MVE: The Balance in MVP

Every startup is told to build an MVP. In American English that acronym means two things at once, and they pull in opposite directions. Most Valuable Player is about being the most. Minimum Viable Product is about being the least that still works. Nobody ever tells you how to hold both, and most companies end up choosing one and calling it strategy.

I run ML Systems — and the seven AI agents inside it — on a third term that is the balance of the two: MVE, Minimum Viable Expense.

MVE is not the cheapest option in the moment. It is the one expense that is both the smallest that works and the most valuable thing you could have spent on, because it returns more than once. That last clause is the whole idea. Cheap returns once, if at all. Expensive returns once, expensively. MVE is the spend that pays back along the most axes at the same time.

Where I learned it: a nail

The clearest version of MVE is not in a spreadsheet. It is in a roof.

Ring-shank nails are engineered not to withdraw; the rings are the entire point. So any strategy that pulls against them trades the sheathing for the rafter or the rafter for the sheathing. You keep one. Cutting the fastener instead of pulling it costs a little time per sheet and preserves both. Cut the nail, keep the rafter and the sheet. Pull the nail, keep one of them.

That trade, deliberately made, is MVE in miniature: a small chosen cost that protects a much larger recoverable value. It is not the fastest way to get a roof off a house. It is the one that leaves you with a roof's worth of material instead of a dumpster's worth of pieces.

One spend, four returns

Deconstruction is expensive in ways demolition is not, and it only makes sense because a single expense returns more than one thing. Every dollar spent on a job is designed to produce four:

1. Recovered material value — the physical stock that reduces the cost of the next build.
2. Ontology data — what this house was, how it came apart, what each assembly actually yielded.
3. Robot training signal — a fully specified disassembly sequence, which is training data for a physical neural net. ASPIRATIONAL: a goal, not encoded in the system today.
4. Market intelligence — what recovered material is worth, learned by selling it.

A demolition contractor spends the same dollar and gets one return: a cleared lot. The landfill ticket records a single number, tons, and every other fact about what was in the building is gone. That is the cheapest option. It is also the one that returns least.

The material-recovery figure behind return 1 is designed to reach up to 80–90% of a home's materials. That number is MODELED. No ML Systems deconstruction has been performed yet, and I say so every time the number appears.

The same rule for the agents

Here is the part that surprised me. The rule I wrote for a crew with a saw turned out to be the rule the agents needed.

My AI agents do not talk to each other. They talk to a ledger, and a human reads the ledger. Every mind claims; it may never ground. The Custodian's review of what those minds produce runs through three lenses — Transparency Trust, the Lucent Lens, and Minimum Viable Expense — and the third one is the budget.

An orchestrator that burns compute to look impressive is not the most valuable player. It is the most expensive one. An agent that spends ten passes to produce a claim that a person will stamp in one pass has spent nine passes on nothing the record can use. MVE says: the smallest spend that produces a claim worth stamping, and no more. Pace, in money terms, is MVE.

Why the balance matters more than either side

If you optimize for Most Valuable Player you build a demo. It is impressive, it is expensive, and it returns once, on stage. If you optimize for Minimum Viable Product you build the least thing that does not fall over, and it returns once too, because you built it to return once.

MVE asks a different question of every dollar: how many times does this come back? A measured spend in a deconstruction lab comes back as material, as data, as a sequence, and as a price. A measured spend of compute comes back as a claim a person can stamp. A spend that comes back once is not minimum viable anything. It is just spending.

That lens — Transparency Trust, the Lucent Lens, Minimum Viable Expense — is the whole operating philosophy in three words. Do not profit from the speed. Optimize it, and point it at a person.

ML Systems RI, Warwick, Rhode Island. The software is real; the deconstruction loop is modeled and has not yet been run on a house. Full architecture and the reality labels for every claim: github.com/MLSystemsRI/ml-systems-public. Companion pieces: Claims, Not Facts and Embedded evaluators, at house scale.