How the notebook is made
The desk's research workflow
Every note this journal publishes is made the same way, and the way is published too — because a workflow you cannot inspect is a promise, and this desk prefers machinery you can open. Four stages, each with a written rule about what the AI layer is allowed to touch. None of it requires paid tools or professional data feeds: the workflow runs on public documents and consumer software, exactly what an individual investor has on a desk at home.
What follows is the operational version of the sequence you met on the front page, expanded for readers who want to borrow it whole.
The four stages, as the desk runs them
Sourcing — decide what counts before anything else
The desk starts by listing the primary material a question touches: the filings and transcripts a listed company publishes itself, the regulator’s pages, the investor-relations releases. A general-purpose model helps index the pile — locating documents, suggesting what a subject’s likely primaries are — but the rule is fixed: AI may locate the sources; it may never be the source. Anything that cannot name its primary home does not enter the notebook.
Extraction — passages and quotes, marked with their pages
This is the stage language models are genuinely built for. The desk asks for exact passages, table rows and quoted sentences, each tagged with the document and page it came from. Paraphrase is allowed only as a second layer beside the quote, never instead of it. If the model cannot point to where it read a thing, the thing is treated as unfiled.
Verification — the human stage, and the slow one
Every number, date and load-bearing comparison is reopened against the original document by a person. The desk asks the four questions in the verification-debt note: where would this live, does the period match, do the units survive, does the wording survive. Confidence from the model is never evidence; only the opened page counts.
Filing — notes built to be audited later
What survives verification is filed in the three-layer format the filing note describes: verbatim quote, plain fact line, dated interpretation. AI drafts structure; the desk signs facts. Every published article is drawn from these filed fragments, so any sentence in it can be re-traced without redoing the reading.
The stages are boring on purpose. Boring survives contact with new tools, market weather and the nightly instinct to skip a step — and the sequence matters more than any software that runs inside it.
Objections the desk answers often
Do you use AI to pick stocks or time the market?
No. The journal is a publication about research workflows, not an investment strategy. AI tools appear in these pages as reading aids — indexers, extractors, paraphrasers — and every published claim that leans on one is re-checked against primary documents by a person before it goes out.
Is any of this personalised investment advice?
No, in either direction: regulated investment advice is out of scope, and so is the looser idea of “just between us, what would you buy?”. The desk answers questions about method — how a claim was checked, how a stage works, what a tool can realistically do to a document. It does not answer questions about your money.
Can I pay the desk for a deeper analysis of a company?
No paid service exists on this site, and none hides behind it: no subscriptions, no commissioned reports, no paid consultations. If a subject keeps arriving in the question queue, it becomes a public note that everyone — including whoever asked — reads at no cost.
Which AI tools does the desk actually use?
The journal deliberately names no products and takes no placements. Articles work in tool categories instead — general-purpose language models for passage work, document readers for long PDFs, spreadsheet utilities for table arithmetic. The stages matter more than brand names, because tools change faster than the habit of checking their output.