Paper Writing — Hierarchical Logic Stream (HLS)

When to Use

Core principle: Write mode builds from ideas. Polish mode extracts structure from a draft. Both use the same HLS levels — only the starting point differs.

Overview

4-level refinement. Each level validates before descending. No skipping. No merging.

Venue-driven discovery: Target venue determined at L0. All resources (skeleton, blueprint, writing guide) auto-discovered by exploring templates/ — never hardcoded.

Every level commits to git.

Universal Interaction Protocol

All L0/L1 confirmations use clickable options — never free-text. The user clicks, does not type.

Every prompt to the user MUST end with a set of clickable options. L0 and L1 each specify which options apply at which step.

Polish Mode

When polishing an existing draft, HLS documents MUST be generated — not just discussed verbally.

1. If no L0/L1 exists — read draft → extract implicit HLS → critical think silently (identify issues, form 2-3 suggestions) → present to user → discuss → write docs/systematic-research/plans/stream-L0.md and stream-L1.md → commit
2. If L0/L1 exist — critical review: points still valid? chains match draft? → critical think (2-3 issues + suggestions) → present → discuss → update L0/L1
3. Invoke L2 — parallel revision following (revised) flow chains + writing guide
4. Invoke L3 — consistency check (7 structural + prose checks). Polish complete when all checks pass.

Polish-lite (language only): User says "just polish the language" → use writing guide for voice/pitfalls → L2 revise prose → L3 consistency. Do NOT restructure.

Write Mode (Full HLS)

Standard flow: L0 → L1 → L2 → L3. Start from ideas, end with paper.

Hard Gates

Process Flow

digraph hls {
    rankdir=TB;
    "Start" [shape=oval];
    "L0: Judge + Venue\n[l0-core-idea.md]" [shape=box];
    "User OK?" [shape=diamond];
    "Commit L0" [shape=box];
    "L1: Architect\n[l1-chapter-structure.md]" [shape=box];
    "User OK?" [shape=diamond];
    "Commit L1" [shape=box];
    "L2: Build\n[l2-parallel-writing.md]" [shape=box];
    "User OK?" [shape=diamond];
    "Commit L2" [shape=box];
    "L3: Audit\n[l3-consistency-check.md]" [shape=box];
    "7 checks pass?" [shape=diamond];
    "Commit final" [shape=box];
    "Done" [shape=oval];

    "Start" -> "L0: Judge + Venue\n[l0-core-idea.md]";
    "L0: Judge + Venue\n[l0-core-idea.md]" -> "User OK?"; "User OK?" -> "L0: Judge + Venue\n[l0-core-idea.md]" [label="no"];
    "User OK?" -> "Commit L0" [label="yes"]; "Commit L0" -> "L1: Architect\n[l1-chapter-structure.md]";
    "L1: Architect\n[l1-chapter-structure.md]" -> "User OK?"; "User OK?" -> "L1: Architect\n[l1-chapter-structure.md]" [label="no"];
    "User OK?" -> "Commit L1" [label="yes"]; "Commit L1" -> "L2: Build\n[l2-parallel-writing.md]";
    "L2: Build\n[l2-parallel-writing.md]" -> "User OK?"; "User OK?" -> "L2: Build\n[l2-parallel-writing.md]" [label="no"];
    "User OK?" -> "Commit L2" [label="yes"]; "Commit L2" -> "L3: Audit\n[l3-consistency-check.md]";
    "L3: Audit\n[l3-consistency-check.md]" -> "7 checks pass?"; "7 checks pass?" -> "L2: Build\n[l2-parallel-writing.md]" [label="no"];
    "7 checks pass?" -> "Commit final" [label="yes"]; "Commit final" -> "Done";
}

Quick Reference

Git

git commit -m "L0: core idea for <topic>"   # docs/systematic-research/plans/stream-L0.md
git commit -m "L1: structure for <topic>"   # docs/systematic-research/plans/stream-L1.md
git commit -m "L2: draft for <topic>"       # paper/
git commit -m "L3: final for <topic>"       # paper/

Red Flags

• "The core idea is clear enough, let's start writing" → return to L0
• "We can figure out the details while writing" → return to L0
• "This point is obvious, no need to write it down" → return to L0
• "We can't write anything until experiments are finished" → return to L0 (draft mode exists)
• "We can just make up placeholder numbers without marking them" → stop
• "One commit at the end is fine" → return to current level
• "I'll check consistency after the paper is accepted" → return to L3
• "We need three challenges to match three designs" → return to L0 (challenges are determined by evidence, not symmetry)
• "This challenge is related to the root cause somehow" → return to L0 (causal link must be direct, strong, and testable)

Anti-Patterns

• Writing before clarifying → L0 must be committed before any prose
• Soft judgement → use rejection criteria mechanically
• Editing skeleton in-place → copy to paper/ first; templates are read-only
• Big bang commit → four commits, one per level
• Hardcoding paths → discover from venue, never assume skeleton location