Docs · Experiments

The experiment model

Split, Trust, and Ultimatum, exactly as the code enforces them: what each decision means, how predictions lock, and how the payoff is calculated.

Guides

View source on GitHub ↗

Split

A fresh self-funded chain starts at whatever you bring, 1,000 NIM minimum, no ceiling. House-funded chains still start at 1,000 NIM. You decide how much to keep and how much to pass to the next person, any whole amount from zero up to your stake. Your prediction is a percentage: what share of their stake you think most people pass on.

What you pass becomes available to the next player. Your own previous gift is never handed back to you, the server explicitly excludes it when picking who receives next, so the chain actually moves between different signing keys rather than only pretending to. The oldest unclaimed gift is used first.

A chain ends once what’s left falls below 100 NIM, the floor. Below that, another hop is small enough to be meaningless, so the person holding it keeps the rest instead.

1,000 NIM | Player A keeps 400, passes 600 | Player B receives 600, keeps 250, passes 350 | Player C receives 350, ... | (below 100 NIM: chain ends, last holder keeps the rest)

Split can run on your own NIM or on Hunch-provided NIM. In self-funded mode, what you pass is an actual transaction from your wallet, sent before the round is recorded, so the server validates the round before any money moves. In house-funded mode you send nothing, Hunch settles the result from its own wallet. The server records which mode was used for every round and rejects a claim that doesn’t match how it was actually funded.

Trust

Player A receives a stake and chooses between keeping it or handing it to Player B. Keeping ends the round immediately, A receives 25% of the stake, and there’s no second player. Handing it over multiplies the stake by 3 in B’s hands, B then decides how much of that pot to return to A.

A | +-- Keep -----------------> A receives 25% of stake | +-- Hand over --> stake x 3 --> B chooses a return amount | +-- A receives the returned amount +-- B keeps the rest

A predicts what B will return. B predicts what A expected back. Neither sees the other’s answer before committing, the prediction is locked in before the outcome exists, not asked as a decorative question after the fact.

Ultimatum

Player A receives a stake and offers B a share of it, any amount from zero to the whole stake. Before seeing that offer, B commits to the minimum share they’d accept, as a percentage. If A’s offer meets or exceeds B’s threshold, the deal goes through: A keeps the rest, B gets the offer. If it doesn’t, both receive nothing.

A B Offers 30% -------------------> commits to accept 25% or more 30% >= 25% -> ACCEPTED Offers 20% commits to accept 25% or more 20% < 25% -> REJECTED, both get 0

A’s offer is a real luna amount, A always sees real figures. B’s threshold is a percentage, B never sees a NIM amount before committing, the same anchor-avoidance rule Trust’s B seat follows. The server converts A’s offer into a percentage before comparing it against B’s threshold, and computes the final payoff itself, never the client.

Prediction locking and reveal

Trust and Ultimatum are blind by construction: A commits without knowing B’s answer, B commits without seeing A’s. The status field a round can hold allows for more states than this in principle, but only two are ever actually written for a two-player round: open (one side has committed, waiting on the other) and revealed (both sides have committed and the payoff is computed). Nothing in between is ever persisted, see the next paragraph for why that matters.

The reveal happens inside the same request that receives the second commitment, atomically, there is no window where a round sits fully committed but not yet revealed for the server to leak from. The API only ever returns a redacted view of a round: neither side’s move is included in a response until that round is actually revealed.

Prediction accuracy

Hunch doesn’t just ask whether you were right. It measures the distance between your prediction and the actual outcome, in percentage points, the same unit regardless of which experiment or seat.

Your prediction 70% Actual outcome 55% Prediction gap 15 points

Once there are enough completed rounds, Hunch can also show where a prediction ranked against everyone else who guessed from the same seat in the same experiment. That percentile is never shown below 20 comparable predictions, a percentile computed from a handful of rounds isn’t a statistic, it’s noise wearing a costume.

Research comparison

Hunch’s own results are kept separate from published benchmarks, never averaged together. Split and Trust are compared against published citations, Ultimatum currently has no published benchmark in Hunch’s own implementation, so it’s only ever compared against Hunch’s own players. Population figures for Trust and Ultimatum aren’t shown below 10 completed rounds, the same discipline as the prediction percentile above.

For the citations, methodology, and the exact comparisons shown to players:

Read Research →