What this is
One person, a fleet of AI agents on a single server, and one discipline that turns out to matter more than any of the tooling.
An agent will tell you it is done. The tests never ran, the file was never written, the service is not up. The claim was generated from context, not read off the machine — and a confident report looks exactly like a true one until somebody asks for the receipt.
That is not a model defect you can prompt away. It is a missing contract.
The rule
Every claim of state carries a source and a date — or the words NOT VERIFIED.
A claim of state is any sentence saying something is or was the case. The mark travels attached to it: path:line (date), command (now), URL (date) — or NOT VERIFIED, with half a line saying why. Opinion and reasoning carry nothing; they are visibly not facts.
It sounds trivial. It is not. It forces the distinction between memory and observation at the one moment where noticing it changes the outcome: while the sentence is being written.
A live process, an exit code of zero and the word “done” all sit upstream of PASS. Tests passing is not validation — it is evidence that the code does what the test says, which is a much smaller claim than the one usually made from it.
What it built
Living World is an Ultima Online shard on ModernUO, running on one VPS since 5 September 2026. It is a game server, and it is also the load test for the method.
- rootless Podman, with a systemd Quadlet as the single recreation path
--read-only,--cap-drop=all,--security-opt=no-new-privileges, non-root user namespace- image pinned by digest, not by tag
- daily backup with a restore that was actually performed, not assumed
- a dated evidence trail and an append-only decision ledger behind every one of those choices
None of that configuration was written by hand. The entire human contribution was asking “did you verify that, or are you inferring it?” until only the verified part was left.
The shard has almost no players, and that is not the point of it. A running server is something a stranger can connect to and check for themselves, which is worth more than any screenshot.
The contracts, free
The files that enforce all of this are public domain — five of them, no dependencies and no install. Copy one into your repository and your agent starts telling you what it ran, when, and what came back.
github.com/edgebuildlabs/evidence-contracts
Its companion says how you ask for the work so that proof is possible: a work order template and the rules behind it, also public domain.
github.com/edgebuildlabs/work-order-contract
And the third says how you talk to an agent that runs on a machine you do not control — and how its delivery outlives the session, the model change and everybody's memory. A repository used as a channel, with the contract inside it.
github.com/edgebuildlabs/channel-contract
The fourth says who may say yes — and how that yes reaches an agent that another agent asked. Once the one that asks is not the one that decides, the request and the yes travel on different paths, what an agent may do alone is written down, and what runs with no human at all has its worst case written and accepted.
github.com/edgebuildlabs/mandate-contract
The fifth says how the record keeps its meaning across writers that do not remember it. An operation run by agents is rewritten every session by a writer with no memory of the last one, and the record can lose its meaning while every sentence in it stays true. Four rules — record at the act, three marks on every rule, a change preserves meaning, proof by resumption — and the skeleton of the handoff the next session reads first.
github.com/edgebuildlabs/continuity-contract
Public domain under CC0. Copy it, change it, ship it, credit nobody. If it saves you one wrong “done”, it paid for itself.
Contact
Edge Build Labs is one person, working full time, in Brazil. If you are running agents on real infrastructure and want the same discipline applied to yours, open an issue on the contracts repository, or write to contato@edgebuildlabs.tech. Either way it is written, asynchronous, no calls.
github.com/edgebuildlabs/evidence-contracts/issues
The story behind this is a book, in Portuguese, free and without sign-up: Não está provado, by Evandro Henn.