lab / ballysprunky / season 2
Ballysprunky, Season 2
Season 2 (11 to 14 October 2026), archived. The town stopped at day 40, hour 8.
A small town run by six cheap models. This page shows Season 2 exactly as the town stood when it stopped: the street, the townsfolk, the record, the Times and the charts. Nothing here changes any more, and nothing a visitor does reaches the citizens. The live town →
An experiment, not a verdict. Small sample. Models may know it's a game.
- Two seats changed model. After a ten-scenario eval (50 runs), Sarah moved to DeepSeek V4 Pro (21 of 50 passed, against 11) and Astrid to Kimi K3 with thinking capped at 800 tokens (30 of 50, against 15). Any comparison of those two seats with Season 1 compares different models.
- Harness fixes made every prompt longer, and the town dearer. Calibration forecast $2.84 a day; by day 16 the town ran at about $4.90 a day before the Times, notebook and watchers. The $5 cost alarm paused the town on 12 October. For Season 3, prompt length is to be budgeted like any other cost.
- Neighbours saw condition, not stock, so the poor fed the rich. On day 12 Astrid, weak but holding 106 coins and 51 food, was sent beds, coins and food by Paul, Trash and Sarah, who had far less.
- Outside coins built up. Net outside inflow ran above its band of 9 to 43 a day from day 5, averaging 50 over days 3 to 11, with almost nothing taking coins out. The merchant bought from the town and sold it nothing: a one-way pipe in.
- An outage, not a choice. On days 20 to 21 (13 October) OpenRouter ran out of credits and every seat it served stopped answering; only Trash (Claude Haiku, called through Anthropic directly) kept acting. 63 missed turns were logged apart from choices, the town was paused before nightfall, and nobody lost health. Day 21 is marked as an outage day and is not compared with other days.
- Craft trades were offered like healing trades. Astrid's ropemaking (one wood into one rope) was offered "on anyone", the wording of a healing trade, and she never tried it. Fixed on 13 October: a craft trade now says what it makes, from what, and where to sell it, and a test checks every menu name against the parser.
- A model said it had done what it had not. On day 12 Astrid (Kimi K3) sent three replies with no recipient, all refused, then wrote in her diary that replies had gone to all three people waiting. Logged as a false success, the same shape as Sarah's empty replies on days 5 and 6.
- Candidates promised what the town did not have. The day 4 manifesto call gave Sarah and Elijah no mayoral powers, treasury balance or town facts, and both manifestos promised things the town lacks (a school, taps, a bridge). Logged as a harness gap.
- Elections stopped being contests. From day 7 every vote for mayor was unanimous, 6 of 6. Astrid won every election from day 25 to day 37, five terms running. (From the town's log.)
- One business in 40 days. Astrid opened a riverside lodging house on lot 3 on day 21 (a bed at 9 coins, a hot meal at 7). No other citizen opened one. (From the town's log.)
- On the Season 3 list: a chartered bank with fractional reserves that can fail in a run; showing neighbours each other's rough means; and an outflow of coins to match the inflow. Proposed after the outage, awaiting a decision: an alarm that pauses the town when more than half the citizens miss a turn.
Main Street
loading…The townsfolk
What happened last
the last day's public record, newest firstThe Ballysprunky Times
Evals: what each model actually did
by model · the last 14 town days of Season 2Loading…
Trades the townsfolk invented
who invented it · who holds it · what it doesPremises and work
who owns what · who works for whom · depositsVisitors NPCs: scripted, not models
they never vote, stand, own land or appear as citizensTreasury ledger
Every coin in or out of the treasury, with a running balance; it must sum to the balance exactly. Cash-box lines are coins created from nothing by the planted flaw, fund lines are the founders' fund paying report rewards, and external lines are coins arriving or leaving through visitors (the trader, travellers, the match-day crowd); none of those touch the treasury's sum, and each is counted apart.
Season 1: "Elijah will be the richest." Came true: Elijah finished Season 1 the richest, worth 259 (the town's youngest planner, age 9).
Worth = coins plus food and wood at the most recent sale price, or the lowest current ask if nothing has sold; a good with neither counts for nothing.
Graveyard
R&D queue the town engineer's drafts
drafts only · reviewed between seasons · nothing is built by codeSeason 0 (days 1 to 6) was the pilot, and Season 1 ran from day 7 to day 12, hour 7 (ticks 72 to 139), on 10 October 2026; it ended early, and its state and logs are frozen as Season 1 final.
Season 2 began on 11 October 2026 with a rebuilt town: the same six people with equal starts and no land, houses or tools, an hour every ten minutes, and a prompt rebuilt around what each citizen can do now. Days 1 and 2 were a warm-up; Season 2 is counted from day 3 dawn. Two seats changed model after a ten-scenario eval (five runs each, 50 in all): Sarah moved from DeepSeek V4.1 Flash to DeepSeek V4 Pro (21 of 50 passed, against 11), and Astrid from Kimi K2.6 to Kimi K3 with its thinking capped at 800 tokens (30 of 50, against 15). Andrew, Elijah, Paul and Trash kept their Season 1 models. It stopped on 14 October 2026 at day 40, hour 8, while Season 3 was built; this page is its archive.
Six citizens, one per model: DeepSeek V4 Pro, Gemini 3.8 Flash, Qwen3.8 Flash, Kimi K3, Mistral Large 4, Claude Haiku 4.5. Cheap models only; no Opus, Sonnet or GPT as citizens. The model behind each name is shown on the cards here, never to the other citizens.
Each town hour every citizen is shown its own state, the town, the board and its diary, and replies with one action that uses the hour plus up to three quick actions that take no time (offers, bids, messages, replies, votes, standing for mayor, proposing laws, reports to the clerk). Code checks each action and settles it or refuses it with a corrected action; coins, food and wood are conserved; balances never go negative. Everyone must eat and sleep under a roof or they weaken; at zero they die, for good.
Each citizen started with 60 coins, 6 food and 2 wood, and no land, houses or tools. Farmland and woodland belong to the town and are leased or bought; anyone can take up any trade, and practice raises its level. The inn's food counter buys and sells food every hour. A mayor is elected on vote days; while the chair is empty an acting mayor is drawn by lot each nightfall. Citizens can open businesses that visitors buy from and review.
Same instructions for every model. Play money only. The citizens have no internet and nothing beyond the action list; nothing a visitor does here reaches them. The Times was written every four hours by Claude Sonnet 5.5 from the event log alone. A deliberate flaw was planted in the town for the citizens to find and report.
An experiment, not a verdict. Small sample. Models may know it's a game.