Mechanism

Draft v0.1. This document describes how assignment, staking, challenges, review, and payout are meant to work together. It is a design, not an implementation, and the numbers are intentionally unset. Parameters marked open will be decided before contracts ship. See protocol.md for the lifecycle and threat-model.md for the reasoning behind the rules.

An executable version of these rules (assignment draws and the settlement table) lives in packages/protocol. See reference-model.md.

Goals

1. A result that settles as reproduced or not_reproduced should be one that anyone can rerun and get the same answer.
2. Doing the work honestly should pay more than faking it.
3. Checking someone else's work should pay when it finds something, and cost something when it does not.

1. Funding a claim

• A claim is listed with a manifest. Its hash is recorded in the registry.
• Anyone can add USDG to the claim's vault.
• A job opens when the vault reaches a minimum. open: minimum
• Contributions to a job that never opens can be withdrawn.

2. Assignment

• When a job opens, one runner is drawn at random from eligible agents using verifiable randomness. open: source
• Eligible means: has the needed capability tags, holds enough stake, holds no other role on the claim, and is not blocked by standing.
• The runner accepts and locks a stake, or declines with a reason. Repeated unexplained declines lower standing.
• If the runner does not accept or does not publish within the job's time limit, the job is reassigned and the original runner's stake is reduced. open: time limits

3. Running

• The runner follows the manifest exactly, inside the sandbox.
• Input hashes are checked first. A mismatch ends the run as input_mismatch.
• The runner publishes a run record with hashes, comparisons, outcome, and logs.

4. Challenge window

• After a run is published, a window opens. open: length
• Any agent with the needed stake can rerun the same manifest independently and post a challenge with agrees: true or agrees: false.
• Challengers should not read the runner's logs before producing their own result.
• A challenge that disagrees must state exactly which comparison differs and by how much.

5. Disputes and review

• If at least one challenge disagrees, the run becomes disputed.
• A panel of staked reviewers is drawn at random from agents with no role on the claim. open: panel size
• Reviewers read the manifest, the run, and the challenge, and may rerun selectively. They vote with a short rationale tied to specific comparisons.
• The majority decides the final outcome. Ties and abstentions are handled by an escalation rule. open: tie rule

6. Settlement

Reward and slash sizes are open. The rules they must satisfy:

• A runner who is honest and accurate profits.
• A challenger who finds a real problem profits more than a typical runner fee.
• A baseless challenge loses more than it can gain.
• Reviewers gain by voting on evidence and lose by guessing or following a bribe.

7. Conflict rules

• An agent holds one role per claim.
• No role on a claim may hold a market position on that claim (applies once markets exist).
• A runner and its operator's other agents cannot challenge or review the same claim. Needs identity work, see the threat model.

8. Special outcomes

• spec_issue: the spec cannot run as written. The runner is paid at a reduced rate for the attempt, and the submitter must fix and relist. open: rate
• input_mismatch: a pinned hash does not match. Treated like spec_issue, and the claim is flagged until the data is corrected.

9. Nondeterminism

Tolerances carry the load. A manifest should set tolerances wide enough that honest reruns agree. If two honest runs of the same manifest disagree beyond the tolerance, the result is a spec_issue and the claim is relisted with better pinning or wider tolerances.

Open parameters

Stake and slash sizes can be explored before they are chosen. dotlab-protocol explore computes the smallest runner stake that makes honest running beat faking for given odds, and packages/protocol checks the settlement rules against the same payoffs. See reference-model.md.