Three systems · One founder
Systems that show their work.
Three systems, built one way: evidence outranks opinion, models are replaceable workers, and the human steps in only by exception.
The common thread
One way of building.
Three systems for three different worlds, on the same three convictions. Each puts a layer of judgment between people and the machinery that does the work.
The human steps in only by exception.
Routine work runs itself. People are called for direction, taste, money and risk, and the question arrives already prepared.
Evidence outranks opinion.
A result counts when evidence says so. Not when its author, a confident model or a majority vote says so.
Models are replaceable workers.
Models, tools and machines keep changing. The system around them is what lasts: its state, its rules, its memory.
MAIOS
Multi-AI Operating System
State the job. MAIOS owns it.
MAIOS isn’t another AI agent. It’s the operating system that runs them.
Give it one plain-language mission. It writes the contract, assembles the right AI workers, has their work checked independently and recovers on its own when something breaks.
- Stage
- Private preview. Its first end-to-end mission was recorded in October 2026.
- People
- You, the Operator. Direction, taste, money and risk.
- System
- MAIOS. Contracts, independent verification, recovery.
- Workers
- AI workers. Qualified before they’re trusted, replaceable after.
“Reconcile our Q3 supplier invoices against the agreed price list and send me every mismatch above 2%, with the evidence.”
- Understand
- Contract
- Plan
- Execute
- Verify
- Deliver
- Builderidleworkingdone
- Writeridleworkingresumeddone
- Verifieridlecheckingdone
- Contract written · 4 acceptance checks fixed before work
- Team chosen · builder, writer, independent verifier
- Builder delivered
reconciliation.csv - Writer interrupted · resumed the same attempt · 0 duplicates
- Writer delivered
mismatch_report.pdf - Verifier checking · independent of both authors
Acceptance is decided by evidence, not by the workers.
Acceptance · C-0147
- Every invoice line parsed or flagged2,318 lines
- Each mismatch cites invoice, line and price-list row57 mismatches
- Totals recomputed independentlymatch
- Held-out sample re-verified0 / 4031 / 4040 / 40
Contract · C-0147
- Objective
- Find every Q3 invoice line priced more than 2% away from the agreed list price.
- Out of scope
- Contacting suppliers · changing the ERP
- Constraint
- Read-only on source data
- Acceptance
- 4 checks, fixed before work started
- Delivers
mismatch_report.pdf·reconciliation.csv
A strong model is a talented employee. MAIOS is the company around it.
MAIOS is model-agnostic. Every AI worker has to prove it can act, not just answer, before it gets real work. Any of them can be replaced without changing how MAIOS thinks.
ROLE≠TOOL≠MODEL
Independent acceptance
Nothing is done until the evidence says so.
Delivering a file isn’t finishing a job. MAIOS accepts work only when a verifier that didn’t build it confirms the exact result meets a bar that was set from your intent before work began. A model saying “done” counts for nothing.
Written before the work
Acceptance comes from your intent and is fixed before anything is built. The builder can’t move the bar.
Checked by someone else
A separate verifier judges the result. The author never approves its own work.
Bound to the exact result
Evidence names precisely what was checked. Missing or contradictory evidence never counts as a pass.
- DELIVERY≠ACCEPTANCE
- AUTHOR≠APPROVER
- EVIDENCE>OPINION
Recovery
Resume. Never rerun.
Processes crash and machines sleep. MAIOS continues an interrupted mission from where it stopped: the same attempt, not a fresh start. Nothing that already took effect happens twice.
- Mission under way worker acting
- Controller killed mid-turn
- Resumed same thread, same attempt
- Delivered once
- Independently verified
- Intent
- 1
- Dispatch
- 1
- Duplicates
- 0
- RESUME≠RERUNInterrupted work continues. It doesn’t start over.
- UNKNOWN≠ABSENTNot seeing an effect doesn’t mean it didn’t happen.
Operator by exception
You’re the CEO. Not the courier.
MAIOS keeps you where your judgment matters and handles the rest. When it needs you, it doesn’t ask what to do. It brings a decision, prepared.
You decide direction, taste, budget, irreversible steps and the risk you’ll accept. MAIOS handles choosing workers, retries and repairs, handoffs, testing and recovery.
The measure that matters: human dependency per verified outcome.
Two suppliers switched invoice currency mid-quarter. Which exchange rate defines a “mismatch”?
- Why now
- 9 invoice lines change status depending on the answer.
- Reversible
- Fully. The report can be regenerated.
- Cost of delay
- The report waits. All other work continues.
- If no answer
- At 15:00 MAIOS proceeds with A and flags it in the report.
One business question, fully prepared. Pick an option to see what happens next.
Proof, not promises
MAIOS built, verified and used its own first tool.
Its first mission was to build an evidence-intake audit tool, have it verified independently, then use it on MAIOS’s own source archive. These numbers come from that recorded run, October 2026.
- 507/507required independent verifier checks passed
- 1·1·0intent · dispatch · duplicates, through a real mid-turn controller kill
- 6/6source documents confirmed unchanged in actual use
- 5Operator touches: the only human involvement in the entire run
Failures are kept too. The first run failed, and a later worker binding was refused, fail-closed. Both causes were fixed and the records preserved. Evidence outranks opinion. Including ours.
Control Tower
AI logistics control tower
Silence the normal. Surface the exception.
Every shipment gets a digital twin. Only the exceptions reach a human.
An exception-first operations layer for international road freight. It sits above the telematics, planning and messaging systems a fleet already runs, keeps one record per trip with the source and confidence of every status, and brings people only what needs them.
- Stage
- Phase 1. The operator interface now runs on simulated data. Nothing is deployed and no live data provider is connected.
- People
- The operations team. Sees the exceptions, not every truck.
- System
- The tower. A digital twin of every trip, each status with its evidence.
- Workers
- Existing systems, and AI. AI reads only free text, voice and photos.
A truck’s status lives in three places at once.
A telematics feed that is sometimes stale, a plan that was right this morning, and a dispatcher’s memory of a phone call. Attention goes to the trucks that are fine. The one that isn’t gets noticed after the delivery slot, the ferry booking or the demurrage window has closed.
Guiding principle“The goal isn’t more messages. It’s the right information with the fewest human and driver touches, with its source and confidence visible.”
What it is
- A decision layer above existing systems of record
- A digital twin per trip, with source and confidence on every status
- An exception desk, not a dashboard
- Button-first driver check-ins, not a chatbot
- Arrival times you can explain line by line
What it deliberately isn’t
- A telematics provider, or a rewrite of the planning system
- A black-box model guessing arrival times
- A system that commits to customers or invoices without human approval
- A driver scoring or ranking engine
Built backend-first
Proven before a single screen depended on it.
The foundation came first and the operator interface came last, on purpose. Every work package passes independent review before it’s accepted. One took five rounds. A documentation check fails the build whenever a document claims more than the code delivers, so the project can’t quietly claim progress it hasn’t made.
Measured 21 September 2026
- 2,342unit tests
- 585integration tests against real infrastructure
- 75mutation probes, and the suite fails if a single one survives
- 38architecture decisions on record
What isn’t claimed. Nothing is deployed to production, no live data provider is connected and no pilot has started, so there are no field results here yet.
Content Factory
Local-first AI media production
Run the factory, not the tools.
AI can generate content. It still can’t run the factory by itself.
Research, writing, images, video, voice, editing, quality control and publishing usually live in separate tools, with a person as the glue. Content Factory is designed to run them as one local-first, model-agnostic system that calls its owner only when human authority is required.
- Stage
- In active development. The architecture and principles are set; the capabilities are being built and qualified.
- People
- The owner. Strategy, taste, capital and true exceptions.
- System
- The factory. State, provenance, quality gates and recovery.
- Workers
- Models, machines, media tools. Routed by capability, not brand.
One claim, followed through a storyEvidence travels with everything built on it.
- SourceInspection report, p. 14Revised report, p. 14CurrentCorrected
- Claim“Cracks grew 2 mm a year”“Cracks grew 3 mm a year”SupportedRe-checkingSupported again
- Script beatBeat 7: the slow failureCurrentStaleRebuilt
- SceneDiagram of crack growthCurrentStaleRebuilt
- ProductsVideo, article and shortVerifiedPublishing heldVerified again
Evidence
Content that can show its work.
Every factual claim is designed to carry its source. A claim without support doesn’t ship unless it is deliberately marked as editorial. Correcting a fact rebuilds only what depends on it.
Restraint
Explanation over spectacle.
Not every scene needs generative video. Real evidence, maps and diagrams come before generated footage.
The editorial test“If removing a scene doesn’t reduce understanding, the scene may only be producing motion.”
Autonomy
Autonomy is earned, not declared.
Each class of task earns its autonomy separately. Nothing is declared autonomous all at once.
- 0Commissioning
- 1Calibrated
- 2Auto with veto
- 3Autonomous
- 4Exception only
The models will change. Build the system that survives them.