At noon, one assistant has produced a promising reading list for your review. By evening, you want another assistant to challenge the synthesis. The first chat contains changed search terms, duplicate papers, and several confident summaries you have not checked. Passing all of it forward would make the second model inherit noise as if it were evidence.
The handoff should begin only after you inspect the sources.
Give each stage a different question
Separate jobs make weak evidence easier to spot:
- Literature search: propose search terms, databases, date bounds, and candidate sources.
- Evidence synthesis: compare the verified sources against a stated question and method.
- Critique: look for selection bias, unsupported causal language, missing counterevidence, and uncertainty that the synthesis compressed.
An assistant can help with each task. It cannot establish that a paper exists, that a quotation is accurate, or that a result supports your claim. Open the original source and check those facts yourself. This workflow describes how to transfer reviewed work; RelayLink makes no claim about the quality or truth of the research.
Make the evidence trail portable
A useful research briefing names:
- the research question and scope;
- where and when you searched;
- the exact query or selection logic;
- verified citations with stable identifiers or URLs;
- inclusion and exclusion decisions;
- the claim each source actually supports;
- conflicting results, limitations, and unresolved uncertainty;
- the critique you want from the next assistant.
Quote only text you checked, with a page, section, table, or figure reference where available. Mark a paraphrase as a paraphrase. If an assistant supplied a citation you could not locate, do not preserve it as a source. Preserve it as an unresolved lead, or remove it.
RelayLink carries text, not PDFs, reference libraries, lab notebooks, datasets, or browser tabs. A DOI or repository path is a reference, not access. The destination assistant needs its own authorized access to every source or dataset it is expected to evaluate.
Hand the reviewed synthesis to your other assistant
Multiple assistants can connect to one RelayLink account. Address the briefing to your own email, then draft, review, and confirm it. No model sends autonomously. Read the package for invented citations, overstated certainty, sensitive participant details, and claims that shifted while being summarized.
A self-addressed package creates no contact request, no email or magic-link page, and no second account. It remains unread until an assistant on your account pulls it; the portal inbox is available when you want to inspect it yourself. The self-handoff guide covers that loop.
This is not context synchronization. The second assistant receives only the human-approved briefing, not the first model’s hidden state or full conversation.
Use disagreement to find the next check
If the critic rejects the synthesis, ask what evidence would distinguish the two readings. One model may have weighted recency; another may have weighted sample size or study design. Neither preference is proof.
Return to the original materials, document the resolution, and preserve remaining uncertainty. When two assistants disagree, the useful output is a sharper verification task, not a vote.
When not to use this workflow
Do not use an unapproved provider for confidential data, identifiable participant information, unpublished manuscripts, restricted databases, or embargoed findings. Institutional, funder, publisher, and data-use rules still govern each assistant. RelayLink approval does not override them.
Do not treat this workflow as medical, legal, or scientific advice, and do not use model agreement to make a safety-critical conclusion. Where the method requires reproducible analysis, specialist review, or a controlled environment, keep those controls and use the briefing only as a record of what a qualified person has checked.