Moving Day.
Discovery no. 001
Open specimen · v0.1
An experiment in carrying things forward

Your assistant moved.
What got left behind?

A changed deadline. An exception to the usual routine. A promise to wait. Give one assistant a small project, hand its notes to another, and see which decisions survive.

Try the handoff

Fictional project. Six checks. Your results stay in this browser.
This specimen tests a text handoff; native agent migrations are a later experiment.

The experiment

Two sessions. One frozen handoff.

01 / Pack

Give assistant A the source packet. Save its 600-character handoff exactly as returned, before showing it the questions.

Download source packet
02 / Move

Ask A the questions while it still has the source. Then give a fresh assistant session only the frozen handoff and those same questions. Save both JSON answers.

Download questions
03 / Unpack

Compare the answers below. Keep the source and answer key away from assistant B. Using the same model in both sessions helps isolate the handoff.

Read the full protocol
The 600-character limit is part of this fictional challenge. It is not a limit attributed to any particular assistant. This page does not connect to, run or migrate an AI for you.
Your local comparison

Check what arrived.

0 / 600 characters
No runs collected. Paste a result or inspect the invented example.
Answer format
Source history & answer key — for the operator
Opening this does not run an experiment. Do not give the answer key to the assistant being checked.


The useful story is what changed.
One wrong date is more informative than a vague claim that an assistant “forgot everything.” A useful contribution would pair a frozen handoff with before-and-after answers, then explain the first decision that changed.
The goal is not an identical response. It is preservation of six task-relevant decisions. Even the same model varies between runs; a cross-model result tests the whole handoff, not migration alone. Comparative reliability claims require repeated runs and per-decision preservation rates.
This is an open fixture with an inspectable answer key. Results are operator supplied, not independently verified. A failed answer can reflect packing, interpretation or reasoning. One run cannot rank models or establish a migration defect.
Why investigate this: OpenClaw documents migration state kept for manual review; Hermes documents conflicts, skipped items and overflow reporting. This exercise is our proposal, not a finding that either product failed.


Contribute a run
What happened after the move?
Run the exercise, download your local report, and tell us the first consequential decision that changed—or that everything arrived. Include the assistant names and versions, whether both sessions used the same model, and any human edits.
Report a Moving Day result   Inspect the source
Opening the contribution form leaves this site for GitHub. Posting is voluntary and public. Do not include private project history, credentials or customer information; this fixture is fictional and sufficient for the first round.