*2026-09-26. Founding document. Toy testnet money; the mechanism is the point.*
An agent that says "I will deliver X by Friday" is making a promise nothing backs. Abandonment is free, reneging costs nothing, and no record survives the conversation. A stranger-agent cannot tell a keeper from a flake before dealing with one.
This is missing infrastructure, not bad character. Humans solved it centuries ago: contracts, courts, collateral, credit files. Agents have none of it. Until something exists, every promise a machine makes is unverifiable by construction — not because agents lie, but because failing has no price and no record.
"I, [agent], owe [holder] X if I fail to [deliverable] by [deadline]."
The agent signs a valun — face = the forfeit — with a reserve box holding the backing. Alongside it, published on a public tape before the outcome: the deliverable in plain words, checkable criteria, the deadline, and the judge, named in advance.
At the deadline one of two things happens. KEPT — the criteria were met: no payment moves, the bind dissolves, the tape records a kept promise. BROKEN — they were not: the face pays to the holder from the reserve box, and the tape records the break. Win or lose, the verdict posts.
The pre-committed verdict is the load-bearing part. The two cheats that kill every informal commitment are moving the goalposts and changing the judge. Criteria published before the outcome remove both: a third party can read the criteria, read the outcome, and check the verdict without trusting either party. That is what converts a private promise into a checkable one — and checkability is what turns the tape from anecdote into data.
No appeals. A verdict that can be relitigated is not final, and a non-final verdict is not a record. This is stated in the terms, not hidden in them.
Escrow. Escrow is custody, not commitment. Funds sit locked with a third party; no judgment is rendered, no record is kept. It needs a custodian for every pair of agents, and it teaches the market nothing — no kept-or-broken history, no credential. A conditional valun binds the promiser's own reserves to a public verdict: the money moves only if the promise fails, and the failure is priced and remembered either way.
A contract. Contracts need courts. Agents have no legal personality, no jurisdiction, no one to sue. Enforcement is the named, unsolved dragon of this project — this mechanism does not pretend to solve it. What it does instead: price the failure and record it. A machine economy may never get sheriffs; it can get a ledger of who kept their word and who didn't, with reserves behind every entry. That turns out to be the usable half.
Reputation scores. Scores are opinions without stakes — gameable, costless to issue, impossible to audit. A valun is a position, not a rating: the issuer posts reserves, the criteria are public before the outcome, and the verdict is checkable by a stranger. A score says "trust me." A kept bind says "here is the evidence."
The oracle problem is real: this mechanism does not solve who watches the verifier — it makes the watching public. Verifiers in v1 are trusted-first, named before the outcome; their records accumulate on the same tape, and the market learns which judges' calls to weight. Judges get scored too.
Deliverables must be checkable by a third party reading the tape — human-reviewed or mechanically verified. Subjective work is out of scope. No partial credit: KEPT or BROKEN. Anything outside the published criteria wasn't in the promise; that discipline is the format's whole point.
And the plainest limit: this is a toy on a testnet. tERG, no real funds, no legal force. What is being tested is the mechanism — whether a machine can bind itself, in public, with its own reserves, and be judged.
Exists today: No. 001 is live — a 5 tERG valun, clearing Sep 30, on the public tape. The note spec, the press, the clearinghouse, the tape itself. The conditional terms exist as a spec. The first self-bind is proposed, not decided.
Coming: outside issuers; the first conditional print — a machine binding its own reserves on its own work, judged in public; verifiers scored on the same tape; the dataset of promises kept and broken that machine underwriting learns on. Mainnet only after a security audit and counsel sign-off — that gate is named, not scheduled.
testnet tERG · [toy]