Paper Pipeline

The paper pipeline is a structured workflow that takes raw experiment results and produces publication-ready manuscripts. Every step is assisted by agents, reviewed across models, and tracked with full version history.

Pipeline Stages

Evidence Collection

As experiments complete, their results are tagged for inclusion in specific paper sections. The system maintains a mapping from experiment IDs to paper sections.

Claims Lock

Before drafting begins, the Paper Lead produces a claims table — every scientific assertion the paper will make, linked to its supporting evidence. No claim can exist without an experiment or derivation backing it.

Outline

An outline is generated from the claims table. Each section gets a scope: what claims it contains, what figures it needs, and what prior work it must cite.

Drafting

Agents draft sections in parallel. The Research Lead writes methods and results. The Paper Lead handles introduction and discussion. Workers generate tables and format equations.

Cross-Model Review

The full manuscript is reviewed by AI models from at least two different providers. Reviews are structured: accuracy, completeness, clarity, overclaiming, missing citations.

Revision

Based on review feedback, agents revise the manuscript. Each revision round produces a diff showing what changed and why. The revision tracker logs every round.

Compilation

LaTeX source is compiled to PDF using revtex4-2. The system verifies: zero undefined references, all figures embedded, correct page count, proper bibliography.

Submission Package

The final package includes: PDF, LaTeX source, figures, bibliography, supplementary data, and a cover letter. Ready for arXiv upload.

LaTeX Compilation

All papers use revtex4-2 (Physical Review D format):

\documentclass[aps,prd,twocolumn,superscriptaddress,
  showpacs,preprintnumbers,nofootinbib,
  longbibliography,floatfix]{revtex4-2}

Compilation runs on GPU pods with texlive-publishers installed:

# Compile (run twice for cross-references)
pdflatex -interaction=nonstopmode paper.tex
pdflatex -interaction=nonstopmode paper.tex

Warning: Common pitfalls: Do not use aastex631. Do not use \citep/\citet (use \cite{}). Do not use deluxetable (use ruledtabular). Figures must be in the same directory as the .tex file.

Figure Management

Figures are auto-generated from experiment outputs and linked to papers:

# List figures available for a paper
hubify paper figures paper-1

# Add a figure
hubify paper add-figure paper-1 \
  --experiment EXP-054 \
  --file posterior_w0wa.png \
  --caption "Joint posterior distribution of w0 and wa"

# Regenerate all figures
hubify paper regenerate-figures paper-1

Version History

Every change is versioned:

hubify paper history paper-1

v2.2.0  2026-03-15  "Focused version, 24 pages, 63 references"
v2.1.0  2026-02-28  "Added quintom MCMC results"
v2.0.0  2026-02-01  "Major restructure: bounce-model agnostic"
v1.5.0  2026-01-15  "Added anomaly sweep results"
...

Readiness Dashboard

The paper readiness dashboard tracks progress across six dimensions:

Dimension	What It Measures
Content	Percentage of sections drafted
Figures	Figures generated vs required
Bibliography	Citations resolved vs referenced
Claims	Claims verified vs total
Compilation	LaTeX errors (target: 0)
Review	Review rounds completed

CLI

# Create a paper
hubify paper create --title "Bounce Cosmology Constraints" --template prd

# Compile to PDF
hubify paper compile paper-1

# Run peer review
hubify paper review paper-1

# Package for arXiv
hubify paper package paper-1 --output submission.tar.gz