Repository ORCID EN
F(AI)²R — FAIR Research with AI in the Loop, Twice DLR Zentrum für Leichtbauproduktionstechnologie (ZLP), Augsburg · Helmholtz-Gemeinschaft · NFDI4Ing · HMC
F(AI)²R

Methodology

build d18a51b8f3

Methodology

This file is the contract between the human author(s) and the LLM agents defined in agents/. It describes how the paper is built, not what it says.

Primary artifacts (invariant)

Four artifacts are primary and must be consistent and up to date at all times:

  1. The manuscript (paper/main.tex, paper/sections/*.tex).
  2. The PROV-O graph (doc/provenance.ttl).
  3. The logbook (doc/logbook.md).
  4. The slide decksslides/pitch-5min.tex (5-minute pitch) and slides/conference-30min.tex (25 + 5 conference talk), Beamer + DLR Corporate Design (slides/style/fair2r-beamer.sty). Rendered PDFs are derivatives.

A working session is not complete until all four are in sync. A commit that updates one of them and leaves the others lagging is a defect, even if the build passes. This invariant supersedes any per-agent style rule elsewhere in this repository. The slide decks are downstream of the manuscript: a claim that appears on a slide must already appear in the paper text.

Pipeline overview

       human goal
            │
            ▼
   ┌────────────────┐
   │  orchestrator  │
   └───────┬────────┘
           │
   ┌───────┴────────┬──────────────┬────────────────┐
   ▼                ▼              ▼                ▼
scientific-    source-       illustration       condenser
 writer        analyzer
   │                │              │                │
   ▼                ▼              ▼                ▼
   ─────────  proposed triples  ──────────  ──────────
                       │
                       ▼
              provenance-curator ──► doc/provenance.ttl
                       │
                       ▼
              (audit pass)
                       │
        ┌──────────────┼─────────────────┐
        ▼              ▼                 ▼
  fair-aligner   layout-scrutinizer  readability-reviewer
        │              │                 │
        ▼              ▼                 ▼
                  human sign-off
                       │
                       ▼
                    git commit
                       │
                       ▼
                    logbook entry

Two passes of AI

  1. Authoring pass. scientific-writer, source-analyzer, illustration, condenser. AI proposes; humans dispose.
  2. Audit pass. fair-aligner, layout-scrutinizer, readability-reviewer, provenance-curator. AI checks the manuscript and the graph against explicit criteria.

Verification ladder

Verification ladder rendered as a finite-state machine, with each rung sized in proportion to the number of fair2r:Claim entities currently at that rung in the live graph.

A fair2r:Claim carries one of:

state meaning
unverified Asserted but no evidence yet attached.
needs-research Evidence-gap acknowledged, search planned.
ai-confirmed An AI agent verified against a vendored or fetched source.
human-confirmed A human author verified the source and the inference.
source-vendored The source itself is in the repo under doc/sources/.

A claim must reach human-confirmed (or source-vendored) before it can appear in main-condensed.tex.

The states form a finite-state machine, monotone non-decreasing modulo explicit retraction (the only legal back-edge, recorded as a prov:Invalidation activity in doc/provenance.ttl). The diagram below mirrors paper/figures/ladder-fsm.mmd, the source-of-truth artefact for the manuscript figure of the same name.

stateDiagram-v2 direction LR [*] --> unverified_external unverified_external: unverified-external needs_research: needs-research lit_retrieved: lit-retrieved ai_confirmed: ai-confirmed lit_read: lit-read unverified_external --> needs_research: recognise gap needs_research --> lit_retrieved: DOI / URL resolves lit_retrieved --> ai_confirmed: AI fetch + quoted snippet ai_confirmed --> lit_read: human reads lit_read --> unverified_external: prov:Invalidation (retraction)

Definition of done for a section

  1. Prose compiles in paper/main.tex with no \todo left.
  2. Every claim has a fair2r:Claim IRI in doc/provenance.ttl at human-confirmed or source-vendored.
  3. fair-aligner audit reports no fails for the section.
  4. Logbook entry committed.

Branch and deployment policy

main is the default branch. All routine work goes via pull request from feature branches into main; direct pushes to main require explicit human-author instruction. The repository's GitHub Pages site is configured with Source = GitHub Actions, deploying via the consolidated .github/workflows/build.yml (site + deploy jobs) on every push to main. No legacy gh-pages branch is used.

Form factor and page budget

F(AI)²R is published as a single short-form paper. The default budget is 10 pages of body, plus references and appendices, which do not count against the budget. There is no separate condensed manuscript; the condenser agent (agents/condenser.md) is repurposed as a page-budget enforcer that fires when make -C paper pages reports a body page count above the budget.

A budget change requires (a) human-author approval recorded in doc/logbook.md and (b) a corresponding update to paper/Makefile (PAGE_BUDGET).

Design system

The site, the paper PDF, and (forthcoming) the slide deck follow the DLR Corporate Design (CD-Handbuch §10.1 Schriften, §4 Printmedien, §10 Wording):

The institutional imprint (Florian Krebs, ORCID 0000-0001-6033-801X, DLR ZLP Augsburg, Helmholtz / NFDI4Ing / HMC) appears identically on the title block of paper/main.tex, the Acknowledgements section, the public site footer, CITATION.cff, codemeta.json, and .zenodo.json. Drift between any pair is a defect surfaced by the FAIR aligner.

Process evolutions during cooperative writing

The methodology above is the current state. The cooperation that produced this paper imposed roughly twenty structurally-significant changes on its own scaffolding while the work was underway; they are worth surfacing because a future cohort following this methodology will land at most of them again. The full chronology lives in doc/logbook.md; the summary below groups by class. The same digest appears as Table 1 in paper/sections/evolution.tex.

When Class What changed Trigger / impact
2026-05-06 Rule Primary-artefact consistency invariant Researcher direction; orchestrator named keystone; FAIR-aligner surfaces desync as audit fail.
2026-05-06 Rule Chapter-per-file rule for .tex paper/main.tex refactored as a thin assembler; one \section{} per file; enabled chapter-level handback.
2026-05-06 Rule Page budget (~10 pp body) and short-form pivot Researcher framing as a position paper; condenser agent repurposed as page-budget enforcer.
2026-05-06 Rule Source-research mandate; Consensus / Scholar default Cut hallucinated-citation surface; paywall-escalation sub-rule shipped together.
2026-05-06 Rule Contribution-tracking rule Researcher question "what is my contribution?"; spawned user-contributions.md and the fair2r:Contribution schema branch.
2026-05-06 Schema fair2r:Contribution class with seven sub-types Operationalises the partition the Author's Note describes; mirrors per-side audit.
2026-05-06 Schema ai-confirmed verification rung added Felt cost of single-tier verification; rung added rather than weakening the existing two.
2026-05-06 Schema fair2r:Slidedeck class; slides as fourth primary artefact Researcher request for two decks; presentation agent owns them.
2026-05-06 Agent presentation agent added to agents/ Slides became binding artefact; new prompt minted rather than overloading scientific-writer.
2026-05-06 Manuscript Position-paper reframing Researcher direction "fair should be more a position paper"; reshaped the contribution narrative.
2026-05-06 Manuscript Author's Note carried over from noheton/Obscurity-Is-Dead Voice-and-stance anchor; the methodology's first-person register lives there only.
2026-05-06 Pipeline Static site builder + Pages workflow + paper-build CI Made the artefact public from day one; cache-bust on the deployed site followed when staleness surfaced.
2026-05-06 Pipeline Marp slide pipeline retired; Beamer-only Researcher direction "only ship beamer"; one toolchain to maintain instead of two.
2026-05-07 Rule DLR Corporate Design pinned for every illustrator tool Default tool theming counts as an unfinished figure; written above the tool list, not below.
2026-05-07 Rule Illustration toolset opened beyond TikZ (matplotlib, mermaid, networkx, graphviz, …) Researcher direction; ladder-populations and contribution-histogram matplotlib figures unlocked.
2026-05-07 Rule No-parentless-claim rule (every claim must name a generating activity) Eight-claim defect found by the verification example; rule landed by the curator pass to prevent recurrence.
2026-05-07 Schema fair2r:repairs object property Curator pass: per-claim repair activity carries the reason as rdfs:comment; reasons are queryable.
2026-05-07 Agent research-protocol agent added to agents/ Standalone prompt for verification-axis research notes; produced provenance-verification.md.
2026-05-07 Pipeline Pages source switched from branch deploy to GitHub Actions Default branch became main; CI is single source of truth for the live site.
2026-05-07 Pipeline LaTeX-fragment stripper in the site builder Markdown sources embedded \emph{}, \textsc{}, \begin{enumerate}; the rendered Pages site needed a clean prose pass.

The catalogue is partial by construction — pure prose edits, single-paragraph content prompts, and figure refinements that did not change the scaffolding are recorded in doc/user-contributions.md (forty-eight typed entries at time of writing) but not promoted into this digest. The cut-off rule is "would a future researcher following this methodology run into the same decision?"; if yes, the row appears here.