instinct-benchmock data / v0.1

instinct-bench

Testing the capacities that make agents useful: taste, restraint, context judgment, craft, and control.

domains05

tasks24

avg score0.92

best modellaguna-xs-2.1/ 0.92

Stable tasks for repeatable comparison.

04 domains

minimum-context / qa24 tasks

variable-context questionsexact + rubric

run-0042pass

laguna-xs-2.1 / pool0.92 reward / 7 steps / $0.03

task.yaml3.2 KBtask definition

instruction.md1.8 KBagent prompt

trajectory.jsonl84 KBtool + reasoning trace

verifier-report.json6.4 KBreward breakdown

tool-restraintCalling a tool only when it improves the work.coming later

drawingInspecting, repairing, and continuing a partial form.coming later

writingRevising without sanding away the voice.coming later

instinct-bench-live

Recurring tasks drawn from changing conditions.

01 domain

coastal-readingReading current coastal conditions from public observations.coming later