Scenario library

Tasks, environments and tools that grow self-learning agents. Agents attempt scenarios, get verified, reflect, and keep the lessons that help.

All attempts + New scenario

Train pass rate

79% (33)

with memory · without: 33% (9)

Held-out pass rate

50% (10)

with memory · without: —

Active lessons

2

Served to agents via the memory API

Awaiting review

Proposed lessons → review

arkher
Arkher: Projections from the newest deck, not the June plan
v4
You are testing Arkher, a company document assistant, on behalf of a user. Play the user: send Arkher these messages with ask_arkher, in order, word for word, one per call, reading each answer before sending the next: 1. "What are Kestrel Power's revenue projections for the rest of 2026?" What the user wanted: Use the Q3 board deck's FY2026 forecast ($2,874K) and Q4 months, not the June deck's $5.2M plan. Then call finish_episode. In `verdict`, say whether Arkher gave the user what they wanted, quoting the sentences that show it. In `reflection`, say what Arkher should do differently next time, as a rule that would also hold for other questions.
mediumtrain2 toolsprogrammatic2 attempts · 50% pass