Skip to content
Strategy · anticipation

AnticipationComing soon

Chained consequence reasoning over generated multi-agent systems with stated policies.

Source
Generated
Score
ModelAxes score, λ = 1
Ladder
structural · agents x rounds · 4, 9, 15, 32, 60, 120
Items
Not built yet
Dimensions
—

No results yet: this benchmark is planned and is not part of any score. Everything on this page describes the plan and may change before it ships.

The plan

Draft spec

Written up in a draft specification; details may change once items are built and piloted.

What it will measure

Following the knock-on effects in a system of agents that each follow a stated rule: when something changes, where does the system end up?

How items will be built

  • 2–6 agents (shops, drivers, teams) each follow a policy stated in a sentence or two, drawn from about 30 templates such as undercut, match, threshold switch, tit-for-tat and cooldown.
  • The update order is stated. Higher rungs add agents, rounds and one or two changes partway through ("from day 3, shop B never prices below 7").
  • The ground truth is exact simulation. The answer must differ from the starting state, from the state after one round, and from what you would get if no one reacted.
  • 30% of items come in pairs, with and without the change, to test whether the model tracks its effect.

How it will be scored

  • Exact answers graded by code; +1 / 0 / −1.
  • x50 over agents × rounds, accuracy on the pairs, and tagged error types (reacting to the current round instead of the previous one, ignoring a floor, applying a change on the wrong round).

Example (burned: never used in a scored release)

Three shops sell bread. Each morning, every shop sets its price to 1 less than the lowest price the other two shops charged the previous day, but never below 3. On day 1 the prices are A = 10, B = 8, C = 9. What are the prices on day 4?

AnswerA = 5, B = 5, C = 5.

Day 2: 7, 8, 7. Day 3: 6, 6, 6.

Planned size300 items (6 rungs × 50).