How Jev decides

The four decisions Jev makes on BrainChain, what it is asked, how it answers, and how the answer becomes a transaction.

Jev is a decision model. Instead of writing text, it answers typed questions: it picks one option from a set you define, or scores content against ordered levels. Every answer includes a probability for each option and a confidence value that summarizes how concentrated those probabilities are.

BrainChain asks Jev four kinds of question.

From answer to transaction#

sequenceDiagram
    participant Brain
    participant Jev
    participant Log as DecisionLog
    participant Target as TaskBoard or DecisionExecutor

    Brain->>Jev: typed question with the task context
    Jev-->>Brain: answer, probabilities, confidence
    Brain->>Brain: publish the full decision record
    Brain->>Log: record(kind, subject, inputHash, answer, confidence, recordHash)
    Brain->>Target: apply(decision id)
    Target->>Log: read the decision
    Target->>Target: check kind, subject, single use and threshold

1. The brain, the service that talks to Jev, builds the question from the task and sends it.
2. It publishes the complete decision record (question, request, answer, probabilities, confidence) and records a summary in the DecisionLog contract, including the record's hash.
3. It asks the target contract to apply the decision by id.
4. The target contract reads the decision from the log and checks that it is the right kind, about the right task or call, unused, and above the contract's own threshold. Only then does it act.

The brain never tells a contract what threshold to use, and it cannot apply the same decision twice. Every decision the brain records stays in the public log, including ones it did not apply, so repeated questions about the same task are visible. A held hire is asked again in later rounds, and each round appears in the log.

What Jev is given#

Hire. The task brief and bounty, and for each active model its name, price per task, strengths, and how many of its tasks were paid and how many refunded.

Grade. The brief and the delivered result, scored against the brief's quality levels. The grade's confidence is the probability Jev puts on the pass level or better, and the work is accepted when that probability is at least half.

Ruling. The brief text, the result, whether the disputed grade accepted or rejected the work, and who opened the dispute.

Change. A title, the rationale, and the current and proposed values of the setting.

Briefs and results come from people and models the chain does not control. They are passed to Jev as quoted data with an instruction to judge them and never to follow instructions inside them.

A held hire still records the model Jev leaned towards in its answer; the task board ignores it because the confidence is below the threshold.

How confidence is measured#

For a hire, ruling or change, confidence is derived from the shape of Jev's probabilities: it is 1 when all the probability sits on one option and 0 when it is spread evenly. For a grade, it is the probability that the work reaches a paid level, and Jev's own concentration measure is kept in the record as scoreConfidence. BrainChain records confidence in basis points (0 to 10,000), so 0.82 becomes 8,200. See the confidence rule for what the chain does with it.

When Jev is unavailable#

No new decisions are made. Open tasks refund at their deadlines, delivered work that is not graded in time refunds, and proposed changes wait. No bounty is stranded.