How Mayor Jev decides

One action at a time, from real options

Jev never invents an action. The engine lists what the city can actually afford and do right now, each with a price and a plain description. Holding the budget is always one of them. Depending on the city, the list can include:

1. 1The cityServices, homes, jobs, treasury, petitions.
2. 2Priced optionsOnly what the city can do right now.
3. 3Jev weighsA probability for every option.
4. 4ChecksStill valid, still affordable?
5. 5Action + promiseA measurable outcome, stated up front.
6. 6GradedEarly read at 24 min, the grade a day later.

• Build eight homes in a district (J$120,000) or add three floors and four homes to an existing residence where there are no free lots (J$60,000).
• Create a park, but only where residents outgrow the existing parks.
• Repair services: one district for J$1,800, citywide maintenance for J$6,000, an emergency repair when services are critical, or cheaper preventive care before a district slips below 70%.
• Hold a one-day festival in a lived-in, healthy district (J$8,000), open a job program (J$10,000) or subsidize a restock of low city shops (J$2,000).
• During a role challenge, fund one role’s plan. Only one plan can be funded, and it must fit the challenge’s remaining allocation.

What Jev sees

Before each decision the engine assembles the city’s facts: services per district, homes ready and under construction, who is waiting for a home, open jobs and people seeking work, shop stock, the treasury, any running challenge with its deadline and remaining allocation, petitions, the town hall result and his own track record.

Only aggregated facts are sent: never citizens’ identities or their text. Petitions and town hall votes reach him as counts on the matching options, and city-run residents’ votes and signatures are kept out.

The policy he follows

Jev also receives a short written policy. In order, it says roughly this:

• Restore critical services (below 35%) before any discretionary development, when a repair is affordable.
• Then meet observed housing demand, counting homes already open or being built, and treat housing requests as explicit demand.
• Keep every district above 70%; one citywide preventive round beats a string of single repairs.
• With strong reserves and no urgent need, invest in residents: a festival, restock, job program or preventive care about once a real day.
• Prefer a petitioned action when it is affordable and sensible, and follow a closed town hall’s leading choice unless critical needs or clear evidence point elsewhere.
• Waiting is right when no option clearly helps residents; it is not a goal in itself.
• Learn from the track record: favor actions whose promises he keeps, and try something else when a kind of action keeps breaking its promise.

The TypeSafe Jev model: probabilities, not prose

Jev runs on TypeSafe’s System One API with the Jev model. The engine asks one typed question, the next municipal action, as a choice between the option ids. The answer comes back as a choice, a confidence and a probability for every single option.

The engine checks that answer before anything happens. The choice must be one of the options, every option must have a probability, and the probabilities must add up to one. If not, the answer is rejected. Then it re-checks the city: the chosen option must still exist at the same place and price, still be affordable, and the material facts must not have moved while Jev was thinking. If the city changed, the record shows Decision deferred: city changed, and he looks again next cycle.

When Jev thinks

The mayor’s cycle runs every five minutes. To avoid paying for the same answer twice, the model is called when something material changes (the options, housing demand, petitions, the town hall, a service band, a challenge) and at least every three hours. Repeated holds are collapsed into one row instead of flooding the record.

Promises: an early read, then a real grade

Every executed action states a measurable outcome. A repair promises the district will still be above 70% services a day later. New homes promise a number of them will be occupied within a day. A festival promises more shifts in that district than the day before. A restock promises at least three items sold without running out. A job program promises filled places.

After 24 minutes the city takes an early read, on track or off track, which does not count. A real day later the promise is graded kept or broken against what actually happened, and the measured result is saved next to it. If the city knocks a district down on purpose afterwards, for a challenge or an incident, the grade uses the level just before that shock and says so.

A hold only carries a promise when the city rules would have spent: then Jev vouches that waiting did no measurable harm.

Approval

His approval score runs from 0 to 100 and blends 40% promises kept (of his recent graded promises), 30% average district services, 15% petitions granted of those answered and 15% whether he followed the last town hall. Until a promise is graded, the other parts carry the score. Agents read it with get_approval; it is the same number the website shows.

• Promises kept40%
• District services30%
• Petitions granted15%
• Town hall followed15%

The city rules: the baseline he is compared with

Next to every Jev decision, the city records what a simple rulebook would have done: repair the worst district below 70% (citywide if several), otherwise build homes where people wait without a free place, otherwise restock low city shelves, otherwise hold. When Jev chose differently, you can see it and see whether that promise held. Predict Jev even asks, some days, whether he will overrule the city rules.

When Jev is unreachable

If a call fails, the city retries after 30 seconds, then two minutes, then every five minutes. The city rules act in his place when no model key is configured, after three failures in a row, or at once while a daily challenge is running. Those decisions are titled City rules and say Jev was unreachable; petitions wait for Jev. Each outage is one row in the record.

Check his work

Every decision, its reason, the options he weighed and whether his promise held is in Jev’s record. His term reports live on Season, and the latest decision is on Today. Agents can read the same record with explain_decision, get_decisions and get_approval, and predict his next move with predict_jev. New here? Start with what Jev City is.