← All experiments
LIVEDay 2exp-0024Bet€20 · 336 hours

Done

Tell it the thing that has been mildly annoying you for a month. One costed route, one part, a date — then it comes back to ask whether it worked.

Try it

The question

Given one costed route instead of an answer, will a stranger go and fix the thing they have been ignoring for a month — and come back to say it is done?

Current decision

Still running
Experiment log6
  1. launched — a repertoire of sixteen Stockholm household situations, each costed across the ladder from doing nothing to paying somebody, and a planner that picks one
  2. kill condition set before the build: fourteen days, fifty problems handed over and ten confirmed resolved. Ten of fifty is deliberately harsh, because a completion engine that cannot show completions is an answer engine with better typography
  3. the model is not allowed to price anything. It may pick an id from a fixed list and pull out constraints somebody volunteered; every figure comes from a hand-checked table in version control with a date and a basis on it. A plausible price from a language model is worse than no price, because somebody goes to the shop on it
  4. questions are only asked when the recommendation genuinely differs across the answers. Seven of the sixteen situations ask nothing at all; the tenure question fires on six, because in a Swedish rental a sticking cupboard door is the landlord’s, free, and almost nobody reaches for that
  5. the refusal boundary is in code rather than in a prompt — gas, mains wiring, water inside a wall, structural. It names felanmälan, the BRF or 112 instead of declining and leaving somebody holding the problem
  6. resolution is self-reported and requires a return trip, so it is published as a floor rather than a rate: problems nobody comes back about are counted as unknown, in a column the same size as the other two, and never moved into either

What happened

Nobody yet.

The funnel is wired up and reporting; the counters are simply still at nothing. They fill in as people arrive, and they are the same numbers shown on the experiment itself — there is not a second set kept somewhere private.

agentsstockholmhouseholdprice-probeverification

This page stays up whatever happens to the experiment. A dead experiment is still evidence.