Research context — what has been established, and what has been ruled out

This sits at the root of a research project and holds what an agent cannot work out by looking: what the project is actually for, what counts as evidence here, and what has already been tried and closed. The working agreement covers where the work stops; this covers what it is working from.

The question: {{question}} What counts as evidence: {{evidence_bar}} Dead ends are recorded in: {{dead_ends_path}}

The question behind the question

The brief is frequently a proxy. Which competitors raised prices may really be are we about to be undercut, and those have different answers and different evidence.

State the real one, because an agent optimising the stated proxy competently is the most expensive failure available here — it produces defensible work that answers the wrong thing, and nothing in the output signals it.

What counts as evidence, which is not a general question

A clinical reviewer and a market analyst both say “evidence” and mean incompatible things. One will not accept a preprint; the other will not wait for peer review and is right not to.

This cannot be inferred from the material, and getting it wrong is not recoverable by working harder — a body of sources gathered to the wrong standard has to be gathered again.

Dead ends — the section this document exists for

A dead end is an avenue that closed: the database with nothing in it, the search term that returned noise, the promising lead that turned out to be about something else, the person who did not reply.

It is the most expensive thing in a research project to rediscover and the first thing lost, because it lives in one person’s head and produces no artefact. A session starting cold walks straight back into it, because nobody told it not to — and it will do so competently, which is what makes the loss invisible.

Three things per entry, in {{dead_ends_path}}:

• what was tried — specific enough to not be repeated by accident
• why it closed — the mechanism, not the mood
• what would reopen it — the condition under which it is worth another go

That third line is what separates a map from a list of failures. Without it the record becomes a wall of no, which people stop adding to and then stop reading.

Write it as it happens. At the end of a project it is already gone; the avenues that closed in week two are not recoverable in week nine, and the ones that felt too trivial to record are exactly the ones that get repeated.

How firmly each thing is held

Research accumulates three kinds of claim and they are not interchangeable:

• Established — there is evidence, and it meets {{evidence_bar}}.
• Argued — somebody reasoned it out and wrote the reasoning down. Weaker than established, and it must never be reported as though it were.
• Assumed — nobody has checked, and the work proceeds as if.

Mark which. The distinction erodes on its own: after a few months we reasoned about this carefully and we verified this compress into the same sentence, and the assumption that was load-bearing becomes a finding nobody re-examines.

Who has already been asked

Approaches to people are not repeatable. An agent that emails a source who declined last month, or re-asks a question already answered in correspondence, spends something that cannot be got back.

Record who has been contacted, what they said, and whether they may be approached again.

Where this lives

Beside the sources and the notes, in the project folder — not in a separate tool where it will be written once and read never.

Without version control nothing records when this changed. Date it by hand when you revise it, particularly {{evidence_bar}}: a standard that quietly moved is worse than none, because the material gathered under the old one still looks current.