Future-House--edison-client--owl-precedent-search

Owl — Scientific Precedent Search (FutureHouse Platform)

Owl is FutureHouse's "has-anyone-ever-done-X" agent (job name PRECEDENT, formerly known as HasAnyone). Where Crow answers what is known and Falcon synthesizes what the field thinks, Owl specifically asks whether a specific thing has ever been reported.

Use this skill when a researcher is about to start an experiment / build a tool / pursue a hypothesis and wants to confirm it's actually novel — and when not, to find the closest prior work.

Prerequisites

• pip install edison-client
• EDISON_API_KEY from https://platform.edisonscientific.com/profile

Minimal usage

import os
from edison_client import EdisonClient, JobNames

client = EdisonClient(api_key=os.environ["EDISON_API_KEY"])
resp = client.run_tasks_until_done({
    "name": JobNames.PRECEDENT,
    "query": "Has anyone tested therapeutic exerkines in humans or non-human primates?",
})
print(resp.formatted_answer)

The answer is phrased as a precedent verdict — "Yes, see X et al. 2023 …" or "No published precedent found, but the closest work is Y et al. 2020 which …" — with citations.

How to phrase queries

Owl works best with specific "has anyone" framings. Compare:

The agent expects the question to imply a precedent test. If the user's phrasing doesn't, rewrite it before sending.

Recipes

Pre-experiment novelty check

hypothesis = "Knock-in of a humanized FOXP3 allele into NSG mice rescues Treg development."
resp = client.run_tasks_until_done({
    "name": JobNames.PRECEDENT,
    "query": f"Has anyone tested whether {hypothesis}",
})
if resp.has_successful_answer:
    print(resp.formatted_answer)
else:
    print("Owl couldn't determine — try Falcon for a broader synthesis.")

Prior-art check before patent / paper draft

Run Owl on the central claim in a methods section to confirm it's actually new.

Finding the closest precedent (when answer is "yes")

Owl always returns the closest matching paper(s). Use those as a starting point for your related-work review. To go deeper, follow up with Crow on the specific paper:

prev_id = client.create_task({
    "name": JobNames.PRECEDENT,
    "query": "Has anyone used base editing to introduce SCD1 mutations in human iPSCs?",
})
follow = client.run_tasks_until_done({
    "name": JobNames.LITERATURE,
    "query": "What were the editing efficiencies and off-target rates reported in the prior work you found?",
    "runtime_config": {"continued_job_id": prev_id},
})

Cost / latency

• ~30 s – 2 min per call
• Credits: similar to Crow (cheaper than Falcon)

Failure modes

• has_successful_answer == False with answer "I couldn't find a clear precedent": treat as soft evidence of novelty, but it could also mean the question wasn't specific enough. Consider asking the user to refine.
• Vague "tell me about X" queries waste credits — rewrite into a "has anyone …" form before sending.