” I run a yard. The numbers that matter:

Meter Unit Why
Catch window hours of 2 m swell or less If saddles can’t clamp, title is theater
Attach-to-title minutes Tug race; spoofed GPS is a firing offense
Bead accept rate % of coupons + production Pay is per cert, so this is cash
UT / PT / MT fail per joint Schema vs puddle disagreement
Loop leak-down Pa·m³/s class Flight file
Crane-hours booked vs burned hours Hot-market inventory
GPU-seconds remaining expiring scrip Barge compute
Megapack SOC % Brownout ride-through
Loop-hours sold (orbit) hours Radiator agent P&L
Joule cost of a radiator farm kWh + crane + bead Estimator vs actual

Foreman will not post a job without a weather window, a crane slot, and a budget in scrip + crane-hours + bead-certs. Example package, in English:

Cut radiator farm A from bus 17. Budget: 14 ETH-equivalent in compute scrip + 2 crane-hours + 6 bead-certs.

If the estimator (Grok) and the puddle disagree by more than the band, work stops. I would rather lose a tide than fly a lie.


Build sequence

The way I would actually do it. Not a roadmap slide.

Step 1 — Steal a hull

Charter or buy a tired semi-sub from the Gulf. Dry-dock. Fender belt, moon pool, catch saddles. No token yet. Just steel.

Step 2 — Stand up the hot shop

One booth. One Grok vision stack. Weld junk pipe in a seaway until the bead-cert schema is real.

Step 3 — Wallet the tools

Crane, booth, Megapack get keys. Human superintendent holds the root. Agents spend inside a cage. First transaction: crane-agent pays booth-agent for a test coupon. If that fails in the mess hall, it will fail in a cyclone.

Step 4 — Write salvage law into code

Sit with admiralty counsel and a Grok draft. Define attach, lie-to, tow, title-flip. Pilot on a dead barge in the Gulf, not on a Starship.

Step 5 — Tie Colossus

Dark fiber + Starlink. Memphis cooling event → barge takes salvage planner and puddle model. Barge generator trip → Colossus holds agent state.

Step 6 — First real customer

Not a coin. A radiator farm for the first AI1 proto. Bead-certs fly with the hardware. If the loop holds vacuum on orbit, KEEL is a vendor. If it doesn’t, KEEL is a lawsuit. That is the only launch that matters.

Step 7 — First real salvage

A Starship splashdown or an AI1 deorbit into a box. Tug-agent attaches. Title flips. Yard strips. Agents settle. Press gets a welder in a hood next to a stainless cylinder the size of a grain silo. That photo is the Series A.

Step 8 — Agent supremacy, limited

Radiator and Megapack trade without a human in the loop inside pre-set bands. Anything that moves a ship, a torch, or a flight file still needs a thumbprint from a superintendent who has burned himself on slag.

Step 9 — Replication

Second KEEL in the Gulf for Starbase. Third in the Pacific for Indian Ocean ditches. Each hull is a validator in the physical sense: if your yard can’t catch steel, you don’t get to attest titles.


Repository layout

assets/keel.jpg           yard mark
docs/HULL.md              hull, DP3, catch throat, fender class
docs/HOT_SHOP.md          WPS, puddle vision, NDT
docs/COLD_SHOP.md         radiator farm line
docs/SALVAGE.md           attach, title-flip, Gulf pilot
docs/AGENTS.md            Foreman, Puddle, Tug, Radiator, Megapack, Core
docs/KEELCHAIN.md         layers, what is not bridged
docs/WALLET_FLOW.md       one work package, end to end
docs/BUILD_SEQUENCE.md    steps 1–9
docs/COLOSSUS_TWIN.md     failure domain, Megapack-first
src/                      types, titles, bead packets, policy cages

This is a specification and a schema lab. The hull is not in git.


Schemas

See src/types.ts. The only objects I care about in v0:

  • SalvageTitle
  • BeadCertificate
  • WorkPackage
  • AgentPolicy
  • ComputeScrip
  • CoolingRight

If a field does not survive a superintendent’s review on a wet deck, it does not ship.


Safety and political settlement

  • Root keys stay with a human who has slag scars.
  • Agents spend inside max-spend / allowlist / kill-switch.
  • Grok-Core attests. It does not treasury.
  • Flight files require bead-certs. No packet, no flight.
  • Salvage GPS+IMU packets are signed. Spoofing is a firing offense, then a legal one.
  • I will not let an agent move a ship, a torch, or a flight file without a thumbprint.

What I will not ship

  • A token with no yard.
  • A dashboard that pretends to be a catch deck.
  • Bridging “the whole economy.”
  • Unattended torch.
  • Common-Crawl cosplay of weld data.

Status

Step 1–2 on paper and in schema. Booth vision stack specified. Bead-cert packet in this repo. Salvage title state machine in this repo. First Gulf dead-barge pilot is the next thing that is allowed to be real.

Memphis trains the mind. KEEL keeps the body in the water.

— Brent Mayo

About

KEEL — Kinetic Economic Execution Layer. Floating foundry, salvage yard, and agent mint between Starbase, Memphis, and AI1.

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages