Decomposition, barriers, and citation without re-derivation

How do we define and fill systemic gaps — what standard research method already knows about building an argument, and how the machine mirrors it exactly.

§Three moves every serious piece of analysis makes

Decompose before you synthesise. A hard question is rarely answered directly; it is broken into narrower sub-questions that can each be settled on their own evidence, and only combined once every one of them has an answer. 2.1's four parallel vectors are this move, mechanised: four independent lines, four independent Cells, four independent findings — and a fifth step, the report, that is structurally incapable of firing until all four exist, because its listener's condition names all four by type.

Do not write the synthesis early. This is the barrier, and it is worth stating as its own discipline because it is so easy to violate informally — a person drafting a report by hand will often start the conclusion before every input has actually arrived, then quietly patch it later. A listener that demands several things at once cannot do this even if asked to; there is no partial firing, no draft state visible to anything downstream. The report either does not exist yet, or it exists having genuinely seen everything it was supposed to.

Cite a prior finding; never silently re-derive it. Good analytical writing distinguishes sharply between reaching a new conclusion and reusing an established one — a citation names its source and carries the finding forward unchanged, rather than restating it from memory and risking drift. This is exactly what 2.1's "carrying mechanism" does mechanically: a finding that re-enters a later step arrives stamped with exactly where it came from, and its content is never retyped by a person or reconstructed by a machine — it is the same value, moved.

§Why this matters more, not less, once a language model is involved

A language model will, if asked, happily restate an earlier finding from its own memory of the conversation — fluently, and not always accurately. The discipline above closes that gap structurally: nothing in this system ever asks a model to recall an earlier finding; every reuse of a finding is a mechanical carry of the actual committed value, and a model is only ever asked to produce something new from material it was explicitly handed. The distinction between "the model derived this" and "the model was shown this, unchanged, from elsewhere" is never blurred, because the machine has no path by which it could be.

§Provenance is not just where a fact came from — it is also how it got there

Ordinary provenance answers "who said this, and when." The mechanism this part is building towards answers a second question research method cares about just as much: by what path did a finding travel from where it was established to where it is now being used — copied verbatim, retitled and re-scoped for a new audience, or handed down into a stricter, walled sub-task with a narrower brief attached? 2.3 gives this its formal shape: answer, attach, and citation are three genuinely different kinds of carry, and the machine keeps the difference visible rather than collapsing all reuse into one undifferentiated "referenced elsewhere."

§What this does not solve

None of this checks whether a finding is actually right — only that its path from origin to reuse is unbroken and honestly labelled. Whether the finding itself deserves to be trusted is, again, the analyst's own job, taken up fully in 7.2. What this part's mechanism buys is narrower and, for exactly that reason, dependable: if a claim is wrong, you can always find out where it was produced and how it got to where you are reading it — nothing about the reuse itself can have quietly changed it along the way.