Skip to content

๐Ÿ”ฌ Deep research: what the genre owes its reader

Assumes the core, ยง1โ€“9 of WRITING-STANDARDS.md. This file holds the writing rules that apply to a multi-perspective deep-research report alone: what the genre is for, why it uses no figures, how its tables and mathematics survive being written in parallel, and what its reread looks for. The element rules are in figures.md, mathematics.md and tables.md; this file repeats none of them.

๐ŸŽฏ What a deep-research report owes its reader

A deep-research report reconciles several perspectives on one question and says where the corpus disagrees with itself. Its reader wants the disagreement, not a smoothed consensus: a contradiction map, the evidence ranked by how well the corpus supports it, and a blind spot named as plainly as a finding. A single-pass literature survey organises a field; this genre interrogates one question from several sides, and a report that reads as one voice has usually lost what it was for.

Two rules apply with particular force, because the report is written by parallel section writers who each see one slice:

  • Terminology is fixed at outline time, not at polish time. Every recurring term has one definition, in one glossary, which every section uses as written. Four writers who each named the same concept differently produce a reconciliation problem that the outline could have prevented (WRITING-STANDARDS.md ยง2).
  • Scope is stated in the report, not held in the author's head. The lead says what the report covers and what it doesn't, including which sub-questions the corpus could not answer (WRITING-STANDARDS.md ยง3 and ยง7).

A deep-research report's unit for multi-source synthesis is the paragraph (WRITING-STANDARDS.md ยง11): a body paragraph closes on two or more citekeys wherever the evidence allows, organised around what its sources agree and disagree about.

๐Ÿ“ No figures, by design

Every other genre calibrates how freely it uses figures.md ยง10; this one uses none. If a section genuinely needs a visual, say so in the report rather than adding one.

๐Ÿ“Š Tables carry this genre's shape

tables.md ยง13 applies, and matters here more than elsewhere: a contradiction map or a per-perspective comparison is this genre's natural shape. Because sections are written in parallel, two writers can pick the same table id: two sections that both wrote <!-- table: comparison --> produce one broken cross-reference in the assembled report. Ids are therefore checked for collisions when the sections are joined, not per section.

๐Ÿ”ข Mathematics: carried over, reconciled once

mathematics.md ยง12 applies, unlike ยง10. A quantity carried over from a source (a threshold, an effect size, a bound) goes in a code span as ASCII with a row in the dossier's math.md, so the pdf sets it as mathematics and the .md stays readable. Each writer sees only its own slice, so the rows are collected into one math.md when the sections are joined, ## Symbols meanings included, and a symbol two sections spelled differently is reconciled to one spelling rather than kept twice.

Number sparingly, and only when the sections are joined. ยง12's numbering rule (standalone, the final step of a derivation, or reused elsewhere) rarely fires within one writer's slice. It is more likely to fire across sections, when two writers independently reported the same bound and a later section reuses it. The equation's marker and its reference are added then, and its id is checked for collisions the same way as the table ids: two writers working in parallel can pick the same short id for unrelated results.

๐Ÿ‘“ The reread

Read the assembled report once as its reader before presenting it (WRITING-STANDARDS.md ยง6): a pass over the whole document, which no individual section writer was in a position to make. Parallel writers leave specific, predictable seams; check for:

  • the same concept named two ways, or one name used for two concepts;
  • a term defined independently in two sections;
  • notation that shifts between sections;
  • the same finding stated at different strengths in two places;
  • a claim one section assumes but only a later section establishes.