Write

RESEARCH.md

[UNGROUNDED]

Prerequisites

1. Read

   RESEARCH.md

   . Apply supervision policy for

   write-start

   (approve → pause; notify → announce and continue; absent → proceed). If writing requires a compromise relative to the current plan, apply

   Idea Changes

   before proceeding.
2. Check Context for experiment results.
   • Results exist → use as empirical basis.
   • No results → ask: proceed with a results-free draft, or wait for experiments?

Narrative Framing

1. Before outlining, state the paper's contribution in one sentence. If you cannot, the framing has not converged — ask the user to clarify the intended story.
2. Build a compact claim-evidence map:
   • Claim — what the paper asserts.
   • Evidence — result, theorem, artifact, or citation supporting it.
   • Section — where the claim will appear.
   • Missing evidence → mark

     [UNGROUNDED]

     and keep drafting only if the user approved a results-free draft.

Drafting Order

1. Follow this sequence strictly:
   1. Outline — section structure with bullet points for each section's argument. The outline should make one coherent story, not a list of experiments.
   2. Section drafts — write each section, starting from the one with the strongest evidence.
   3. Self-consistency check — verify that claims in the abstract match claims in the body, numbers are consistent across sections, and the conclusion follows from the results.
   4. Citation pass — see below.

Grounding Rule

1. Every quantitative claim, dataset claim, baseline claim, theorem claim, and method-detail claim must trace to a specific source in

   RESEARCH.md

   Context or a verified citation. If a claim cannot be traced, mark it inline:

   [UNGROUNDED]

   . Leave the marker visible for the user — do not silently remove it.

Style Pass

1. After drafting, polish for publication-ready academic prose:
   • Keep subject and verb close.
   • Use concrete verbs and consistent terminology.
   • Replace vague nouns like "performance" with the actual metric when known.
   • Remove filler and inflated language: "groundbreaking", "revolutionary", "delve", "pivotal", "landscape", "notably", "importantly", "it is worth noting".
   • Avoid generic conclusions such as "opens new avenues" unless the avenues are specific.
   • Run a reverse-outline test: read the first sentence of every paragraph in sequence. They should form a coherent argument on their own.

Citation Pass

1. For every reference: search arXiv, resolve the DOI, or use another trusted source. If unverifiable, mark

   [UNVERIFIED]

   . Never generate citations from memory. Fabricated citations are academic misconduct.

Output

1. Default output:

   paper.md

   and

   refs.bib

   at the project root. Escape hatch: if a

   .tex

   file already exists at the project root, ask whether to write there instead.
2. Before declaring done, check:
   • Abstract is self-contained.
   • Introduction claims have supporting evidence in the body.
   • Numbers are consistent across sections.
   • Terminology is consistent.

   • [UNGROUNDED]

     and

     [UNVERIFIED]

     markers are listed for the user.

Post-Draft

1. Auto-invoke

   review

   . Write the score to

   RESEARCH.md

   Context. Apply stop policy (

   target_reached

   ). In wild mode, continue if configured; otherwise ask: "Review score is X/10. Want a revision pass?"

Partial State

1. If

   paper.md

   already exists, ask whether to continue from it or start fresh.

Constraints

• Never fabricate results, citations, baselines, or venue claims.
• Never hide

  [UNGROUNDED]

  or

  [UNVERIFIED]

  markers.
• Do not turn

  write

  into compilation, figure generation, or experiment execution; delegate those to the relevant skill.

Example

review