Choose one of three approaches. The steps for that option appear below.
Choose an option above. On a phone, swipe the tabs left or right to see all options →
Each material assumption has a named source, an uncertainty and a falsification question. The research lead signs off.
Assemble the source set
Draft the ledger
Check rows against originals
Draft the register
Lead signs off
Use: You
Assemble the recommendation draft, model inputs and supporting sources as one source set; record scope, permission limits and source IDs.
Check every file opens and is lawfully accessible or owner-authorised.
Use: Claude
Load the whole source set and ask for the evidence ledger in one pass.
What to paste into the tool
Using every supplied source, build a ledger. Evidence ledger rows with source_id, source_date when known, passage_or_cell_location, extracted_value_or_quote, evidence_status, permission_note and unresolved_question; original units and time zones kept. Use only the supplied material. Quote exactly and give the source ID and location for every line. Separate what a source says, what is inference and what is not known yet. Do not send, publish, buy, book, change any account or run code.
Use: You
Check every ledger row against the originals and mark rows that fail.
Compare every extracted passage, date, identifier and numeric field with the original; flag missing or unreadable inputs.
Use: Claude
Ask for the assumption register: separate observed inputs, estimates and value judgements and trace downstream dependence.
What to paste into the tool
From the checked ledger only, draft the assumption register. Label each item as observed input, estimate or value judgment. For each material assumption give provenance, uncertainty, a falsification question and the consequence if it is wrong. Use only the supplied material. Quote exactly and give the source ID and location for every line. Separate what a source says, what is inference and what is not known yet. Do not send, publish, buy, book, change any account or run code.
Use: The research lead
Sign off or return the draft, keeping evidence status gaps visible. The research lead checks materiality and unsupported certainty.
Each material assumption has a named source, an uncertainty and a falsification question.
Each material assumption has a named source, an uncertainty and a falsification question. Only the checked ledger reaches the assistant.
Fill the ledger by hand
Second reader checks
Draft the register
Trace lines, finalise
Use: You
Read the originals and fill the evidence ledger by hand in a spreadsheet; no AI tool sees the raw sources.
Compare every extracted passage, date, identifier and numeric field with the original; confirm scope, source access and permission limits.
Use: The second reader
Check the ledger against the originals and return disputed rows; confirm each quote, date and number matches the source; list anything unreadable.
Use: Gemini Notebook
Give only the checked ledger; ask for the assumption register citing ledger row IDs.
What to paste into the tool
You have only the ledger below, not the sources. Draft the assumption register citing ledger row IDs. Label each item as observed input, estimate or value judgment. For each material assumption give provenance, uncertainty, a falsification question and the consequence if it is wrong. For anything the ledger does not support, write a question and do not fill it in. Use only the supplied material. Quote exactly and give the source ID and location for every line. Separate what a source says, what is inference and what is not known yet. Do not send, publish, buy, book, change any account or run code.
Use: The research lead
Trace each line back through the ledger row to the original; finalise, keeping evidence status gaps visible. The lead checks materiality and unsupported certainty.
Each material assumption has a named source, an uncertainty and a falsification question.
Each material assumption has a named source, an uncertainty and a falsification question. You write the register.
Split pack per source
Source card per chat
Combine cards by hand
Draft the register by hand
Challenge the draft
Lead signs off
Use: You
Split the pack into one file per source document and register each file.
Check no file lost pages or tables in the split and that permission limits carry over to each file.
Use: ChatGPT
For each source document, open a fresh chat with only that file and ask for a source-document card.
What to paste into the tool
This chat holds one source document only. List exact passages relevant to the research question with source ID and location, and mark each as stated, not known yet or ambiguous. Use only the supplied material. Quote exactly and give the source ID and location for every line. Separate what a source says, what is inference and what is not known yet. Do not send, publish, buy, book, change any account or run code.
Use: You
Combine the source-document cards into the ledger by hand and note conflicts between sources.
Compare every extracted passage, date, identifier and numeric field with the original; flag missing or unreadable inputs.
Use: The research lead
Draft the assumption register by hand from the ledger: separate observed inputs, estimates and value judgements and trace downstream dependence.
Use: ChatGPT
Challenge the human draft: list lines with no ledger support and requirements or criteria it does not cover.
What to paste into the tool
Review the draft against the ledger. List every line without a ledger row, every missing item, and anything stated more strongly than the ledger allows. Do not rewrite the draft. Label each item as observed input, estimate or value judgment. For each material assumption give provenance, uncertainty, a falsification question and the consequence if it is wrong. Use only the supplied material. Quote exactly and give the source ID and location for every line. Separate what a source says, what is inference and what is not known yet. Do not send, publish, buy, book, change any account or run code.
Use: The research lead
Act on the challenge list, then sign off or return the register, keeping evidence status gaps visible. The lead checks materiality and unsupported certainty.
Each material assumption has a named source, an uncertainty and a falsification question.
Best when: One assistant drafts both from one source set
What it does: Claude drafts ledger and register; people check
Strong at: A single trail from source to draft
Limit: Uploads capped at 20 files per chat, 1000 pages per PDF
Free plan: Yes, limited (session usage cap resets every 5 hours)
Best when: Raw sources stay with people
What it does: Gemini Notebook works from the checked ledger only
Strong at: Each draft claim traces to a ledger row
Limit: Anything missed in the ledger is unrecoverable
Free plan: Yes, limited (50 sources per notebook)
Best when: Distinct sources, human authorship wanted
What it does: ChatGPT reads per source and challenges; people write
Strong at: Authorship and judgement stay with the reviewer
Limit: The card-join step carries the join risk
Free plan: Yes
Before you paste into any AI tool, remove any confidential information. Put it back at the end.
Deciding between ChatGPT and Claude for this? See the comparison.
Use only material you may lawfully use or have the owner's permission to use. Treat the output as a research suggestion for your own review.
Do not let the tool send, publish, buy, book, change an account, start monitoring or run live code on its own. You keep every decision and any later approval.
Keep the source IDs and where each passage or cell came from. Keep what the source says, your own inferences and the gaps separate.
Before the brief counts as complete, confirm each material assumption has source details, an uncertainty note and a falsification question, and the evidence trail has no open items. Check nothing overstates how sure the evidence is.
Check the result against your original inputs. Approve it before sharing or acting.