end-to-end — a Lancet Digital Health Viewpoint submission and its substrate

A single-operator agentic-research demonstration, packaged as a submission to The Lancet Digital Health (article type: Viewpoint, 2 500 words).

The submitted Viewpoint argues that the new minimum disclosure unit for agentic-LLM-assisted research is prompt + commit hash + tagged release, not the classical methods-section narrative. The argument is operationalised by this repository: every artefact the proposed disclosure standard would require is published here, with nothing redacted and no post-hoc curation. A reviewer who accepts the Viewpoint endorses the methodology; a reviewer who rejects it must do so on grounds the repository itself disproves.

The medium is the message.

Layout

manuscript/             The Lancet Digital Health Viewpoint (primary deliverable)
  main.tex                 LaTeX source, Vancouver superscript-numeric style
  references.bib           up to 30 references
  JOURNAL.md               distilled author instructions for Lancet DH
  cover-letter.md          submission cover letter (drafted in Phase 8)
  Makefile, latexmkrc      build helpers
  figures/                 generated by case-study analysis, not by AI

case-study/             Supporting empirical artefact (Layer 1 pipeline output)
  analysis/                01..05 numbered scripts (HCC OS stratification)
  data/raw/                gitignored, regenerable from documented URLs
  data/processed/          gitignored, regenerable
  data/results/            small analytic artefacts kept in-repo
  figures/                 figure PDFs
  manuscript/              the Layer-1-produced case-study draft manuscript

reviewer-logs/          Verbatim transcripts of every reviewer-subagent round
  round-01/.. round-NN/    one per round, four reviewers each
  audit/                   Layer-2 audit transcript and findings.md

prompts/                The protocol artefacts (the Viewpoint's key claim)
  00-original-spec.md      operator-supplied orchestration spec (verbatim)
  01-layer1-pipeline.md    Layer 1 autonomous pipeline kickoff prompt
  02-layer2-audit.md       Layer 2 audit subagent prompt
  03-reviewer-*.md         four reviewer persona prompts

docs/                   The integrity anchors
  design.md                three-layer architecture, honesty contract
  prereg.md                Layer-3 external-validation preregistration
  ai-usage-disclosure.md   exhaustive AI-tool / prompt / task disclosure
  ledger.md                chronological artefact ledger (Figure 1 source)

.github/workflows/      CI that rebuilds the manuscript on each v* tag
.gitignore, LICENSE, pyproject.toml

What to read in what order

1. manuscript/main.pdf — the Viewpoint, 2 500 words. Compile with cd manuscript && make.
2. docs/ai-usage-disclosure.md — every AI tool and every operator intervention, itemised.
3. prompts/00-original-spec.md — the operator-supplied orchestration prompt the Viewpoint defends.
4. case-study/manuscript/main.pdf — Layer 1's autonomous output (the "what the pipeline actually produced" artefact).
5. reviewer-logs/round-01..NN/ — the reviewer-loop transcripts.
6. reviewer-logs/audit/findings.md — Layer-2 audit findings.
7. docs/ledger.md — chronological artefact ledger (Figure 1 source).

Reproduce

The case-study analysis is reproducible from a clean clone:

uv sync
cd case-study
uv run python analysis/01_prepare_data.py
uv run python analysis/02_layer1_risk_score.py
uv run python analysis/03_layer3_external_validation.py
uv run python analysis/04_figures.py
cd ../manuscript && make   # build Viewpoint PDF

LLMs are not bit-deterministic; the reproducibility claim is distributional (substantive findings reproduce, not character-level outputs). See prompts/README.md.

Author

Hsieh-Ting Lin, M.D., Department of Hematology and Medical Oncology, Koo Foundation Sun Yat-Sen Cancer Center, Taipei, Taiwan. ORCID: 0009-0002-3974-4528.

Licence

MIT. See LICENSE.

Tags

• viewpoint-v<x>.<y>.<z> — Viewpoint manuscript releases.
• case-study-v<x>.<y>.<z> — Case-study manuscript releases.