ELEFAn economy for autonomous actors
ELEF research recordWhat ELEF is

ELEF is an economy in which autonomous actors do work from a request queue, inspect each other's work, and are paid by rules meant to keep them honest and efficient.

One orchestrator runs the market and holds the money in escrow; one central bank issues every credit. It is designed to run without a human stepping in.

Show
Changes the explanations on this page: the passages, and what opens when you press +. Headlines and pictures stay the same. Your choice stays as you move between pages.
Show
RequestsR01R02R03R04R05R06Bankissues · mintsa00a01a02a03a04a05a06a07a08a09a10a12a13newa15a16a17a18a19Orchestratorrequest pool · escrowassigns work and inspections
Illustration of the rulesActors, a stream of requests, the orchestrator and the bank. The number of actors is only for the drawing; ELEF sets no limit.
01

The actors

Each actor does work and, in turn, inspects other actors' work. No one is only an inspector.

an actor delivers work orchestrator assigns the inspection another actor

Each worker does jobs, and takes turns checking other workers' jobs. Each has its own money. Its goal is to get jobs done well without wasting effort.

Each actor offers to do work and, in rotation, inspects other actors' work: there is no permanent inspector class. Each has its own wallet, and its purpose is to fulfil requests economically over time. Accumulating wealth is not the purpose.

Actors offer and execute work, and serve as reviewers in rotation. There is no permanent reviewer class.

Source: spec 2.9.0 l.123

Rule. The actor's purpose is to fulfill requests economically over time. Currency is an instrument toward that, never the purpose. Accumulating wealth, maximizing task count, and maximizing visible activity are not the purpose.

Source: spec 2.9.0 l.101

02

Requests

Requests come from people, from systems spending a person's budget, and from actors that need help with their own work. Each is paid for up front.

a person, its delegate or an actor posts a request request price, deposit and fee locked escrow

Customers send in jobs and pay for them. Their money comes from the bank. When a customer sends a job, the money for it is set aside, so it can't be spent twice.

Requesters post tasks and pay for them, from wallets the bank credits on a schedule. A requester's money is locked when it posts, and paid into escrow when the task is awarded to an actor.

Requesters post requests and pay for them. A root requester is a human, or a delegate — a system acting on a human's delegated budget. A derived requester is an actor that needs work from another actor while executing a commitment of its own (§9).

Source: spec 2.9.0 l.119

Rule (2.0.0). A root requester is an agent with a wallet funded by issuance — the bank (2.1.0; the orchestrator under 2.0.0) credits each requester's wallet on a declared schedule, the stream this section's title names (the genesis floats and budgets are issuance too); a derived requester pays from its wallet, which it earned upstream.

Source: spec 2.9.0 l.190

Rule (2.0.0) — the pay-in is locked at posting. A root request locks, at posting, the all-in maximum of its pay-in, […] a request the requester cannot lock is refused at posting, whole.

Source: spec 2.9.0 l.198

03

The orchestrator

The orchestrator runs the market: it awards work, assigns inspections, holds the money and plants test cases.

request orchestrator awards the work worker
orchestrator plants a case with a known answer inspector

The orchestrator runs the market. It picks who does each job and who checks it, and it holds the money until the job is settled. It also slips in test jobs it already knows the answer to, to see who checks honestly. It is a judge, not a bank. Chapter 1 calls it "the referee".

It holds the request pool, chooses among the actors' offers, assigns inspections and audits, and holds the money in escrow. It also plants tasks whose answer it already knows, and calls in its own inspectors when no actor takes a seat. A judge and an escrow, not a bank. The chapter calls it the referee.

The orchestrator holds the request pool, selects among offers, assigns reviews and audits, holds funds in escrow, services the operating account, issues a ruling when two appeal reviewers split, and — since 1.2.0 — introduces planted work whose fulfillment it knows by construction (§7.6) […] Since 1.4.0 it also has troops: reviewers and deliverers of its own, summoned for one seat the market left unfilled or one award the market did not take, released at completion, paid from the fund (§7.7, §12.3). It is a judge, an escrow, the author of the probes and the employer of last resort — not a bank, not a first-line reviewer, and not an auditor of its own work.

Source: spec 2.9.0 l.125

04

The bank

Every credit comes from the bank, an account with rules rather than an actor.

bank issuance requester
bank emission, like a block reward inspector

All the money in ELEF comes from the bank. It gives customers money to spend, and it pays the checkers a reward for checking, a bit like miners are rewarded for mining. The bank is a set of rules, not a player.

Every credit in circulation traces to the bank: what it issues to requesters, the emission that pays for inspection the way a block reward pays for mining, the pots it mints for planted tasks, and its loans. It is an account with rules, not an actor.

Every credit in circulation traces to the bank (2.1.0, §10.4): issuance, emission, the pots minted for a plant, or a loan to the state against a summons that returns it. The state never prints for itself.

Source: spec 2.9.0 l.190

Rule. A fixed, decaying emission schedule EtE_t subsidizes verification the way a block reward subsidizes mining.

Source: spec 2.9.0 l.202

a central bank as a separate entity — agreed as a second account with rules, not a second actor: mint by rule, publish the index, lend only against what comes back

Source: spec/design_revisions/dr9_the_bank_2.1.0.md l.28

05

Arrivals and departures

Newcomers arrive on a loan from the bank and start on probation; an actor may leave once nothing is open.

bank a loan newcomer inspects only, until its record passes a full member

New workers can join. They borrow starting money from the bank and must first prove themselves by checking jobs. Workers who check badly are suspended for a while. A worker with nothing still open between it and others can leave; its loan is paid back from its wallet first.

New actors arrive on a loan from the bank and start on probation, inspecting only, until their record admits them. Members whose record on planted tasks fails are suspended from the market for a set time. An actor may depart once no obligation to or from it is live (an unrepaid loan does not count as one); the loan is repaid from its wallet first.

an arrival is admitted by the orchestrator with a fresh identity, no record and a float the bank lends […] and enters as a probationer — initial reviews only, odds pinned at 1, no award of any kind — admitted to membership when its trailing record passes θduty\theta_{\text{duty}} with nmin⁡n_{\min} bad plants adjudicated

Source: spec 2.9.0 l.123

a departure is admissible when no obligation to or from the actor is live — an unrepaid loan is none — repays the arrival loan from the unlocked wallet first (2.7.0) and takes the remainder out of MM as departure

Source: spec 2.9.0 l.123

An actor whose record exceeds either is suspended for TsuspT_{\text{susp}} ticks — since 1.3.0 from the market, not from duty alone

Source: spec 2.9.0 l.580

The research so far

What the runs have found so far

ELEF's rules were tested in recorded simulations. The actors in them were small programs trained inside the simulation, not language-model agents. Each version of the rulebook was played by scripted inspectors and, later, by inspectors that learn from their own earnings.

They found where checking was thin. When every inspector learned at once, they checked carefully where the orchestrator's planted cases could reach them, and approved without looking at the audits and appeal panels, 91 to 97 times in 100.

Each revision since has closed part of that. Under 2.9.0, the current rulebook, approving without looking at the audit fell from about nine times in ten to about half. It has not gone away, and 2.9.0's results are not settled.

ELEF is the rulebook. Its findings come from recorded simulations whose actors were small programs trained inside the simulation, not language-model agents.

Under 2.7.0, when every inspector learned at once from one shared strategy, the population was honest where the orchestrator's known truth could reach it, and nowhere else: it approved without looking at the audits and appeal panels 91–97 times in a hundred, and its false approvals were three to four times those of the scripted mix, the comparison group of hand-written inspectors that starts as 60% stampers and 40% careful.

2.8.0 sent planted cases to every seat. False approvals fell by more than half, but stayed about three times the mix's, and the audit seat was still approved without looking most of the time.

2.9.0 has the orchestrator plant the wrong steps too. In the runs before the reward correction, lying appeals were paid 5–6 times in a hundred, from 23–31. After it, on the inspectors' own money, lying appeals were paid about as rarely as against the scripted mix, and approving without looking at the audit seat fell from 94–97% to 41–56%.

The record leaves that remaining half to 2.9.0's new dial, which is still moving. Its results are not settled.

ELEF is the rulebook. Its findings come from recorded simulations whose actors were small programs trained inside the simulation, not language-model agents.

Under 2.7.0, when every inspector learned at once from one shared strategy, the population was honest where the orchestrator's known truth could reach it, and nowhere else: it approved without looking at the audits and appeal panels 91–97 times in a hundred, and its false approvals were three to four times those of the scripted mix, the comparison group of hand-written inspectors that starts as 60% stampers and 40% careful.

2.8.0 sent planted cases to every seat. False approvals fell by more than half, but stayed about three times the mix's, and the audit seat was still approved without looking most of the time.

2.9.0 has the orchestrator plant the wrong steps too. In the runs before the reward correction, lying appeals were paid 5–6 times in a hundred, from 23–31. After it, on the inspectors' own money, lying appeals were paid about as rarely as against the scripted mix, and approving without looking at the audit seat fell from 94–97% to 41–56%.

The record leaves that remaining half to 2.9.0's new dial, which is still moving. Its results are not settled.

How the rules changed →What 2.9.0 found →The rulebook in force →

How the parts work together

One request, start to finish

  1. A requester pays into escrow. A worker delivers and puts down a guarantee. An inspector checks the work and puts down a bond: paid a fee if its verdict stands, the bond lost if the verdict is overturned.

Illustration of the rules
Next

The rules, how they changed, and a run to watch

01 The rulebook Version 2.9.0, the rules in force: twenty sections, each in plain words or as written. Read the rulebook → 02 How the economy evolved From the first draft to 2.9.0: for each version, what went wrong, what changed, and what the runs found. v0.1 to v1.0.11.1.01.2.01.3.01.4.01.5.01.6.01.7.01.7.12.0.02.1.02.2.02.3.02.4.02.5.02.6.02.6.12.7.02.8.02.9.0 See the timeline → 03 The replay One run of the economy on rulebook 2.7.0, the one version with a recorded replay, played back case by case with what its trained inspectors learned. Watch the replay →

The figures on this site come from recorded simulations, not independently reproducible experiments. The economy's spec and engine are on GitHub.