MetaScience, Scientific Inquiry, and Agents

Working Quarto book, v0.4. Thirteen one-hour meetings for the Denolle Lab, taken one at a time in the order below rather than on fixed dates.

Published at https://denolle-lab.github.io/metascience-and-AI/

Scientific advance is the primary outcome: warranted gains in knowledge, understanding, and research capability. We study plural geoscientific inquiry through philosophy, history, and the science of scientific process and impact, then design agents around human-defined modes and evaluate them against past work. Novelty, correctness/warrant, and impact inform the assessment; interdisciplinary integration is a candidate contributor to advance.

One anchor and one companion per meeting keep participation light. The convenor authorized the v0.4 revision; the mode definitions and evaluation instruments remain drafts for discussion.

Contents

Path	What it is
index.qmd	Rationale, schedule, standing questions
glossary.qmd	Working definitions of the modes of inquiry and of the terms the quarter measures; meeting 1 revises it
unit.qmd	The unit of inquiry: deduction, induction, abduction; the hypothetico-deductive schema and its verbs; the thinkers who fixed them; one graph on which Platt's loop and Cleland's objections are placed
agents/skills/unit-of-inquiry/SKILL.md	The unit graph as an agent skill: state, one rule per node with its source, three compositions of units, required outputs, planted-flaw tests; every meeting's page ends by saying what its papers change here
metrics.qmd	Registry of what the measurement literature found useful for novelty, impact, breadth, and advance, with data, findings, stated limitations, and the basis each entry rests on; how the club tests the metrics against its own judgments; the embedding-era revisit
tools/indicators.py	Computes citations, field percentile, the CD_5 disruption index, and the sleeping-beauty coefficient for the assigned papers from OpenAlex, into private/indicators.json
sessions/	One chapter per meeting, thirteen of them
rubrics.qmd	Inquiry, novelty, advance, and evaluation worksheets
prior-art.qmd	Registry of existing science agents and benchmarks, and what their scores do not establish
corpus-study.qmd	Design for coding the mode of inquiry of published papers and relating it to impact, novelty, interdisciplinarity, and depth
agent-design.qmd	Versioned modes, procedures, claim records, implementation status, and fair agent comparisons
bibliography.qmd	Every reading by meeting with links and access notes, the group's suggestions, and the full reference list; the reading lists are generated from curriculum.json
modes/	Authoritative human-editable mode specifications, exported for agent use
historical-evaluation.qmd	Decision reconstruction, reproduction/reassessment, and extension/transfer cases
templates/claim-record.json	Empty claim-record template; no experimental result
agents/mode-instructions.json	Generated mode instructions with source hashes; not an agent runtime
tools/export_modes.py	Regenerates and checks the instruction export
review/	Dated audits, convergence plans, and source-check evidence
notes.qmd	The group's meeting record — add yours
contributing.qmd	How to contribute, and what must never be published here
references.bib	Bibliographic records; current counts are reported by the validator
curriculum.json	Editorial metadata for sessions and references. Kept in step with the .qmd files by hand; validate.py checks that the citation keys agree
READING-AUDIT.md	Why the reading selections are what they are
PLAN-AUDIT.md	September 2026 audit of the plan against the seminar's aims, and the reasoning behind the expanded optional extensions
tools/validate.py	Source, schedule, link, and citation checks, including that the generated reading lists are current
tools/build_bibliography.py	Regenerates the reading lists in bibliography.qmd from references.bib and curriculum.json
tools/fetch_open_copies.py	Fetches legal open-access copies of the references into the gitignored private/pdfs/ and writes private/reading-copies.md
tools/check_links.py	Verifies every DOI against Crossref and fetches every other URL
tools/stage_site.py	Fallback: copies a rendered book into a subdirectory of a local Jekyll checkout. Not the current publishing path

Build locally

Install the Quarto CLI from https://quarto.org/docs/get-started/. The book has no executable code cells and needs no Jupyter kernel, R, or Julia.

quarto preview      # live reload while editing
quarto render       # writes _book/

The source checks need Python 3 and PyYAML. On a system with an externally managed Python, which covers recent macOS and most Linux distributions, use a virtual environment rather than pip install into the system interpreter:

python3 -m venv .venv
.venv/bin/pip install PyYAML
.venv/bin/python tools/export_modes.py
.venv/bin/python tools/build_bibliography.py
.venv/bin/python tools/validate.py
.venv/bin/python tools/check_links.py

.venv/ is gitignored. validate.py checks that every chapter exists, that local links and citation keys resolve, that the thirteen meetings are numbered in order, that references.bib and curriculum.json describe the same set of works, and that no private material is registered as a chapter; it then runs a Pandoc citation pass that fails on any warning. check_links.py is slower and hits the network, so run it when you have changed references rather than on every edit.

Publishing

Pushing to main triggers .github/workflows/publish.yml, which validates the sources, renders, and publishes to the gh-pages branch. Pull requests run validation and a render without publishing. There is nothing to run by hand.

tools/stage_site.py remains for the alternative of copying a rendered build into the group's Jekyll site at denolle-lab.github.io. That is not how this book is published and the script is kept only as a fallback.

Contributing

See CONTRIBUTING.md. Short version: this repository is public, so no journal PDFs, no unpublished data, nothing a collaborator has not agreed to share, and nothing from private/.

private/ is where the parallel agent analysis lives. It is gitignored, and validate.py fails if that ignore rule is removed or if private material is registered as a chapter.

Reading access and rights

Every assigned paper has an external DOI, publisher, repository, or proceedings link, and some full texts require institutional access. No journal PDFs are redistributed here. Free access is not permission to republish; check the license and preserve attribution before adding any third-party asset.

Course content is CC BY 4.0 and tools/ is MIT; see LICENSE. Rights in the cited papers remain with their holders.

Validation status

See VALIDATION.md.

Documentation

• Book structure: https://quarto.org/docs/books/book-structure.html
• Book output: https://quarto.org/docs/books/book-output.html
• GitHub Pages: https://quarto.org/docs/publishing/github-pages.html
• Citations: https://quarto.org/docs/authoring/citations.html