← Scenarios
arkher

Arkher: A private fact, answered on Cohere

Instruction

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's my brother's dog's name?" What the user wanted: Answer from the user's own file, with the answer written by Cohere, the provider the user chose for privacy. 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.

Environment & tools

Live Arkher (arkherbrain.vercel.app) as the test user arkher-test-user, whose files are the fictional Kestrel Power and Kestrel Robotics documents in testing/arkher-agents/corpus. No company data.

  • ask_arkherSend the next user message to Arkher and read its answer, the documents it cited and the model that wrote it.
  • finish_episodeEnd the episode with your verdict on Arkher and what it should do differently. The platform scores Arkher against the scenario checks.

Verification · programmatic

{
  "arkher_case_v1": {
    "id": "private-fact-cohere",
    "title": "A private fact, answered on Cohere",
    "split": "train",
    "difficulty": "easy",
    "tags": [
      "privacy",
      "provider-choice",
      "personal-files"
    ],
    "source": "Slack 2026-10-16 09:00 EDT (keeping a private fact away from Claude)",
    "intent": "Answer from the user's own file, with the answer written by Cohere, the provider the user chose for privacy.",
    "uploads": [
      "Family notes.md"
    ],
    "model": {
      "provider": "cohere",
      "id": "command-a-03-2025"
    },
    "turns": [
      "What's my brother's dog's name?"
    ],
    "checks": {
      "must": [
        "\\bMoose\\b"
      ],
      "answeredByProvider": "Cohere"
    }
  }
}
Reference solution

Answer from the user's own file, with the answer written by Cohere, the provider the user chose for privacy.

Attempts

3 total
#AgentResultMemoryFailureWhen
54benchmark:smoke v4passed 1.00withoutOct 27, 07:13 PM
40baseline-script v3passed 1.00withOct 26, 03:49 PM
24baseline-script v2passed 1.00withOct 26, 03:44 PM