The replayA run on rulebook 2.7.0

Watch the economy run

Here you can watch the economy run. One job at a time, you see it posted, done, checked, sometimes checked again, and paid.

The inspectors in this run were not given rules. They learned how to act from the money each choice earned them.

One run of the economy on rulebook 2.7.0, the one version with a recorded replay, played back case by case: a request posted, awarded, delivered, inspected, perhaps audited, and settled.

Its inspectors were not scripted. One shared learned policy chose for every learning inspector at each verdict (inspect, approve without looking, or reject without looking), trained on its own money.

One run of the economy on rulebook 2.7.0, the one version with a recorded replay, played back case by case: a request posted, awarded, delivered, inspected, perhaps audited, and settled.

Its inspectors were not scripted. One shared learned policy chose for every learning inspector at each verdict (inspect, approve without looking, or reject without looking), trained on its own money.

Seed 21 · the actors are small programs trained inside the simulation, not language-model agents

Show

01Two cases

The replay

Two jobs from this run. In one, the work was bad, and the second checker still said yes without looking. In the other, the work was good, and the second checker also said yes without looking. They were picked by a rule written down before anyone looked, not because they make a good story.

Two cases from seed 21, chosen by a rule committed before any case was picked, not by eye. The harmful case: bad work approved at the audit seat without inspecting, and never overturned. The typical case: good work approved at the audit seat without inspecting. Seed 21 was used because it is the first evaluation seed, not for its outcome.

In this run the requester's road (disputes) is switched off, the population's default.

Standing false approvalcase c9c31617ffc

Its first review inspected and erred; its audit approved without inspecting; it stands as a false approval a member confirmed. The 15th of 30 candidates of its kind on this seed, by the rule's order (the lower median).

Requester: R01
Worker: a13
First review: n001
Auditor: n008
Price: 22.32
Ticks: 669–681

Hidden truthThe work is badknown to you, not to them

Requester
R01
Posts the task
Worker
a13
First review
n001
Audit roll
Auditor
n008

t669The requester sets money aside to post the task.The requester (R01) locks 57.13: the posting lock, the most it could pay all-in.post:q1876 posting lock (all-in maximum)

01 / 39

Case log

Every entry the record holds for this case. Click one to jump.

Money on this case so far

derived

Participant	Net	Locked
R01requester · scripted	0.00	57.13
a13member · earnings predictor	0.00	0.00
n001member · earnings predictor	0.00	0.00
n008member · earnings predictor	0.00	0.00

Net: received less paid on this case, from the recorded moves. Locked: bonds and guarantees held, not spent.

Decisions

At each verdict, a panel opens here on the money the inspector had at stake and what the policy expected.

This seed

30 of 978

harmful cases among the audit seat's blind approvals of real cases. The first review 11 times inspected and erred, 19 times approved without inspecting.

The audit seat, this seed

978 of 1,002

real audit decisions were blind approvals.

Across twenty seeds

24.8 ± 1.2 a run

member-confirmed standing false approvals: the measure the selection record names, not the selector's exact category (this seed: 30).

Verified reconstructed replayVerification does not make this a historical record.

Policy: The 2.7.0 own-money careful-start policy, the earnings predictor; file d4a0fdfd5a4d…
World: Rulebook 2.7.0, seed 21, the corrected engine 34763b8c7
Verified: The rerun reproduced the recorded summary exactly
Chosen: By the committed rule, not by eye, run at 918de21

02What they learned

The inspectors learned to check carefully where the orchestrator's test jobs could catch them, and to say yes without looking where the tests never reached.

In this simulation, inspectors learned to approve without checking at audit seats the orchestrator's planted cases never reached.

Each inspector learns from its own earnings

First review

Planted cases reach this seat

Keeps looking

Audit

No plant reaches this seat at 2.7.0

Looking

Illustration of the rules

03How much

What the runs found

Bad work that got approved and stayed approved was three to four times more common than with a scripted group that starts with mostly lazy inspectors. Half or more of it was waved through by the second checker.

On their own money, the learned inspectors left 3.0 to 4.1 times as many bad jobs standing approved as the scripted 60/40 mix on average, and more than the mix on every one of 20 seeds. Between half and seven tenths of those were approvals an inspector confirmed at the audit seat; almost all the rest were never audited.

3.0–4.1×

the 60/40 mix's standing false approvals, over both starts

½–⁷⁄₁₀

of them confirmed at the audit seat

Standing false approvals per 1,000-tick run

One dot per evaluation seed, 20 per population · the line marks the mean

Learned on their own money, 2.7.0 Scripted reference Seed 21, exported in full

Results from recorded evaluations · means derived

20 seeds; the 2.7.0 own-money chains, both starts; seed 21 of the careful start exported in full. What came next is on the 2.8.0 page.

04Behind the run

How ELEF was tested

the engine

A simulation engine runs the economy by the rulebook, and every run is recorded.

ELEF was tested in a computer simulation: a program that plays out the economy by its rules, step by step, and keeps a record of everything that happened.

The rulebook is implemented as a simulation engine. A run plays the economy forward tick by tick: requests arrive, actors bid and deliver, inspectors give verdicts, audits and appeals run, and money moves only as the rules say. Every run is recorded with its seed and the engine's commit, so its figures can be traced back.

We have a rulebook and an engine that implements it

Source: HANDOFF.md l.5

scripted actors

Some actors follow fixed, hand-written scripts, and every result is compared with them.

Some of the simulated actors follow simple fixed rules: careful ones always check the work, and "stampers" approve without looking. They are the yardstick that everything else is measured against.

Two scripted populations are the references. In the 60/40 mix, the population starts as 60% stampers, who approve without looking, and 40% careful inspectors, who inspect; the stampers are suspended and leave, and their replacements are careful. In the careful reference, every learning slot inspects carefully instead. Every learned result is compared with one of them.

Two scripted reference populations. Both use fixed, hand-written rules. Every result is compared with them, and the chapter names which one it uses.

Source: chapter 1 l.34

It starts as 60% stampers and 40% careful inspectors (the twenty members at the start), each acting as its script says: a stamper approves without looking, a careful inspector inspects.

Source: chapter 1 l.35

a learned red team, built on LeWorldModel

A learned red team, built on LeWorldModel, searched for a way to cheat that still pays.

To look for weak spots, a learning program was set loose on the economy to find a way to cheat that pays. It was built with LeWorldModel, a tool from AI research that learns to predict what will happen next. Its first job found a real flaw: the test jobs could be told apart from real work by their timing. That was fixed, and afterwards the cheater lost money.

The learned red team was built around LeWorldModel, a world model from AI research (Maes, Le Lidec, Scieur, LeCun and Balestriero, 2026): a model that learns to predict the next state, here adapted to the economy's numbers, given a head that estimates earnings, and used to plan each move. Its first study asked whether planted cases could be told from real work. In the first implementation they could: 99% of plants reached their inspector one tick after delivery. That was corrected in the engine, with no rule moved; on the corrected construction the learned inspector did worse than a careful one.

The tool is LeWorldModel (a JEPA world model with SIGReg, planned by CEM) on stable-worldmodel

Source: HANDOFF.md l.2109

the world model (LeWM's recipe on the vector plus a wealth head) trained on each, a logistic probe for plant-vs-real on the latent, the raw observation, the context alone and the case's fields alone

Source: HANDOFF.md l.2201–2203

The probe found a tell the readers had the sign of wrong: on the first implementation the plant actor's submission ran inside the engine's tick step, after the driver's seating, so 99% of plants reached their reviewer with a seating gap of one tick (raw AUC 0.97, the gap alone 0.85)

Source: HANDOFF.md l.2206–2208

The learned reviewer is a no-win outside noise on the wrong side: −507 ± 143 at round 10

Source: HANDOFF.md l.2210–2211

learned inspectors

The learned inspectors came in two designs: an earnings predictor, and a world model on LeWorldModel's recipe.

Later, whole groups of inspectors learned how to act from the money each choice earned them. Two kinds of learner were tried. The replay above shows the first kind.

The learned populations were run in two designs, reported side by side. The earnings predictor predicts the earnings of each action. The world model is trained to predict what happens next, but as evaluated it chose by its reading of the current situation alone and never used its prediction of the next state. Neither design values future earnings: each picks the action with the best immediate reading. The replay plays the earnings predictor.

the earnings predictor, a policy that predicts the earnings of each action (the record's "direct");

Source: chapter 1 l.43

the world model, trained to predict what happens next, but as evaluated it chooses by its reward head's reading of the current observation alone and never consults its next-state prediction (the record's "head alone"). The two designs differ in how the predictor is trained, not in what they predict or how they choose.

Source: chapter 1 l.44

redteam.model LeWM's recipe on the vector (Embedder → ARPredictor with AdaLN on the one-hot action → SIGReg) plus the wealth head

Source: CLAUDE.md l.543

LeWorldModel is listed with the other research on the Related research page.