# Research claim ledger

Original editorial worksheet by Use AI for X. Proposed evaluation, not performed: all observation fields are blank. This is a way to inspect evidence, not a tool ranking or a completed research answer.

Start from https://useaiforx.com/for/research-with-citations/#example . The fixed question is: “What does published research say about open-plan offices and face-to-face communication at work?”

## Prepare two separate rounds

1. In **discovery**, ask exactly that question in each eligible tool. Record the account plan, actual mode, date and original answer file before follow-ups. Standard search and a paid research mode are different conditions.
2. In **fixed_sources**, use the same five legally accessible documents in each tool that accepts them. Select only those documents and exclude web searches or newly imported material from this round. Gemini Notebook also documents an experimental agentic mode that can search the web; if your account mode cannot stay inside the selected set, record this round as unavailable. Name that set in `source_set_id` and keep a separate document list with URL, title, author/year and retrieval date. Do not invent five studies or treat inaccessible full text as read.
3. Keep the two rounds separate. The CSV offers three claim slots per tool per round, twelve in all; delete an unused round or copy rows for another tool. Tool A and Tool B are placeholders, not selected products.

## Inspect one claim at a time

Save the initial answer unchanged. Copy the exact claim and original URL, then read the source yourself. Record the supporting passage and page/section, study population, setting and time period. Keep quotations short and retain the source link.

Use `access_status`: full_text, abstract_only, not_accessible or not_found. Use `reviewer_decision`: supported, partly_supported, contradicted, insufficient or conflicting_sources. A missing passage is insufficient even when its page opens successfully. An inaccessible source cannot establish support. For two sources that conflict, duplicate the claim ID into a second row, record each passage, and explain the conflict; do not average the values or pick the one you prefer.

Write a reason and a corrected claim that retains the study conditions. Keep the reviewer and date attached. A correlation in one workplace must not become a universal causal claim. Do not fill in an observation until you have actually made it.

Google's citation instructions let readers open a quoted passage in context; this helps locate evidence but does not decide whether it supports your wording. Source: https://support.google.com/gemininotebook/answer/16179559?hl=en (checked 2026-09-27; citation navigation only).

## Separate evidence review from export access

Check the selected tool's profile and the plan currently available in your account before starting. The local CSV needs no paid export. If a plan does not include your requested RIS, BIB or CSV export, record that restriction and keep your manual ledger. A manual copy is not a successful structured export. Do not infer paid research access from a free account.

## Acceptance and handover

Each selected claim needs an exact source location, access status, reason and reviewer decision. Passing means you completed the review faithfully, not that every claim was true. Remove or qualify unsupported wording in a separate corrected answer. Retain unresolved conflicts explicitly. Stop if a decision would depend on unread evidence.

Hand over the initial answer, this completed ledger, the source-set list, the corrected answer and any export actually produced on the recorded plan. Record corrections separately from first-answer quality. No pass percentage or source count substitutes for reading the decisive claims.

When opening the CSV, import columns as text. Source quotations may begin with spreadsheet formula characters; keep them as text when adding them. No completed file is sent back to Use AI for X.