Logbook

Append-only, newest at the bottom. One entry per working session, dated in ISO-8601. The logbook is the human-readable shadow of provenance.ttl; if you change the manuscript, add an entry here.

Format:

## YYYY-MM-DD — short title
*Author:* <human or agent name>
*Touched:* <files>
*Decision / outcome:* <one paragraph>
*Next:* <bullet list>

2026-05-06 — Repository initialised

Author: claude-opus-4-7 (under direction of repo owner) Touched: whole tree (initial commit) Decision / outcome: Bootstrapped F(AI²)R as a paper-plus-pipeline repository. Adopted the meta-structure distilled from noheton/Obscurity-Is-Dead: separation of paper/ (artifact), doc/ (process), and agents/ (LLM workforce); machine-readable metadata at the root (CITATION.cff, codemeta.json, .zenodo.json); a Makefile with stable targets; a PROV-O graph in Turtle as the spine of provenance; a verification ladder for claims. The IoT-security domain content of the source repository was deliberately excluded — only the meta-practices were transferred. Next: - Replace placeholder author block in CITATION.cff, codemeta.json, paper/main.tex. - Draft Section 1: Introduction via agents/scientific-writer.md. - Run agents/fair-aligner.md against the empty manuscript to baseline the audit report.

2026-05-06 — Primary-artifact consistency rule added

Author: claude-opus-4-7 (under direction of repo owner) Touched: CLAUDE.md, doc/methodology.md, agents/README.md, all ten files under agents/. Decision / outcome: Codified an invariant: the manuscript (paper/), the PROV-O graph (doc/provenance.ttl), and the logbook (doc/logbook.md) are primary artifacts and must remain consistent and up to date at all times. Every agent prompt now carries a "Primary-artifact consistency (binding)" block tailored to its role. The orchestrator is named ultimate custodian of the invariant; the provenance-curator is named keystone; fair-aligner will surface desync as a fail. Next: When agents/scientific-writer.md produces its first section draft, exercise the rule end-to-end as the test case.

2026-05-06 — Site builder, Pages workflow, paper-build CI

Author: claude-opus-4-7 (under direction of repo owner) Touched: site/static/style.css, site/index.md, scripts/build_provenance_site.py, scripts/requirements.txt, .github/workflows/pages.yml, .github/workflows/build-paper.yml, .gitignore. Decision / outcome: Added a static-site generator that renders doc/*.md, agents/*.md, and doc/provenance.ttl (parsed via rdflib) into _site/, mirroring the layout of noheton/Obscurity-Is-Dead's public site. Mermaid blocks render client-side via CDN. Added two GitHub Actions workflows: pages.yml builds and (on main) deploys via the official Pages actions; build-paper.yml compiles paper/main.tex and paper/main-condensed.tex with xu-cheng/latex-action@v3 and uploads PDFs as artifacts. Local smoke test: 20 pages built clean from the current sources. _site/ added to .gitignore. Next: Draft the paper using the methodology distilled from Obscurity-Is-Dead; trigger both workflows by pushing.

2026-05-06 — First-draft manuscript distilled from prior repo

Author: claude-opus-4-7 (under direction of repo owner) Touched: paper/main.tex, paper/references.bib, twelve files under paper/sections/, agents/scientific-writer.md, doc/logbook.md, doc/provenance.ttl. Decision / outcome: Drafted the full long-form manuscript by distilling the methodology and FAIR thinking of noheton/Obscurity-Is-Dead (same author), with all IoT-security domain content abstracted away. The framing of F(AI)²R is corrected to the source author's intent: FAIR axes with an (AI) factor multiplied through them, squared because each axis demands two passes (authoring and audit). The contribution claim is stated as the integration of eight individually unoriginal practices into a single discipline. The paper is structured as: Introduction; Background (FAIR, FAIR4RS, FAIR4ML, PROV-O, PRISMA, GRADE, AI-disclosure norms); The F(AI)²R Pattern (axes; ten-stage pipeline; verification ladder as FSM with PRISMA/GRADE crosswalk; handback discipline; mirror discipline); Eight Integrated Practices; Implementation in this Repository; This Paper, Generated by this Process; Process Evolution During Cooperative Writing; Failure Modes Named and Addressed; Discussion; Related Work; Conclusion; Acknowledgements; three appendices. main.tex was refactored into a thin assembler that \inputs one file per \section{}. Per researcher request, the Scientific Writer prompt gained a binding "File granularity" block codifying chapter-per-file with a 100-400-line soft target. Per researcher request, two new sections were added: self-reference.tex (an explicit attestation that the paper was produced by the pipeline it describes) and evolution.tex (a candid record of how the process itself changed during cooperative writing, for transparency about human-AI interaction). All section files measured: largest is pattern.tex at 221 lines, well within the granularity rule. Total ~1473 lines of LaTeX, well under the 50-page upper bound. References include canonical anchors (Wilkinson 2016, W3C PROV-O, FAIR4RS, datasheets, model cards, PRISMA 2020, GRADE, ICMJE, stochastic parrots) plus base-rate citations for fabricated-citation literature (Walters & Wilder 2023; Magesh et al. 2024) flagged TODO-VERIFY at lit-read. The FAIR4ML reference is an honest placeholder per the asset-honesty rule. Next: Trigger the build-paper workflow by pushing; review the compiled PDF for layout defects; promote lit-retrieved references to lit-read over the next sessions; populate the fair2r:Claim triples for each section in the provenance graph (currently three seed claims plus the invariant claim; many more named in the prose now require curation).

2026-05-06 — Short-form pivot, DLR design integration, sustainability + AI-rise sections, imprint

Author: claude-opus-4-7 (under direction of repo owner) Touched: every section file under paper/sections/ (compact rewrite); new paper/sections/sustainability.tex; new paper/style/fair2r.sty; rewritten paper/main.tex; paper/Makefile (drop condensed target, add pages target with PAGE_BUDGET); .github/workflows/build-paper.yml (drop condensed compile, add page-count report); deleted paper/main-condensed.tex; paper/references.bib (add liang2024mapping, kobak2024delving, else2023chatgpt, vannoorden2023chatgpt); site/static/style.css (rewritten on DLR tokens); site/static/dlr/ (vendored DLR colors_and_type.css, four Frutiger TTFs, dlr-logo.svg, dlr-logo-white.svg); scripts/build_provenance_site.py (DLR layout template — utility bar, header with logo + org line, blue nav, hero, sidebar, content, dark-grey footer ribbon with imprint); site/index.md; CITATION.cff, codemeta.json, .zenodo.json, README.md (with the GitHub Pages URL https://noheton.github.io/f-ai-r/), paper/sections/acknowledgements.tex (with the Florian Krebs / DLR ZLP / Helmholtz / NFDI4Ing / HMC imprint); agents/condenser.md (repurposed as page-budget enforcer); agents/scientific-writer.md (page-budget rule + DLR voice rules + British English); agents/README.md; doc/methodology.md (form factor and design section); doc/submission-plan.md (single-form venues, including NFDI4Ing / HMC / RDA / FORCE11); doc/todos-for-publication.md (presentation deferred to P3 with explicit Pages-URL on slides requirement). Decision / outcome: Pivoted to the short-form-only manuscript per researcher request. The long/condensed split was retired; the condenser agent was repurposed as a page-budget enforcer that fires when make -C paper pages exceeds PAGE_BUDGET (default 10). Section files were rewritten compactly. New section sustainability.tex (resource cost, frontier-model dependence, equity, long-term archivability) added per researcher request. New intro paragraph added on the recent rise of AI-modified prose at journals and outlets (Liang 2024, Kobak 2024, Else 2023, Van Noorden 2023), motivating why a discipline like F(AI)²R is necessary now. The DLR Corporate Design system (CD-Handbuch §10.1, §4, §10) was integrated end-to-end: tokens vendored under site/static/dlr/; site CSS rewritten on those tokens (utility bar, DLR-blue nav, hairline borders, square corners, photography-led imagery); LaTeX style file paper/style/fair2r.sty carries the equivalent typography (helvet→Arial fallback) and accent colours into the PDF. Imprint updated with the institutional identity: Florian Krebs, ORCID 0000-0001-6033-801X, DLR ZLP Augsburg, Helmholtz / NFDI4Ing / HMC. The GitHub Pages URL https://noheton.github.io/f-ai-r/ is now in the README, in CITATION.cff, in codemeta.json, in .zenodo.json, and on the public site footer; a deferred TODO records that the same URL must appear on the imprint slide and title-slide footer when the slide deck is built. Next: Run the build-paper workflow; verify the compiled PDF stays within the 10-page body budget; merge PR #1; mint Zenodo DOI; replace TODO-VERIFY references with lit-read once the human author has read them.

2026-05-06 — Mermaid graph fixed for the GitHub renderer

Author: claude-opus-4-7 (under direction of repo owner) Touched: doc/provenance-graph.md, scripts/build_provenance_site.py. Decision / outcome: Repaired three GitHub-Mermaid-renderer incompatibilities in doc/provenance-graph.md: (a) the subgraph Plans (prompts) title with bare parentheses was rejected by the parser and is now subgraph plans["Plans (prompts)"]; (b) node labels containing colons (e.g. [fair2r:HumanResearcher]) need to be wrapped in double quotes (e.g. ["fair2r:HumanResearcher"]); (c) edge labels containing colons (e.g. -- prov:wasGeneratedBy -->) were rewritten to the pipe form (e.g. -->|"prov:wasGeneratedBy"|). These are the same constraints documented at https://docs.github.com/get-started/writing-on-github/working-with-advanced-formatting/creating-diagrams#creating-mermaid-diagrams.

The site builder (scripts/build_provenance_site.py) had a related defect that appeared only after the diagram was quoted: the markdown library was HTML-escaping the contents of fenced ```mermaid blocks (turning " into " and <br/> into <br/>), which would have broken client-side Mermaid rendering on the public site. The fix extracts Mermaid blocks before markdown runs, replaces them with sentinel tokens, and re-injects the original (un-escaped) source after rendering inside <div class="mermaid"> wrappers. This is verified locally: _site/provenance-graph.html now contains raw " and <br/> inside the Mermaid block, which Mermaid's CDN loader can parse. Next: Visual confirmation on the deployed Pages site after merge.

2026-05-06 — Promote the GitHub Pages URL to a top-of-README callout

Author: claude-opus-4-7 (under direction of repo owner) Touched: README.md. Decision / outcome: The Pages URL was already at line 6 of README.md, tucked into a structured contact list. Researcher reported it as missing — the structural placement made it easy to miss. Promoted to a dedicated "Read on the web" section directly under the tagline, with the URL rendered as a large markdown link. The horizontal rule below the section preserves the structured contact list that follows. A first edit included a globe emoji (🌐) in the heading; immediately reverted to obey the DLR voice rule "No emoji. Anywhere." that the Scientific Writer prompt now binds. Next: If the Pages URL is still not where the researcher expects, fix it in one more pass.

2026-05-06 — CI fix, provenance-analysis section, pipeline figure, interactive explorer, draft-PDF release, sources, consent gate

Author: claude-opus-4-7 (under direction of repo owner) Touched: paper/style/fair2r.sty (full rewrite to fix CI), paper/main.tex (DRAFT mode default; Helmholtz/NFDI4Ing/HMC separator), paper/Makefile (add provenance target), paper/sections/intro.tex (correct base-rate framing per source analysis), paper/sections/failure-modes.tex (split base-rate attribution), paper/sections/pattern.tex (insert pipeline figure), new paper/sections/provenance-analysis.tex, new paper/sections/_generated/*.tex (auto-generated), new paper/figures/pipeline.tex (TikZ ten-stage pipeline figure), new scripts/provenance_analysis.py (graph-to-LaTeX-fragments extractor), new site/provenance-explorer.md (vis-network interactive viewer), site/index.md (splashy hero with PDF CTA, card grid), site/static/style.css (CTA buttons, card grid, explorer canvas), scripts/build_provenance_site.py (JSON exporter for the explorer; new nav entry; updated PAGES list), .github/workflows/build-paper.yml (add Python step; publish DRAFT-watermarked PDF as a rolling latest-draft GitHub Release on main), .github/workflows/pages.yml (gate Pages-specific steps on main so working-branch builds do not fail when Pages is not yet configured for the repo), paper/.gitignore (ignore sections/_generated/), paper/references.bib (correct vannoorden2023chatgpt DOI from d41586-023-03907-5 to d41586-023-03930-6, second author Webb not Perkel, page 509; correct ravi2024fair4ml author block to Castro/Garijo and the RDA FAIR4ML WG with Zenodo DOI 10.5281/zenodo.14002310), doc/sources.md (full sources index, produced by a research subagent — most canonical references promoted to lit-read, remaining items at lit-retrieved flagged TODO for human reading; magesh2024legal figures corrected to 17–34% for legal RAG, distinct from Walters/Wilder's 18–55% for full-citation generation), new doc/publication-consent.md (gate file, status: not yet given). Decision / outcome: Diagnosed and repaired the paper-build CI failure on PR #2: the LaTeX style file had bugs (the \ProvidesPackage self-name carried a style/ prefix; xkeyval option processing was incompatible with the actual options; \arrayrulecolor was used without loading colortbl; sansmath was loaded for no benefit; a custom quote redefinition risked breaking section content). Rewrote with a simpler \providecommand{\faiarVariant}{a} mechanism, loaded colortbl and textcomp, dropped the fragile dependencies, kept the DLR-Blue accent and mid-grey H1 typography. Default mode is \faiarDraft=true, applying a "DRAFT — not yet peer-reviewed" watermark on every page via draftwatermark.

Added the Provenance Analysis section per researcher request, backed by scripts/provenance_analysis.py which queries the graph at build time and emits four LaTeX fragments under paper/sections/_generated/: rung-distribution, section-coverage, source-strength (with ghost-citation count), and recent-activities. The section will stay current automatically as the graph evolves.

Added the interactive provenance explorer at provenance-explorer.html: a vis-network node-link diagram populated from a JSON projection of the graph (100 nodes, 220 edges at the time of writing), with class-filter, label/IRI search, and click-to-inspect. The JSON exporter sits alongside the existing tables view; both are generated by the site builder.

Added the TikZ pipeline figure (paper/figures/pipeline.tex) in pattern.tex, showing the ten stages colour-coded per pass (authoring blue, audit green, curation yellow, orchestration / site grey) with the dashed handback edge. The figure uses the DLR colour ramps from the style file, so the visual identity carries through the PDF.

Improved the homepage with a hero CTA-row (linking the latest DRAFT PDF, the repository, and the explorer), a callout that explains the DRAFT status and the local-replay loop, and a three-card grid pointing to Methodology, Provenance, and Agents.

Added draft-PDF distribution to the build-paper workflow: on pushes to main, the freshly-built (watermarked) PDF is published as a rolling latest-draft GitHub Release, giving a stable download URL the README and site link without rerunning CI per reader. The watermark and the consent gate make the "unpublished but readable" status unambiguous.

Made the Pages workflow robust against repos where Pages is not yet enabled: working-branch builds still validate the Turtle, build the site, and upload it as a workflow artefact, but skip the Pages-specific configure-pages and upload-pages-artefact steps unless the ref is main. This was the cause of the working-branch build failures observed on PRs #2 and #3.

Added the Sources index at doc/sources.md (researched by an isolated source-analyser subagent against the verification ladder of doc/methodology.md). Most canonical references advanced to lit-read; the remainder sit at lit-retrieved pending human reading. The research surfaced two factual corrections in paper/references.bib (the vannoorden2023chatgpt DOI and the ravi2024fair4ml author block) and one phrasing correction in paper/sections/intro.tex and paper/sections/failure-modes.tex (the 18–55% figure was conflating Walters/Wilder with Magesh et al.; the two are now cited distinctly with their actual ranges).

Added publication-consent.md as the explicit gate for any publication surface beyond the public Pages site and the DRAFT-watermarked release. Status: not yet given. Zenodo deposition, arXiv submission, conference upload, and watermark removal all require an appended consent record. Next: Visual confirmation of the rebuilt PDF; visual confirmation of the new homepage hero and the explorer; on receiving explicit publication consent, prepare the Zenodo + arXiv submission packages.

2026-05-06 — Site cache-busting + Consensus/Scholar mandate in the source workflow

Author: claude-opus-4-7 (under direction of repo owner) Touched: scripts/build_provenance_site.py, site/provenance-explorer.md, agents/source-analyzer.md, doc/research-protocol.md. Decision / outcome: The deployed Pages site appeared to serve a stale bundle after a fresh deploy — typical browser / CDN cache. Added a build-time cache-bust version computed as a SHA-1 over the files that affect rendered pages (site/static/style.css, doc/provenance.ttl, site/index.md, site/provenance-explorer.md); each rendered page references static/style.css?v=<hash> and carries a <meta name="build-version" content="<hash>"> tag. The explorer JS reads that meta and appends the same ?v=<hash> to the static/provenance.json fetch, so the interactive viewer participates in the same cache-busting scheme as the CSS. A softer <meta http-equiv="Cache-Control" content="no-cache, must-revalidate"> is also emitted as a fallback. The footer prints the build version so a reader can verify which bundle their browser is rendering. The source workflow now mandates a peer-reviewed-corpus search tool first, web search second. Default: the harness Consensus / Scholar tool (over 200M peer-reviewed papers across Semantic Scholar, PubMed, Scopus, and arXiv). Fallbacks: Semantic Scholar / OpenAlex / arXiv direct, then web search for grey literature (editorials, venue policies, USCO / UrhG, W3C / RDA recommendations). agents/source-analyzer.md and doc/research-protocol.md now both carry the binding tool catalogue. Promotion through the verification ladder still requires an actual fetch of the source plus a quoted snippet for ai-confirmed, and a human reader for lit-read. Next: Await the source-verification subagent; merge its results in a follow-up commit.

2026-05-06 — Second-pass source research, scientific-repository tool list, institutional-access channel

Author: claude-opus-4-7 (orchestrator) + source-research subagent Touched: paper/references.bib (eleven new entries), paper/sections/intro.tex, paper/sections/background.tex, paper/sections/failure-modes.tex, paper/sections/sustainability.tex, doc/sources.md, agents/source-analyzer.md, doc/research-protocol.md, new doc/sources-needing-institutional-access.md, doc/provenance.ttl. Decision / outcome: The second-pass source-research subagent ran against the verification ladder of doc/methodology.md and returned eleven new BibTeX entries, all identifier-verified, all flagged TODO-VERIFY at lit-read pending human reading: luccioni2024power, patterson2021carbon, li2023thirsty, strubell2019energy, birhane2022values (sustainability paragraph); shumailov2024collapse, alemohammad2023mad (model-collapse row in the failure-modes table); chen2023drift (frontier-model dependence in sustainability); thorp2023chatgpt (intro editorial position); usco2023ai (US Copyright Office 2023 Federal Register guidance) and urhg2 (German UrhG §2) covering the legal-honesty-about-authorship paragraph in background.tex. Each new citation lands with an inline \todo[inline]{verify} placeholder so the human author can confirm at lit-read. The subagent terminated with a 529 overload after 45 tool calls, having completed the canonical work; nothing was lost.

The source workflow is extended with a full catalogue of scientific repositories: arXiv, OpenAlex, Crossref, PubMed / EuropePMC, IEEE Xplore, Springer Link, ACM Digital Library, ScienceDirect / Elsevier, plus the standards-track surfaces W3C / RDA / NIST / ISO. agents/source-analyzer.md and doc/research-protocol.md both carry the binding catalogue.

The institutional-access channel is a new file doc/sources-needing-institutional-access.md — append-only, schema-tight. When a source is necessary for a load-bearing claim but the full text is paywalled and the abstract is insufficient, the agent appends a structured request (bibkey, identifier, why-needed, specifically-need, abstract-ok?, status). The human author has DLR institutional access via TIB / ZB MED / Helmholtz e-journals / direct publisher subscriptions, and supplies the PDF; once supplied, the source moves to source-vendored under doc/sources/<bibkey>/. A pending request older than 30 days is a soft warn from the Aligner; a claim unverified for more than 90 days must be hedged in prose or removed. This is the explicit escalation path that replaces the silent downgrade / synthesised-quotation failure mode named in §\ref{sec:failure-modes}. Next: Visual review of the eleven new citations; on the next session, advance some of them to lit-read and curate per-claim prov:wasDerivedFrom triples in the graph.

2026-05-06 — Author's Note carried over from Obscurity-Is-Dead; position-paper reframing in progress

Author: claude-opus-4-7 (under direction of repo owner) Touched: new paper/sections/authors-note.tex, paper/main.tex, doc/logbook.md, doc/provenance.ttl. Decision / outcome: Researcher requested that F(AI)²R reflect the author's notes from the methodological ancestor and read more as a position paper. A first piece landed: a front-matter "Author's Note — Advice for reading" adapted from the source paper, with five subsections: This is a position paper, not a survey; Open to constructive criticism — actively; Frameworks evolve with capabilities; What surprised me about the cooperation (a finding for the recursive case study, naming the human-deciding / model-drafting partition); Tokens out, tokens in (machine-readable sources are committed under permissive licences so a future agent can compress or extend without laundering provenance through a binary PDF). The note is wired into paper/main.tex between the abstract and the introduction. The remaining position-paper restructuring (numbered Position section; Anticipated Objections section; Statement of authorship and AI use; sharpened conclusion; updated prose in paper/sections/intro.tex and paper/sections/discussion.tex; prompt and methodology updates) is queued for a follow-up commit. A separate observation, also from the researcher mid-session: "I wonder how much of the work I'm actually doing. What's my contribution." The model offered a perspective in chat: the human does the deciding, the abstracting, the position-taking, the catching of drift, and the staking of reputation; the model does the drafting, the cross-referencing, the structuring, and the catching of second-order errors. The decisions are doing the load-bearing work; the drafting is the scale multiplier. Whether that question and its answer should appear in the paper itself (as a recorded mid-writing aside in the What surprised me note or in evolution.tex) is left for the researcher to decide on the next session. Next: If the researcher endorses, fold the contribution-of-the- human paragraph into the Author's Note; complete the position- paper restructuring (Position section, Objections section, Statement of authorship).

2026-05-06 — Position-paper completion + Contribution entity + user-contributions / user-observations logs + back-fill

Author: claude-opus-4-7 (under direction of repo owner) Touched: paper/sections/conclusion.tex (sharpened), paper/sections/discussion.tex (marked superseded; preserved for history), paper/sections/acknowledgements.tex (marked superseded); paper/sections/position.tex (new — five numbered claims); paper/sections/objections.tex (new — six anticipated objections with charitable responses); paper/sections/statement-of-authorship.tex (new — adapts the source paper's Statement of independence for DLR/Helmholtz/NFDI4Ing/HMC institutional context, plus AI-use disclosure); paper/sections/authors-note.tex (extended What surprised me about the cooperation with the user's question and the model's perspective on the human contribution, per researcher "yes add it"); paper/sections/evolution.tex (new sub-section Field notes from the second cooperative writing project with three promoted observations: researcher-role shifting toward experimentation; prompt-craft is closer to programming than to drafting; responsibility includes the rooms after publication); paper/main.tex (re-ordered with new sections; old discussion and acknowledgements no longer included); new doc/user-contributions.md (schema + back-fill of 21 prior human contributions reconstructed from git log, conversation, and logbook); new doc/user-observations-log.md (schema + back-fill of 8 meta-observations including this turn's: token budgets, experience-meta, researcher role, prompts feel easier, responsibility-uptake, gambling metaphor, prompt accretion); doc/provenance.ttl (new fair2r:Contribution / HumanContribution / AIContribution / MetaContribution schema extension; new properties fair2r:contributionType, fair2r:contributionLeverage, fair2r:promotedToPaper; 21 back- filled hc:* HumanContribution instances; 8 back-filled MetaContribution instances; new claim claim:contribution-tracking-rule); CLAUDE.md (new Contribution tracking (binding) block); agents/orchestrator.md (orchestrator made responsible for the contribution-log step at the end of every session); scripts/build_provenance_site.py (new user-contributions and user-observations site pages and nav entries). Decision / outcome: The user posed two interleaved requests: (a) fold the what's my contribution? question into the paper, and (b) make contribution-tracking a binding rule with a new entity in the schema. Both landed. The Author's Note now contains the question, the model's perspective, and the explicit naming of the partition. The Contribution schema makes the partition machine- readable: every material decision, correction, prompt, rule, or observation now has a typed entity with a leverage rating and a "promoted-to-paper" flag (for observations). Two new site pages mirror the markdown logs. The position-paper restructuring is also complete: a five-claim Position section near the front, a six-objection Objections section near the back, a sharpened Conclusion that absorbs the submission-conditions paragraph from the retired Discussion, and a Statement of authorship and AI use that adapts the source paper's Statement of independence for an institutional setting. Three field-note observations from the second-time cooperative-writing experience are promoted into paper/sections/evolution.tex; the rest stay in doc/user-observations-log.md per the curation rule the file documents at the top. Local smoke test: provenance.ttl parses to 976 triples; site builds to 23 pages; section files total 1478 lines of LaTeX (well under the 50-page upper bound, with the 10-page body budget still tight; a follow-up condenser pass will trim). Next: Trigger the build-paper workflow on main; await the visual confirmation of the new sections; fold the remaining pending observations into the paper if the researcher endorses; add a companion doc/ai-contributions.md so the partition the Author's Note describes is symmetric.

2026-05-06 — Accelerator-or-blender observation logged

Author: claude-opus-4-7 (under direction of repo owner) Touched: doc/user-observations-log.md, doc/provenance.ttl, doc/logbook.md. Decision / outcome: Researcher posed the cognition question: "I can build a vision in my head, mumble ideas as they come and go, and LLMs make sense of that and put it into a mostly / sometimes / usually useful structured context. I wonder what that means for 'cognition' in biological systems — is GenAI some kind of accelerator of human potential, or is it just a big cultural blender?" The model offered a perspective in chat: probably both, conditional on where in the cognitive stack the model is allowed to operate. LLMs interpolate fluently within the manifold of prior text and accelerate when they stay on the prose-craft side of the partition; LLMs do not extrapolate off the manifold and become cultural blenders when they cross into the abstracting layer. The Author's Note's partition (the human does the deciding, the model does the drafting) is therefore not just a description but a coupling rule. The honest test for any collaboration: ask whether the abstraction would have been thinkable without the model. The observation is logged in doc/user-observations-log.md as a strong candidate for paper graduation (sharpens What surprised me about the cooperation) and mirrored as hc:obs-accelerator-or-blender in the graph at promotedToPaper="pending". The prompt-accretion entry's italic conclusion was inadvertently overwritten in the first edit and has been restored in the same commit. Next: On researcher endorsement, extend the Author's Note's What surprised me about the cooperation paragraph with the coupling-rule sentence.

2026-05-06 — Author's Note coupling-rule graduation; third-pass source-research dispatched

Author: claude-opus-4-7 (under direction of repo owner) Touched: paper/sections/authors-note.tex (coupling-rule paragraph appended to What surprised me about the cooperation), doc/provenance.ttl (hc:obs-accelerator-or-blender flipped from promotedToPaper="pending" to "yes" with explanatory rdfs:comment; new hc:third-pass-research HumanContribution), doc/user-contributions.md (new entry). Decision / outcome: The researcher endorsed graduation of the accelerator-or-blender observation. The Author's Note's What surprised me about the cooperation paragraph now contains both the question and its answer, plus a second-order finding stated as a coupling rule: the cooperation works as an accelerator when the model stays on the prose-craft side of the partition; it collapses toward a cultural blender when allowed across into the abstracting layer. The honest diagnostic is recorded inline: ask whether the abstraction would have been thinkable without the model. A third-pass source-research subagent was dispatched in parallel to build further sources (manifold / extrapolation; distributed cognition / extended mind; prompt-engineering literature; HMC / NFDI4Ing position papers; verification of inference-cost references that remain at lit-retrieved) and to run a context-data pass (orphan fair2r:HumanContribution checks; dangling-cite check across paper/sections/*.tex; section-graph mismatch checks; unactioned-Next: todos). Its results will land in a follow-up commit. Next: Fold the subagent's bibliography additions and the context- data findings; commit + PR + merge; trigger the build-paper workflow on main to refresh the latest-draft PDF.

2026-05-06 — Three illustrations + 11 new sources + section-graph repair

Author: claude-opus-4-7 (under direction of repo owner) Touched: new paper/figures/ladder-fsm.tex, new paper/figures/axes.tex, new paper/figures/coupling-rule.tex; paper/sections/pattern.tex (figures wired in §3.1 Axes and §3.3 FSM); paper/sections/authors-note.tex (coupling-rule figure wired; four new \cite{} keys: clark1998extended, hutchins1995cognition, clark2025extending, anderson2024homogenization); paper/sections/intro.tex (\cite{sadasivan2023reliably} added to the detection-arms-race sentence); paper/sections/evolution.tex (\cite{kuteeva2024diversity} for humanities-exception; \cite{reynolds2021prompt,liu2023prompt} for prompt-as-programming); paper/sections/statement-of-authorship.tex (\cite{curdt2025hmc} and \cite{schmitt2020nfdi4ing} added to the institutional-context paragraph); paper/references.bib (11 new entries, all lit-retrieved); doc/sources.md (new entries indexed); doc/provenance.ttl (four new fair2r:Section IRIs for section-position, section-objections, section-authors-note, section-statement-of-authorship; one prov:Invalidation for ent:section-discussion; ten new fair2r:Source entities; three new fair2r:Figure entities; new act:rev-readability-figures activity; new hc:readability-figures HumanContribution). Decision / outcome: Researcher noted the paper is "still a bit hard to read" and asked for more illustrations. Three TikZ figures were added on the DLR token palette so they fit the existing visual grammar: (a) the verification ladder as a finite-state machine with five colour-coded rungs and the dashed retraction back-edge; (b) the F(AI)²R axes as a 4× 2 grid showing the (AI)² factor multiplied through every axis with one row per pass; (c) the coupling rule from the Author's Note rendered as a two-layer partition with accelerator / cultural-blender outcomes and the diagnostic question along the bottom. Each figure replaces a paragraph that had been carrying load by prose alone. A third-pass source-research subagent ran in parallel and added eleven new bibliography entries (Clark & Chalmers 1998, Hutchins 1995, Clark 2025, Anderson et al. 2024, Kuteeva & Andersson 2024, Reynolds & McDonell 2021, Liu et al. 2023, Sadasivan et al. 2023, Curdt et al. 2025 HMC, Schmitt et al. 2020 NFDI4Ing) and ran a context-data pass that caught three section-graph mismatches: ent:section-discussion existed in the graph after its file was retired; the four position-paper section files had no corresponding fair2r:Section IRIs. All three issues are repaired in this commit. The subagent reported zero dangling \cite{} keys, zero orphan hc: IRIs. The Consensus tool reports 3 searches remaining this month (resets June 1st); the project should plan further peer-reviewed verification accordingly. Next: Trigger the build-paper workflow on main to refresh the latest-draft PDF with three new figures and the citations; visual confirmation; advance the eleven new lit-retrieved entries to lit-read once the human author has read them; on endorsement, draft a companion doc/ai-contributions.md so the partition the Author's Note describes is symmetric.

2026-05-06 — Domain-ontologies extension + reproducibility-crisis + reviewer-AI policies + reading-queue generator

Author: claude-opus-4-7 (under direction of repo owner) Touched: new §3.5 in paper/sections/pattern.tex Extensibility through domain ontologies (citing SOSA / SSN, OM-2, QUDT; carries the German fragment "konkrete standardisierte Methode" as a load-bearing technical term in the source-author voice); paper/sections/intro.tex (\cite{ioannidis2005,pineau2021reproducibility} for the reproducibility-baseline anchor); paper/sections/background.tex (\cite{neurips_llm_policy,iclr_llm_policy,aclrr_llm_policy} for the reviewer-side AI policies); paper/references.bib (8 new entries: 3 ontologies, 2 reproducibility-crisis, 3 venue policies); doc/provenance.ttl (8 new fair2r:Source entities; 3 new fair2r:Claim entities; hc:domain-ontologies, hc:gap-fill-reproducibility-and-policies, hc:reading-queue HumanContributions); new scripts/build_reading_queue.py (auto-generates doc/reading-queue.md from the graph plus the bib, sorted by load-bearing weight); new auto-generated doc/reading-queue.md (44 entries); paper/Makefile provenance target now also runs the queue generator; .github/workflows/build-paper.yml regenerates both fragments and the queue; scripts/build_provenance_site.py adds the Reading queue page to the public site nav. Decision / outcome: Researcher proposed a new claim: "ontologies might actually help to make this process more efficient and reproducible (first example prov, but for example sensor data and OM-2 (units of measure) ontology — konkrete standardisierte methoden ... potential)". The claim graduated to paper text as a new subsection §3.5 in pattern.tex. Three sources cited with \todo[inline]{verify}; bib entries added with TODO-VERIFY notes; new fair2r:Claim claim:domain-ontologies-extension recorded at human-confirmed. The German fragment was carried over verbatim to keep the voice consistent with the source author's house style. The same researcher endorsed two of the three gaps the model had flagged earlier (reviewer-side AI policies; reproducibility-crisis literature). Five further bib entries added; citations inserted into intro.tex and background.tex. Two further fair2r:Claim entities recorded. The researcher then asked: "do we have a list of reading tasks for the researcher?". A new generator scripts/build_reading_queue.py was written; it reads doc/provenance.ttl and paper/references.bib, finds every source not yet at lit-read, sorts by the number of dependent fair2r:Claim entries, and emits doc/reading-queue.md. The queue currently lists 44 entries: 32 at lit-retrieved, 10 at ai-confirmed, 2 graph-cleanliness defects without rungs. The queue is wired into the build pipeline (Makefile + CI) and surfaced as a tab on the public site so the researcher always has a current, prioritised list. Local smoke test: provenance.ttl parses to 1184 triples (up from 1096); reading-queue.md regenerates cleanly with 44 entries; 8 new bib entries, all lit-retrieved and identifier-verified to the extent possible without full-text reading. Next: Trigger build-paper on main; await visual confirmation of the new §3.5 and the additional citations; advance the 24 lit-retrieved entries the human author has time to read; fix the 2 graph-cleanliness defects (sources without rungs).

2026-05-06 — §3.6 forward-looking subsection: bioinformatics precedent + research-infrastructure vision

Author: claude-opus-4-7 (under direction of repo owner) Touched: paper/sections/pattern.tex (new §3.6 Looking forward: provenance graphs as research infrastructure), paper/references.bib (new entry ashburner2000go), doc/provenance.ttl, doc/user-contributions.md, doc/logbook.md. Decision / outcome: Researcher proposed two paired forward-looking claims:

Claim 1. Future research success might depend a lot more on provenance models for hypothesis development, graph interconnectivity (Helmholtz graph supercharged) and clever experimental validation.\\ Claim 2. Bioinformatics was driven there earlier by the at-the-time "obscene amounts of data" leading to semantic tagging becoming the norm to identify genes.

Both graduated to paper text as a single new subsection §3.6 of pattern.tex, immediately following §3.5 Extensibility through domain ontologies. The bioinformatics precedent is anchored on Ashburner et al. 2000 (the canonical Gene Ontology paper, Nature Genetics 25:25–29, DOI 10.1038/75556) — a high-citation, high-recognition reference that establishes the semantic-tagging-as-norm-under-data-pressure case in one move. The Helmholtz cross-centre graph reuses the existing curdt2025hmc citation rather than adding a fresh entry. The marketing-register verb "supercharged" was rephrased to "fully-instrumented" to fit the DLR voice rules in agents/scientific-writer.md. Two new fair2r:Claim entries (claim:bioinformatics-precedent, claim:future-research-infrastructure), both at human-confirmed because both rest on researcher-supplied abstraction rather than on AI-confirmed reading of the underlying sources. New act:rev-research-infrastructure activity; new hc:research-infrastructure HumanContribution with the verbatim researcher quote in the rdfs:comment. New ent:section-pattern-research-infrastructure fair2r:Section entity wired with prov:wasGeneratedBy. The §3 (Pattern) section is now a full position-paper spine: §3.1 axes, §3.2 pipeline, §3.3 verification ladder, §3.4 handback + mirror disciplines, §3.5 extensibility through domain ontologies, §3.6 looking forward. Next: Trigger the build-paper workflow; verify §3.6 reads cleanly in the compiled PDF; promote the bioinformatics-precedent and future-research-infrastructure claims into the Conclusion's call-to-action if the researcher endorses; advance ashburner2000go and curdt2025hmc to lit-read once the human author has read them in full.

2026-05-06 — "The longer arc" paragraph promoted into the Conclusion

Author: claude-opus-4-7 (under direction of repo owner) Touched: paper/sections/conclusion.tex, doc/provenance.ttl, doc/user-contributions.md, doc/logbook.md. Decision / outcome: Researcher endorsed (with "yes") the model's proposal to promote claim:future-research-infrastructure into the Conclusion. New paragraph "The longer arc" inserted between the "Call to action" and the closing aphorism. The paragraph restates the bioinformatics-precedent / forward-research-infrastructure argument from §3.6 in compressed form, names the three properties the schema is designed to support (provenance models for hypothesis development; graph interconnectivity across institutional knowledge graphs; clever experimental validation against the graph), and frames F(AI)²R as "the entry ticket, not the destination". The graph carries a new act:rev-conclusion-longer-arc activity linking the claim to its second appearance (the conclusion section, in addition to its primary home in §3.6); a new hc:promote-longer-arc HumanContribution records the researcher endorsement. The position-paper closer now ends on the forward- looking note before the aphorism, which strengthens the position-paper register asked for two sessions ago. Local smoke test: provenance.ttl parses to 1262 triples (up from 1250); reading queue stable at 45 entries. Next: Trigger build-paper on main to refresh the latest-draft PDF with the new Conclusion paragraph.

2026-05-06 — Latency-as-feature observation logged

Author: claude-opus-4-7 (under direction of repo owner) Touched: doc/user-observations-log.md, doc/provenance.ttl, doc/logbook.md. Decision / outcome: Researcher posed a meta-observation about cooperative writing: "the experience of using LLM for work would be fundamentally different if results were near-instantaneous; this prompt / meta-conversation planning keeps the ball rolling", then extended it with "and allows for short context switches — e.g. checking on a coding agent". The model offered a perspective in chat: latency is doing real cognitive work (formulation pauses are exactly the abstracting-layer work the partition needs), and the wait window is not idle but productive parallel-tracking time amortising the cognitive cost of context-switching across slow loops the human author is already running. A counter-claim is recorded honestly: instantaneous response could be a different kind of beneficial (closer to typing assistance, model-as-keyboard rather than model-as-conversation-partner). The observation is logged as hc:obs-latency-as-feature (fair2r:MetaContribution, leverage high, promotedToPaper pending) and points to two candidate paper edits: a sentence in authors-note.tex's coupling-rule paragraph naming latency as part of the defence, and a candidate ninth integrated practice (the supervisory-glance / multi-tracked-research pattern) marked for a future revision. Next: On researcher endorsement, fold the latency-as-feature sentence into authors-note.tex; consider whether the supervisory- glance pattern warrants graduation to a numbered practice or remains as a field-note in evolution.tex.

2026-05-06 — Two slide decks (DLR Marp theme) + presentation agent + slides promoted to fourth primary artefact

Author: claude-opus-4-7 (under direction of repo owner) Touched: slides/pitch-5min.md, slides/conference-30min.md, slides/Makefile, slides/.gitignore, slides/static/dlr/dlr.css, slides/static/dlr/mermaid.css, slides/static/dlr/assets/{slide_back.png, title_back.png, dlr-logo.svg, dlr-logo-white.svg}, new agents/presentation.md, new .github/workflows/build-slides.yml, CLAUDE.md, doc/methodology.md, doc/provenance.ttl, doc/user-contributions.md, README.md, site/index.md. Decision / outcome: Researcher requested two slide decks — "every paper needs a slide deck. I want you to design 2 slide decks. A 5 minute pitch deck and a 25+5 conference deck for the paper. can you generate it using the DLR DESIGN system. both should be treated as primary artifacts. included in ci. create a fitting presentation agent". Implementation: The DLR design-system bundle's canonical slide kit is ui_kits/marp/ (Marp + DLR theme, Markdown → PPTX/PDF). Vendored the theme (dlr.css, mermaid.css) and the layout assets (background plates, logos) under slides/static/dlr/. Two source decks committed: pitch-5min.md (six slides, ~50 s each, claim per slide, no figure overload) and conference-30min.md (~28 slides, section-divider slides between the position-paper's natural breaks, backup slides included). Both use only the DLR-theme layout classes (title, section-divider, toc, thanks); no inline HTML. Both close on the position-paper aphorism (pipeline-team-methodology / specialisation-at-the-seam / naming-the-practice). A slides/Makefile runs the Marp CLI via npx, producing dist/<deck>.{pdf,pptx,html}. The .github/workflows/build-slides.yml mirrors the paper-build pattern: it installs Node + Chromium + Marp, renders both decks to all three formats, uploads the workflow artefact, and on main publishes a rolling latest-draft-slides release with the PDF + PPTX downloads. The README and the site index now carry direct download links to all four files. A new agents/presentation.md is the slide-deck custodian. It binds the same primary-artefact-consistency rule the other agents follow, plus a slide-specific addendum: a claim on a slide must already appear in the paper text; a slide that contradicts the paper is a defect. The agent also restates the DLR voice rules and the variant-locking rule (one chapter accent per section). The biggest invariant-level change: slides are promoted to the fourth primary artefact. CLAUDE.md and doc/methodology.md now name four primary artefacts (manuscript, PROV-O graph, logbook, slides). A new fair2r:Slidedeck class extends the schema; both decks have fair2r:Slidedeck IRIs in the graph. A new claim:slides-as-primary-artefact formalises the rule. Local smoke test: provenance.ttl parses to 1331 triples (up from 1276, +55 for the new schema and entries). The slide builds cannot be smoke-tested locally without a Node + Chromium install; the CI workflow is the canonical render path. Next: Trigger build-slides on main after merge to verify the DLR theme renders cleanly; on the first set of rendered decks, spot-check the section-divider slides' background plates and the title slide's accent rule; replace any pending source rungs that the conference deck's content depends on (e.g. magesh2024legal, walters2023fabrication) with lit-read once the human author has read them in full.

2026-05-06 — Coding-with-LLMs structure observation logged

Author: claude-opus-4-7 (under direction of repo owner) Touched: doc/user-observations-log.md, doc/provenance.ttl, doc/logbook.md. Decision / outcome: Researcher posed a meta-observation about coding cooperation: "In coding with LLMs somehow structure becomes more important, like guiding and shaping a river of code — it's more about visions, feature sets, dependencies." The model offered a perspective in chat: the observation is the coding-counterpart of the writing observations earlier in this log. The same partition appears (human at the architectural layer, model at the executable layer); the same coupling rule transposes (accelerator vs cultural blender, with the diagnostic "would the architecture have been thinkable without the model?"); the bottleneck shifts from production to integration when generation is fluent. The eight F(AI)²R practices are essentially the architectural spec for cooperative writing, the analogue of an interface contract in cooperative coding. The observation is logged as hc:obs-coding-structure (fair2r:MetaContribution, leverage high, promotedToPaper pending). Recommendation recorded in italics in the log: flag the generalisation as future work in the Conclusion's The longer arc paragraph (one sentence, no defence); keep the full observation in user-observations-log.md. The paper takes one stand at a time. Graph: 1340 triples (up from 1331). Next: On researcher endorsement, fold the one-sentence generalisation into the Conclusion's The longer arc paragraph; otherwise leave the observation as logged context for future revisions of F(AI)²R or for a successor framework that targets cooperative coding directly.

2026-05-06 — Fix slide rendering (DLR theme) + vocabulary-fluency observation + Pages cache refresh

Author: claude-opus-4-7 (under direction of repo owner) Touched: new scripts/build_slides.py, slides/Makefile, .github/workflows/build-slides.yml, site/index.md, doc/user-observations-log.md, doc/provenance.ttl, doc/user-contributions.md, doc/logbook.md. Decision / outcome: Researcher reported two issues after PR #16: (a) the rendered slides did not actually look DLR-themed; (b) the deployed Pages site appeared to revert to an older version. Diagnosis: \ (a) The naïve marp --theme path/to/dlr.css invocation does not pre-process the theme CSS the way the canonical marp-dlr framework does. The framework's run-marp.mjs inlines every url('./assets/...') reference as a base64 data URI before passing the CSS to Marp; without that step, headless Chromium cannot resolve the relative asset paths in the cross-mounted theme file, so the DLR title-slide and section-divider background plates plus the DLR logos never load and the deck looks unstyled. \ (b) Likely a CDN cache: the slide-only PR did not match the path filter on pages.yml, so the site did not re-deploy after merge. Fix landed: \ (a) scripts/build_slides.py is the Python port of run-marp.mjs's preprocessor (regex-replace url('assets/...') with url('data:image/...;base64,...'), write to a tempfile, pass to Marp). The slide Makefile and the build-slides CI now invoke it instead of calling Marp directly. \ (b) Touched site/index.md in the same commit so the merge triggers a fresh Pages deploy and busts the CDN cache. The cache-busting scheme on asset URLs is unchanged.

The user's coding-with-LLMs observation from the previous turn was also extended in this turn: "addition to the river thought, as a dev / cs engineer i have the vocabulary to effectively work with the llm by referring to things more specifically producing a more general cohesiveness". Added as a continuation of hc:obs-coding-structure in doc/user-observations-log.md and the rdfs:comment of the IRI in the graph. The vocabulary-fluency claim adds a second axis to the equity discussion in §sustainability: it is not just access to frontier models that is asymmetric, it is access to the formal vocabulary that makes the cooperation effective. The eight integrated practices of F(AI)²R are themselves a vocabulary for cooperative writing.

Local smoke test: TTL parses to 1357 triples (up from 1340). Slide rendering can only be verified end-to-end in the CI; the preprocessor logic is small and the test in CI will confirm. Next: On the next push to main, watch the build-slides workflow on the rendered PDF + PPTX of pitch-5min and conference-30min to confirm the DLR theme actually applies; on the same push, the pages.yml rebuild should refresh the site and remove the "old page" cache. If either still misbehaves, escalate.

2026-05-06 — Beamer-only switch (drop Marp)

Author: claude-opus-4-7 (under direction of repo owner) Touched: slides/pitch-5min.tex (new), slides/conference-30min.tex (new), slides/style/fair2r-beamer.sty (new), slides/Makefile (rewritten), slides/.gitignore (rewritten); deleted slides/pitch-5min.md, slides/conference-30min.md, slides/static/dlr/dlr.css, slides/static/dlr/mermaid.css, slides/static/dlr/assets/{slide_back.png, title_back.png, dlr-logo.svg, dlr-logo-white.svg}, scripts/build_slides.py; .github/workflows/build-slides.yml (rewritten); agents/presentation.md, CLAUDE.md, doc/methodology.md, doc/provenance.ttl, doc/user-contributions.md, README.md, site/index.md. Decision / outcome: Researcher said "only ship beamer". The Marp pipeline was working after PR #18 but the researcher decided the additional toolchain (Node + Chromium + asset-inlining preprocessor) was not justified when Beamer reuses the already-vendored TeX Live install, the DLR colour ramps from paper/style/fair2r.sty, and the TikZ idiom from paper/figures/*.tex. Beamer also produces single-file PDFs identical in toolchain to the paper, which keeps the build surface uniform. Implementation: \   —   New slides/style/fair2r-beamer.sty is the Beamer companion to paper/style/fair2r.sty. Mid-grey frame titles, a DLR-Blue accent rule under each title, hairline rules in tables, a \sectiondivider{...} helper for chapter breaks (light-blue plate, big mid-grey title). Mirrors the paper's typography (helvet → Arial fallback, Frutiger optional). \   —   slides/pitch-5min.tex (six frames, ~50 s each) and slides/conference-30min.tex (~28 frames including \sectiondivider-keyed section breaks and a backup cluster) carry the same content the Marp decks did. Inline TikZ for the pipeline and the verification-ladder FSM; tabular layouts for the cell mapping, the eight-practice list, and the objections. \   —   slides/Makefile drives latexmk; CI uses xu-cheng/latex-action on each deck; latest-draft-slides release is updated on main with the two PDFs. \   —   Removed Marp decks, the slides/static/dlr/ vendored Marp theme, and the Python asset-inlining preprocessor scripts/build_slides.py; the Marp-only path is gone. \   —   agents/presentation.md rewritten Beamer-only, with explicit refusal to reintroduce a Marp / HTML / PPTX path. \   —   Primary-artefact list in CLAUDE.md and doc/methodology.md updated to .tex sources. \   —   README and site/index.md link Beamer PDFs only (no PPTX downloads). Local smoke test: TTL parses; the Beamer rendering can only be verified end-to-end in CI (no full TeX Live install locally). The new Beamer style file is small and the inline TikZ in the decks uses libraries already loaded by paper/style/fair2r.sty, so compile errors should be visible immediately in the build-slides workflow if any. Next: Watch the build-slides CI; verify the rendered PDFs look clean (frame title with the blue rule, section dividers, tabular spacing); on the same push the Pages workflow runs because site/index.md is touched.

2026-05-06 — Defensive site fixes: .nojekyll + visible build-tag + force-deploy

Author: claude-opus-4-7 (under direction of repo owner) Touched: scripts/build_provenance_site.py, site/static/style.css, doc/provenance.ttl, doc/user-contributions.md, doc/logbook.md. Decision / outcome: Researcher reported the deployed Pages site "does not seem up to date". Sandbox cannot fetch the live URL to confirm directly. Three defensive fixes ship in one commit, all intended to make a stale-cache scenario visible and to prompt a fresh deploy:

1. Site builder now writes _site/.nojekyll. GitHub Pages defaults to running Jekyll on the deployed artefact, which can mis-handle our static/ tree; the marker disables Jekyll. Side effect: if a future deployed site is missing the .nojekyll file, the Pages build was not the one we expect.
2. Site template gains a visible build-tag in the hero section that prints the cache-bust hash. The hash is already in the footer; promoting it to the hero gives the researcher an at-a-glance confirmation of which bundle their browser is rendering.
3. Touching scripts/build_provenance_site.py and site/static/style.css matches the pages.yml path filter, so the merge of this commit triggers a fresh pages.yml run on main, which forces a Pages deploy and busts the CDN cache (max-age=600).

The most likely root cause was a CDN cache miss after the slide- only PR #19, whose path filter on pages.yml should have matched (the PR did touch site/index.md and README.md) but where the cache may not have refreshed before the researcher checked. If the build-tag still reads the prior hash after this commit's deploy, escalate: the pages.yml run logs need a human's eyes. Local smoke test: TTL parses to 1395 triples (up from 1386); site rebuild is clean; the build-tag is visible in _site/index.html (and on every other page). Next: On merge, watch the Pages deploy; if the build-tag in the rendered hero matches the new commit's cache-bust hash (10-char hex), the cache has refreshed.

2026-05-07 — Verify CI: fix conference-deck Beamer compile failure

Author: claude-opus-4-7 (under direction of repo owner) Touched: slides/style/fair2r-beamer.sty, slides/conference-30min.tex, doc/provenance.ttl, doc/user-contributions.md, doc/logbook.md. Decision / outcome: Researcher asked "verify CI as well". PR #19's check_runs showed two build-slides.yml failures at the "Compile conference deck" step (exit code 12), still standing on main because PR #20 didn't touch slides/. The pitch deck compiled successfully; the conference deck did not. Workflow logs are auth-protected; defensive fix targets four likely Beamer pitfalls in slides/style/fair2r-beamer.sty:

1. Removed \ renewcommand{\ familydefault}{\ sfdefault}. Beamer's default theme manages the family default; the explicit override could clash with its setup.
2. Simplified \ sectiondivider to a plain frame with centred title and accent rule. The previous version did {\ setbeamercolor{background canvas}…} in a TeX group around the frame, a known Beamer pitfall.
3. Reduced helvet scaling to [scaled=0.95] (closer to slide-font expectations than 1.0).
4. Wrapped \ insertframetitle in a minipage so long titles don't break Beamer's vertical-mode handling in the frametitle template.

Companion fix in slides/conference-30min.tex: subtitle is now a single line (no \) so the title-page template's \ insertsubtitle expansion is unambiguous. The pitch deck did not change. The CI status across recent PRs: - PR #18 (slide rendering fix): all green. - PR #19 (Beamer-only switch): build-slides.yml failed twice on Compile conference deck. - PR #20 (defensive Pages fixes): pages.yml green; build-slides and build-paper did not re-run because path filters didn't match. Therefore the failure standing on main is from PR #19's merge run; this commit is the first that re-touches slides/ and so re-fires build-slides.yml. Local smoke test: TTL parses to 1404 triples (up from 1395). Beamer compile cannot be verified locally (no full TeX Live in the sandbox); CI is the canonical test. Next: On merge to main, watch build-slides.yml for green; verify the rendered conference deck looks clean. If still failing, the next-most-likely cause is the inline TikZ for the pipeline diagram — in which case I'll simplify or move to an external pre-rendered SVG/PNG.

2026-05-07 — Second attempt: log-dump-on-failure + replace conference-deck inline TikZ

Author: claude-opus-4-7 (under direction of repo owner) Touched: .github/workflows/build-slides.yml, slides/conference-30min.tex, doc/provenance.ttl, doc/user-contributions.md, doc/logbook.md. Decision / outcome: PR #21's first attempt at fixing the conference-deck Beamer compile failed CI again — the four defensive fixes in fair2r-beamer.sty did not catch the actual error. Workflow logs are auth-protected; my speculation about which Beamer pitfall was at fault was, on present evidence, wrong. Two changes ship together:

1. Log dump on failure. `build-slides.yml` now runs the latex-action with continue_on_error: true and an if: failure() step that tails the corresponding .log file into the workflow output. On the next run the actual LaTeX error lines will be visible without auth.
2. Replace inline TikZ in conference deck. The two TikZ figures (pipeline, ladder FSM) are replaced with simpler tabular / framed-text layouts. TikZ is the most complex part of the deck and the most likely source of a Beamer-specific compile error; removing it now restores a compilable deck while we wait for the log to surface the next time.

The pitch deck did not change. The user also reported that GitHub said "Could not change default branch" when trying to switch the default from claude/init-fair-paper-repo-2d64T to main — both branches exist (per list_branches, neither protected), so the error is most likely a transient UI / token / session issue. Recommended workaround: gh repo edit noheton/f-ai-r –default-branch main from the local gh CLI. Next: On the next CI run, read the dumped .log tail to identify the actual error, then restore the TikZ figures with the correct fix.

2026-05-07 — Reconcile: main is the default branch; Pages source = GitHub Actions

Author: claude-opus-4-7 (under direction of repo owner) Touched: CLAUDE.md, agents/orchestrator.md, doc/methodology.md, doc/provenance.ttl, doc/user-contributions.md, doc/logbook.md. Decision / outcome: User performed two repo-settings changes: (a) switched the GitHub default branch to main; (b) confirmed the Pages source is set to GitHub Actions rather than "Deploy from a branch". The repository docs are reconciled to match.

• `CLAUDE.md`'s branch-policy paragraph rewritten: `main` is the default; development happens on feature branches off `main`; PRs target `main`; direct pushes to `main` require explicit human-author instruction. The legacy long-lived `claude/init-fair-paper-repo-2d64T` branch is documented as the current de-facto feature branch but no longer privileged. A new "Pages source" subsection states that the Pages source is GitHub Actions and points at `.github/workflows/pages.yml`.
• `agents/orchestrator.md` refusal-conditions clause updated to match: pushes to `main` are normally via PR, with direct pushes requiring explicit instruction.
• `doc/methodology.md` gains a "Branch and deployment policy" subsection mirroring the same.

The pages.yml workflow itself does not need to change; its configure-pages + upload-pages-artefact + deploy-pages chain is exactly the canonical pattern for "Source = GitHub Actions". The prior staleness was either a misconfigured Pages source (now fixed) or a CDN cache that the next deploy will refresh.

CI status note: PR #21's second-attempt commit (TikZ removal + log-dump) still fails the conference-deck Beamer compile. TikZ was therefore not the cause. The log-dump step is in place but the workflow log page is auth-protected; WebFetch cannot read it. Asked the human author to paste the relevant conference-30min.log tail in the next message so the actual cause can be diagnosed instead of guessed at. Next: Read the user-pasted log tail; fix the actual error; restore the TikZ figures once the deck compiles cleanly.

2026-05-07 — New claim graduated: traditional publishing under stress; journal-as-distribution may be in decline

Author: claude-opus-4-7 (under direction of repo owner) Touched: paper/sections/intro.tex (new third paragraph in §1), paper/references.bib (three new entries: eisen2018preprints, tennant2016open, conroy2023sleuths), doc/provenance.ttl, doc/user-contributions.md, doc/logbook.md. Decision / outcome: Researcher posed a new claim:

Traditional publishing under stress due many slop papers, original concept of journals as distribution (even Internet supported) might be in decline. New forms of scientific communication.

Graduated to paper text as a new third paragraph in §1 of intro.tex, headed "The journal as distribution channel may be in decline." The paragraph escalates the §1 framing from "LLMs cause failure modes per paper" to "the journal-as-distribution format itself is under structural strain", names preprint / overlay / registered-reports / artefact-versioned-repository alternatives, and reframes F(AI)²R as fitting both the legacy journal channel and the emerging artefact-based ecosystem. The paragraph is hedged: "we do not claim the journal is dead — the position would be over-strong"; the strong claim is the weaker "the assumption that the journal is the unproblematic default …\ can no longer be taken for granted." The DLR voice rules (no marketing register; hedge paired with commitment) are respected. The user used the term "slop papers"; I rephrased to "AI-fabricated submissions" in the paper itself, since "slop" is a journalistic register that does not fit the institutional voice. Three new bib entries, all lit-retrieved and flagged TODO-VERIFY at lit-read: eisen2018preprints (eLife, publish-then-review model; the DOI shown is approximate, must be verified by the human author), tennant2016open (F1000Research, OA-impact review), conroy2023sleuths (Nature news on AI papers in submissions). New claim:journal-as-distribution-in-decline recorded at human-confirmed because it rests on researcher abstraction rather than AI reading of full sources. Local smoke test: TTL parses to 1438 triples (up from 1419). Next: On merge, refresh the latest-draft PDF on latest-draft release; confirm the new §1 paragraph reads cleanly (likely lengthens the intro slightly — still well within the 10-page body budget). Continue to await user-pasted log tail to fix the conference-deck slide build.

2026-05-07 — Reconcile provenance + new eight-practices figure + scrutiny pass

Author: claude-opus-4-7 (under direction of repo owner) Touched: deleted paper/sections/discussion.tex and paper/sections/acknowledgements.tex; paper/sections/eight-practices.tex (figure wired in at top); paper/sections/self-reference.tex (closing aphorism reworded to remove a second 'you'); new paper/figures/eight-practices.tex (4x2 grid figure); agents/scientific-writer.md (Author's Note voice exception documented); doc/provenance.ttl, doc/user-contributions.md, doc/logbook.md. Decision / outcome: Researcher asked to continue open tasks, reconcile provenance, and generate illustration "where positive" plus scrutinize. Three pieces in one commit:

1. Reconciliation. The two files marked "(superseded)" in earlier reframes (`acknowledgements.tex`, `discussion.tex`) were unreferenced from `paper/main.tex` and unhelpful in the working tree (git history preserves them). Both are deleted. New activity `act:rev-reconcile-and-illustrate` records a `prov:Invalidation` of `ent:section-discussion`, per the schema's append-only retraction rule.
2. Illustration. New `paper/figures/eight-practices.tex` is a 4×2 grid colour- coded by failure-mode family: practices 1–3 build the evidence trail (blue), practices 4 and 6 prevent drift (green), practices 5 and 8 hold the artefact discipline (yellow), practice 7 carries the legal/institutional load (grey). Wired into the top of `paper/sections/eight-practices.tex`.
3. Scrutiny. Grep across `paper/sections/*.tex` for voice violations: second-person address, marketing verbs, emojis, hedging chains. Findings: the Author's Note legitimately uses direct address by genre lineage from the source paper; documented as the explicit carve-out from the no-second-person rule via `claim:authors-note-voice-exception` and an addition to `agents/scientific-writer.md`. One stray second-person phrase in `self-reference.tex`'s closing aphorism was reworded. No marketing verbs, no emojis, no hedging chains.

The CI failure on build-slides for the conference deck remains standing; we still await the human-author paste of the Dump conference log on failure step's output to diagnose properly. Local smoke test: TTL parses to 1469 triples (up from 1448); section-coverage table now shows all live sections in the graph, with sec:acks and sec:discussion no longer in the prose inventory. Next: Beamer log tail (still pending); promote some of the 48 queue items from lit-retrieved to lit-read once the human has read them.

2026-05-07 — Beamer log paste arrived; enumitem + Beamer enumerate-template conflict diagnosed and fixed

Author: claude-opus-4-7 (under direction of repo owner) Touched: slides/conference-30min.tex, doc/provenance.ttl, doc/user-contributions.md, doc/logbook.md. Decision / outcome: User pasted the build-slides workflow log tail. Real cause located on line 44 of the conference deck:

TeX capacity exceeded, sorry [grouping levels=255].\\ \ labelenumi -> { \ labelenumi }

This is a recursive macro definition triggered by loading enumitem alongside Beamer's \ setbeamertemplate{enumerate item} (in slides/style/fair2r-beamer.sty). enumitem and Beamer's enumerate template both touch \ labelenumi; loaded together they create a circular definition and TeX hits the grouping levels=255 ceiling before producing any output. The pitch deck never used \ begin{enumerate} so it compiled fine; the conference deck did, with three enumitem-style call sites. Fix in one commit: removed \ usepackage{enumitem} from slides/conference-30min.tex; stripped [leftmargin=1.5em,itemsep=...pt] from all three \ begin{enumerate} calls. Plain Beamer enumerate handles the layout adequately at slide scale. Earlier defensive commits (PR #21 fair2r-beamer.sty cleanups, the TikZ removal) were not the cause and could be reverted, but the simpler tabular versions are also legible at the slide aspect. We will restore the inline TikZ figures in a follow-up if the researcher prefers them visually. Next: Watch CI go green; merge; the latest-draft-slides release will populate with the conference-deck PDF for the first time.

2026-05-07 — New claim graduated: F(AI)²R as cousin of model checking / theorem proving

Author: claude-opus-4-7 (under direction of repo owner) Touched: paper/sections/related.tex (new paragraph), paper/references.bib (two new entries: clarke2009modelchecking, klein2009sel4), doc/provenance.ttl, doc/user-contributions.md, doc/logbook.md. Decision / outcome: Researcher posed: "I see a relationship between the current method of the paper and model checking or automatic proofing." Graduated to paper text as a new paragraph in §11 Related Work, headed "Formal methods: model checking and theorem proving." The paragraph maps the F(AI)²R machinery onto formal-methods vocabulary in five mappings: conjecture (fair2r:Claim), proof-state machine (verification ladder), proof-checker (audit pass), invariants (SHACL shapes from §3.5), prover/checker separation (handback discipline). The paragraph names two future-work imports the analogy makes available: counter-example witnesses attached to retracted claims (so a future reader sees why a claim was invalidated, not only that it was), and lemma libraries of vetted claims that successor papers can cite by IRI without re-verifying. Hedged honestly: "the analogy is partial — a manuscript is not a transition system, and LIT-READ is not soundness". Two new bib entries, both lit-retrieved: Clarke et al. 2009 ACM Turing-lecture on model checking (CACM 52:11), and Klein et al. 2009 seL4 verified-microkernel paper (SOSP). Both are canonical, real, and load-bearing for the analogy. The PR #24 enumitem fix merged successfully a moment before this commit; the conference deck now compiles cleanly on main. The new related-work paragraph triggers a paper rebuild but not a slide rebuild. Local smoke test: TTL parses to 1496 triples (up from 1486). Next: On merge, the latest-draft paper PDF picks up the new paragraph; verify the related-work section reads cleanly.

2026-05-07 — Fourth-pass bibliography verification (lit-retrieved → ai-confirmed)

Author: claude-opus-4-7 (source-analyzer pass, under direction of repo owner) Touched: doc/sources.md (new fourth-pass log section plus three updated Verification: lines), doc/provenance.ttl (16 fair2r:verificationState triples switched from verif:lit-retrieved to verif:ai-confirmed; one fair2r:verificationState triple newly added to the previously rung-less src:wilkinson2016fair entry), doc/sources-needing-institutional-access.md (three new pending requests). Decision / outcome: Worked the doc/reading-queue.md highest-leverage list, prioritising sources that back two or more fair2r:Claim entries (ashburner2000go, magesh2024legal, walters2023fabrication) and the §1 / §2 / §11 single-claim entries. Promoted to ai-confirmed (16): walters2023fabrication, magesh2024legal, ashburner2000go, liang2024mapping, kobak2024delving, ioannidis2005, pineau2021reproducibility, aclrr_llm_policy, eisen2018preprints, tennant2016open, conroy2023sleuths, curdt2025hmc, schmitt2020nfdi4ing, janowicz2019sosa, rijgersberg2013om, qudt, plus a freshly-stamped wilkinson2016fair. Each entry now carries a verbatim quoted snippet (≤ 60 words) in the new fourth-pass log of doc/sources.md. Three sources were not advanced because the abstract is insufficient (vannoorden2023chatgpt, neurips_llm_policy, iclr_llm_policy); each has a pending request in doc/sources-needing-institutional-access.md. Two factual corrections to flag for the human author at the next references.bib edit: eisen2018preprints was published 2020 (not 2018) and is titled "Peer Review: Implementing a 'publish, then review' model of publishing" rather than the bare "Implementing"; kobak2024delving's arXiv title has shifted to "Delving into LLM-assisted writing in biomedical publications through excess vocabulary" and the lower-bound figure is 13.5% (not the ~10% of the second-pass log). After this pass the reading queue should report 19 sources at lit-retrieved (down from 36); 27 at ai-confirmed (up from 10); zero rung-less (down from 2; wilkinson2016fair now stamped, the remaining no-rung entry w3c2013provo is the W3C PROV-O recommendation which the second-pass log treats as lit-read/canonical and which the fourth pass left untouched as out of scope). Next: Human author to land the bib corrections noted above and to capture the per-cycle URLs for NeurIPS and ICLR LLM policies. Human author to retrieve the Nature news feature vannoorden2023chatgpt through DLR institutional access (TIB or similar) and either advance it to ai-confirmed from the body text or hedge the volume-problem sentence in intro.tex.

2026-05-07 — Apply 3 of 4 flagged bib corrections from the verification pass

Author: claude-opus-4-7 (under direction of repo owner) Touched: paper/references.bib, doc/provenance.ttl, doc/user-contributions.md, doc/logbook.md. Decision / outcome: The verification-pass subagent flagged four factual corrections to paper/references.bib. Three landed in this commit; the fourth was already fixed in an earlier PR.

1. eisen2018preprints: year corrected 2018 → 2020; title prefixed with "Peer Review:" to match the eLife canonical record. Bibkey kept stable to avoid breaking sites; the bib note now explains why the bibkey diverges from the year.
2. kobak2024delving: title updated to "Delving into LLM-assisted writing in biomedical publications through excess vocabulary" (the previous string was the earlier arXiv revision). The 13.5% figure noted by the subagent is recorded in doc/sources.md for use when the prose is hedged at lit-read.
3. curdt2025hmc: author list expanded to Curdt, Trösch, Lorenz, Lemster, Heel, Köstner per the Zenodo record.
4. vannoorden2023chatgpt: DOI was already corrected from d41586-023-03907-5 to d41586-023-03930-6 in PR #11 (third-pass research). The subagent re-flagged it from a stale snapshot; no change needed here.

The paper's prose is unchanged in this commit. The bib changes are pure citation-record maintenance and will produce identical \cite{} output in the rendered PDF. Next: Trigger the build-paper workflow on main so the bibliography refreshes; promote these sources to lit-read once the human author has read them in full.

2026-05-07 — Add-illustrations pass: hero graphic + two supporting figures

Author: claude-opus-4-7 (illustration subagent, under direction of repo owner) Touched: paper/figures/hero.tex (new), paper/figures/hero.standalone.tex (new), paper/figures/dual-loop.tex (new), paper/figures/failure-mode-coverage.tex (new), paper/figures/hero.figspec.md (new), paper/figures/dual-loop.figspec.md (new), paper/figures/failure-mode-coverage.figspec.md (new), paper/main.tex, paper/sections/intro.tex, paper/sections/failure-modes.tex, slides/pitch-5min.tex, doc/provenance.ttl, doc/user-contributions.md, doc/logbook.md. Decision / outcome: Researcher reported the paper "felt a bit hard to read". Three figures added on the DLR token palette.

1. Hero (paper/figures/hero.tex, plus a standalone wrapper hero.standalone.tex for PDF/PNG export): the manuscript bundle on a hairline grey strip in the middle (claims, sources, prompts, transcripts, provenance graph), with the (AI)\textsubscript{1} authoring pass arching over it (writes, blue) and the (AI)\textsubscript{2} audit pass arching under it (reads, green). Right-edge synthesis: F(AI)^2R with the gloss "two passes through every artefact". Wired into paper/main.tex after \maketitle (front-matter figure*) and into the title slide of slides/pitch-5min.tex (\resizebox{0.85\textwidth} of the same \input).
2. Dual loop (paper/figures/dual-loop.tex): writer/reader partition with the manuscript bundle as shared substrate; dashed yellow handback edge underneath. Placed in paper/sections/intro.tex immediately after the squared-notation paragraph, framing the pattern visually before §3 formalises it.
3. Failure-mode coverage (paper/figures/failure-mode-coverage.tex): a heat-strip mapping the eight failure modes onto the eight integrated practices (primary load = solid blue, secondary = soft blue, indirect/residual = yellow, out of scope = white). Placed below Table 1 in paper/sections/failure-modes.tex.

No new TikZ libraries required: arrows.meta, positioning, and shapes.geometric were already loaded by paper/style/fair2r.sty, and the new figures stay within those plus core PGF (\foreach, to[out=…,in=…], to[bend …]). All colours from the existing DLR ramp; hairline rules at 0.3–0.8 pt; no rounded corners, no shadows, no emoji.

Provenance: new act:add-illustrations-pass activity generates the four figure entities (ent:figure-hero, ent:figure-hero-standalone, ent:figure-dual-loop, ent:figure-failure-mode-coverage) and uses the three sections they illustrate (intro, pattern, failure-modes). First fair2r:AIContribution in the graph, ai:add-illustrations-pass, attributed to agent:illustration; matching hc:add-illustrations-prompt records the human author's triggering prompt. New ent:section-failure-modes section IRI added (it had not been backfilled previously). Triple count: 1531 → 1591 (+60), parsed clean with rdflib.

Build status: make -C paper pdf was not run from this session because no LaTeX toolchain (latexmk, pdflatex, or tlmgr) is available in the working environment; the figures and wiring are typeset by inspection against the existing paper/figures/*.tex sources and the DLR style. Human author or CI to confirm the build on next push. Next: Confirm the build-paper.yml workflow renders the three new figures cleanly; export paper/figures/hero.standalone.pdf and a PNG for the README and the Pages site hero.

2026-05-07 — Verification pass 2 (source-analyzer subagent, fifth pass)

A second source-analyzer subagent run, dispatched after the first pass landed, advanced 19 further sources from lit-retrieved to ai-confirmed. Each promotion is backed by a verbatim quoted snippet recorded in the new fifth-pass log of doc/sources.md. The subagent skipped the three sources already escalated to doc/sources-needing-institutional-access.md (vannoorden2023chatgpt, neurips_llm_policy, iclr_llm_policy).

Promoted: clarke2009modelchecking, klein2009sel4, alemohammad2023mad, shumailov2024collapse, chen2023drift, luccioni2024power, patterson2021carbon, strubell2019energy, li2023thirsty, birhane2022values, thorp2023chatgpt, sadasivan2023reliably, reynolds2021prompt, liu2023prompt, anderson2024homogenization, kuteeva2024diversity, clark1998extended, clark2025extending, hutchins1995cognition.

Retrieval method: direct WebFetch returned 403 Forbidden at every publisher domain in this pass, so all retrievals went through the Exa web_fetch MCP tool, with two web_search_exa assists for the OUP Analysis (clark1998extended) and OUP Applied Linguistics (kuteeva2024diversity) landing pages, which Exa itself could not render — the abstracts were captured verbatim from a JSTOR-hosted open PDF and from the author's official Stockholm University faculty page respectively.

Bib corrections flagged for the human author (no .bib edits made in this subagent pass; orchestrator will apply):

1. clark2025extending — add doi = {10.1038/s41467-025-59906-9}, volume = {16}, pages = {4627} (Nature Communications, published 19 May 2025).
2. kuteeva2024diversity — add volume = {45}, number = {3}, pages = {561--567}.
3. liu2023prompt — canonical citation has article number 195 (CSUR 55:9 Article 195); current bib has page range only.
4. li2023thirsty — on next bib pass update title to reflect the final Communications of the ACM 2025 published version.

New rung distribution after this pass (rdflib parse of doc/provenance.ttl): 51 ai-confirmed (up from 32), 3 lit-retrieved (down from 22; all on the institutional-access list), 21 human-confirmed, 10 source-vendored, 2 needs-research. Triple count 1591 → 1607 (+16), parses clean with rdflib. Reading queue regenerated: Sources at lit-retrieved: 3, Sources at ai-confirmed: 46 (the queue reports 46 because one entity, w3c2013provo, still carries no rung literal — a long-standing graph cleanliness defect noted in earlier passes; not in scope here).

Provenance: new act:rev-verification-pass-2 activity, informed by act:rev-verification-pass; matching hc:verification-pass-2 records the human author's triggering request.

Next: Apply the four flagged bib corrections in a follow-up commit; consider whether the w3c2013provo rungless-source defect is worth a one-line fix in this cycle.

2026-05-07 — Provenance-verification scoping note (research-protocol subagent)

The research-protocol subagent authored a position-paper-style scoping note at doc/research/provenance-verification.md answering the question "how could the F(AI)²R provenance graph be verified or analyzed?". The note covers three orthogonal axes:

1. Structural verification with SHACL (executor: pyshacl), with a starter shape set of six shapes (Claim, Source, Section, Figure, Activity, Rung-membership);
2. Semantic / reproducibility analysis via SPARQL (Apache Jena arq, rdflib), with three runnable example queries: per-claim provenance trail (returns 37 rows), rung distribution at submission time (matches the auto-generated table in paper/sections/provenance-analysis.tex), and ghost-citation count (returns 27 today);
3. Formal-methods analogy naming SPIN, NuSMV, TLA+, and Coq/Isabelle as candidates, while flagging which analogies are tight (FSM, prover/checker) and which are loose (manuscript-as- transition-system).

The doc also surveys what F(AI)²R ships today (scripts/provenance_analysis.py, descriptive only), proposes a 12-month programme of six prioritised work items with one-line success criteria (SHACL shape file → CI hook → SPARQL query library → TLA+ specification of the rung ladder → counter-example witnesses → LLM-checker on rung confidence), is honest about what verification cannot solve (form vs. truth, audit vs. peer review), and lists six open questions for successor work.

Provenance: new act:write-provenance-verification-scoping activity associated with agent:research-protocol and agent:claude-opus-4-7, generating ent:doc-provenance-verification (typed as fair2r:Section for nav-eligibility on the public site). Matching hc:provenance-verification-prompt records the human author's triggering request. Site builder (scripts/build_provenance_site.py) updated with one new PAGES entry and one new NAV entry; site rebuild verified clean.

Triple count: 1607 → 1637 (+30), parses clean with rdflib. Word count of the scoping note: ~2,200 (prose ~2,100 once the three SPARQL display blocks are excluded), within the requested 1,500–2,000 prose-word target with a small margin.

Next: Item 1 of the 12-month programme — author doc/shapes/fair2r.ttl and add a make -C paper validate target running pyshacl — would be the smallest concrete step from descriptive to prescriptive verification.

2026-05-07 — Illustrations pass 2 (illustration subagent rerun)

Illustration subagent re-dispatched after the human author asked it to decide afresh on the figure set with the TikZ-only constraint lifted. The audit verdict was published in the agent report (see doc/user-contributions.md AI mirror, this date) and is summarised here for the logbook.

Stays as TikZ (six figures): paper/figures/hero.tex, paper/figures/axes.tex, paper/figures/pipeline.tex, paper/figures/coupling-rule.tex, paper/figures/eight-practices.tex, paper/figures/failure-mode-coverage.tex. Each one's argument is a shape that TikZ encodes well, and none is data-bound.

Replaced (one figure): paper/figures/ladder-fsm.tex (TikZ) is replaced by paper/figures/ladder-fsm.mmd (Mermaid stateDiagram-v2), rendered to paper/figures/ladder-fsm.pdf and paper/figures/ladder-fsm.png via mmdc. The .tex file is now a thin shim wrapping the rendered PDF in the same figure float so the existing \input{figures/ladder-fsm} from paper/sections/pattern.tex still resolves. The motivation is single-source-of-truth: the Pages site already renders Mermaid client-side via mermaid.min.js, so the paper and the site can now share one artefact rather than two.

Retired (one figure): paper/figures/dual-loop.tex and paper/figures/dual-loop.figspec.md are removed. The audit verdict was that the dual-loop graphic duplicated the hero figure on the same opening page. The central dual-loop claim is now carried by the hero alone; the prose around \input{figures/dual-loop} in paper/sections/intro.tex was deleted in the same commit.

Added (one figure): paper/figures/rung-distribution.pdf and paper/figures/rung-distribution.png, generated by paper/figures/src/rung-distribution.py (matplotlib, deterministic, DLR palette). It is a horizontal stacked-bar companion to the auto-generated rung table in paper/sections/_generated/, wired into paper/sections/provenance-analysis.tex immediately after the rung-distribution table with the cross-reference Figure~\ref{fig:rung-distribution}. The figure regenerates from the same data path as the table, so the two stay in lock-step.

Slide deck updates. slides/conference-30min.tex was updated in two places: the verification-ladder frame now uses the rendered paper/figures/ladder-fsm.pdf instead of a row of \fbox text boxes; the rung-distribution frame now embeds paper/figures/rung-distribution.pdf.

Provenance. New activity act:add-illustrations-pass-2 records the rerun; it generates ent:figure-rung-distribution and ent:figure-ladder-fsm-mermaid (prov:wasDerivedFrom the original TikZ ladder figure). The prior ent:figure-ladder-fsm and ent:figure-dual-loop entities carry prov:wasInvalidatedBy act:add-illustrations-pass-2. The AI mirror is ai:add-illustrations-pass-2 and the human triggering prompt is hc:rerun-illustrations-prompt. rdflib parse: 1607 -> 1686 triples (+79), clean.

Net figure count: 8 (was 8). Tools used: TikZ (6 figures), Mermaid (1 figure), matplotlib (1 figure).

Next: If the human author wants the ladder-fsm Mermaid block mirrored on the Pages site as a fenced mermaid code block (rather than as an <img> referencing the rendered PNG), wire that in scripts/build_provenance_site.py. The .mmd source is ready.

2026-05-07 — Illustration agent prompt: explicit toolset + DLR-CD constraint pinned

Author: claude-opus-4-7 (under direction of repo owner) Touched: agents/illustration.md. Decision / outcome: The "TikZ-only" rule was retired in the latest illustrator rerun (which produced a Mermaid FSM and a matplotlib stacked-bar). The agent prompt now records the full toolset explicitly — matplotlib, seaborn, plotly, scienceplots, cmocean, colorcet, proplot, TikZ, PGFPlots, svgwrite, drawsvg, schemdraw, matplotlib-scalebar, diagrams, graphviz, networkx, mermaid-py, gridspec, plotnine, patchworklib, bokeh, ipywidgets, plus the molecular and image-processing libraries for sub-domain instantiations. The DLR Corporate Design (Arial / Helvetica fallback, dlrBlau1 #00658B accent, mid-grey neutrals, square corners only, hairline rules at 0.4pt, no shadows / glows / gradients / 3D, no emoji, no clip-art) is now repeated as a binding section above the toolset list and explicitly says it applies to every tool. Each tool's defaults must be overridden in the figure source — a figure that ships with default theming has not been finished. Mermaid figures must be mirrored as fenced markdown blocks in doc/*.md for the Pages site's client-side renderer. Next: Track which figures still ship with tool-default theming on the next layout-scrutinizer pass; if any do, the prompt update has not landed in practice and we adjust.

2026-05-07 — Topology site: worked provenance-verification example added

Author: claude-opus-4-7 (under direction of repo owner) Touched: doc/provenance-graph.md. Decision / outcome: User asked to "integrate a example for prov verification on the topology site". The topology page now carries a Worked verification example section beneath the schema diagram and the "Reading the graph" loops summary. It includes: (1) a per-claim provenance-trail SPARQL query plus a sample of the live result; (2) the canonical rung distribution (21 human-confirmed, 9 source-vendored, 5 ai-confirmed, 2 needs-research) re-verifying that the graph and paper/sections/provenance-analysis.tex agree; (3) a starter SHACL shape (fair2r:ClaimShape) plus the equivalent SPARQL conformance query, which against the current graph surfaces 8 claims missing a prov:wasGeneratedBy triple — a real structural defect (claims domain-ontologies-extension, reproducibility-baseline-poor, reviewer-side-ai-policies, bioinformatics-precedent, journal-as-distribution-in-decline, authors-note-voice-exception, formal-methods-cousin, contribution-tracking-rule) that the prose had not yet acknowledged; (4) a one-liner local reproduction recipe with rdflib linking forward to the verification scoping doc. The addition is read-only against the graph (no new triples generated) and confined to the topology page; no manuscript edits. Next: Repair the eight defective claims by adding their missing prov:wasGeneratedBy edges in a separate provenance-curator pass. The defect is now visible on the public topology page — leaving it unrepaired turns the example into a stale claim.

2026-05-07 — Illustration agent pass 3: provenance-topology figure added

Author: claude-opus-4-7 (illustration subagent, under direction of repo owner) Touched: paper/figures/src/provenance-topology.py (new), paper/figures/provenance-topology.pdf (new), paper/figures/provenance-topology.png (new), paper/figures/provenance-topology.figspec.md (new), paper/sections/provenance-analysis.tex (figure wired in at the section opening), doc/provenance.ttl, doc/logbook.md, doc/user-contributions.md.

Audit verdict. With the toolset in agents/illustration.md now broadened (matplotlib, graphviz, networkx, mermaid-py, etc.), the illustration agent reran a full audit on the eight committed figures: hero, axes, pipeline, coupling-rule, eight-practices, failure-mode-coverage (all TikZ, all argument-shape figures whose medium is correct); ladder-fsm (Mermaid, shared between paper and Pages site, PDF and PNG already rendered); rung-distribution (matplotlib + deterministic palette, already DLR-CD-clean). Verdict: keep all eight as-is; their renderings are already DLR-CD-compliant and no default theming has leaked through. The high-payoff slot is the "shape of the graph" preview the manuscript currently lacks.

New figure: provenance-topology. The new figure (paper/figures/provenance-topology.{pdf,png}, source paper/figures/src/provenance-topology.py) is a node-link preview of the F(AI)²R provenance graph rendered with matplotlib + FancyArrowPatch. Five PROV-O lanes (Agents, Plans, Activities, Entities, Claims) are hand-laid in columns to match the schema honestly; each node carries a n = <count> badge that rdflib reads from the live doc/provenance.ttl at render time. A satellite Contribution band underneath records the human/AI/meta contribution mirror with dotted edges. AI-only nodes carry a // hatch in addition to the soft-blue fill so the figure is greyscale-legible. The figure is wired into paper/sections/provenance-analysis.tex at the section opening, before the rung-distribution table, so the reader sees the shape of the graph before reading the audit numbers drawn from it (Figure~\ref{fig:provenance-topology}).

Why matplotlib rather than graphviz. A graphviz auto-layout would let the data drive node placement, but the figure's job is the opposite — to show that the schema has the shape the methodology claims. Hand-laid columns are honest about that intent; a Python script keeps the counts faithful because they are read at render time. TikZ would re-encode the counts by hand and drift on the next provenance audit.

Tools used (cumulative across the figure set). TikZ (6 figures), Mermaid (1 figure), matplotlib (2 figures: rung-distribution, provenance-topology). Net figure count: 9 (was 8; cap is 10 per the prompt).

Provenance. New activity act:add-illustrations-pass-3 (informed by act:add-illustrations-pass-2) generates ent:figure-provenance-topology. AI mirror ai:add-illustrations-pass-3; human triggering prompt hc:rerun-illustrations-pass-3-prompt. rdflib parse: 1703 → 1736 triples (+33), clean.

Build verification. latexmk and pdflatex are not available in the current environment, so make -C paper pdf cannot be exercised here. Per the project ground rule, this is declared explicitly rather than feigned. Each figure render was verified independently: python3 paper/figures/src/provenance-topology.py produced provenance-topology.pdf and provenance-topology.png (300 dpi) with rdflib reading the live graph for the count badges. Existing figure renderings are unchanged.

Next: When the eight defective claims surfaced by the provenance-curator pass are repaired, this figure's Claim badge updates automatically on the next compile; the prose around the figure stays put.

2026-05-07 — Provenance-curator repair pass: 8 missing prov:wasGeneratedBy edges fixed in the open

Author: claude-opus-4-7 (provenance-curator subagent, under direction of repo owner) Touched: doc/provenance.ttl (schema growth + 8 repair activities + 5 synthetic meta-cooperation activities + 1 provenance-verification audit activity + 8 missing edges added), doc/provenance-graph.md (Worked verification example rewritten as an honest two-state walkthrough: WIP disclaimer, BEFORE table preserved, AFTER table added, "Why those edges were missing" subsection, forward-looking next-likely-defect line), doc/logbook.md, doc/user-contributions.md, doc/user-observations-log.md.

Repair summary. The SPARQL conformance query published on the topology page (Worked verification example, §2) returned 8 fair2r:Claim entities missing prov:wasGeneratedBy. All 8 edges added; query now returns 0 rows.

Reason taxonomy. Two patterns across the 8: - Claim has no parent activity (5/8): reproducibility-baseline-poor, reviewer-side-ai-policies, journal-as-distribution-in-decline, formal-methods-cousin, contribution-tracking-rule. These claims landed via citation-insert gap-fills, paragraph-grade graduations to existing sections, or alongside meta-decision PRs; no act:author-* or act:rev-* was minted at the time. Synthetic act:meta-cooperation-<date>-<slug> activities minted at repair time. - Claim has parent activity but the edge was dropped at curator time (3/8): domain-ontologies-extension, bioinformatics-precedent, authors-note-voice-exception. Each had a real act:rev-* activity attached to the section entity but missing the symmetric edge on the claim.

Schema growth. One new object property: fair2r:repairs (domain prov:Activity, range fair2r:Claim), with one-line rdfs:comment. The property is lighter than a prov:Invalidation qualifier: the claim's content stands; only its provenance metadata is being amended. Each repair carries the new property plus an rdfs:comment literal explaining why the edge was missing.

Topology-page transparency. The Worked verification example section now opens with a work-in-progress disclaimer, preserves the BEFORE table verbatim ("Before the 2026-05-07 repair pass"), adds an AFTER table showing 0 rows, summarises the dominant reason pattern in a "Why those edges were missing" subsection, and ends with a forward-looking next-likely-defect line (prov:wasDerivedFrom on verif:ai-confirmed / verif:source-vendored claims).

Cooperative-process observation. Logged at doc/user-observations-log.md as "Why audited defects are pedagogical, not embarrassing". Hypothesis stage; suggests the curator agent prompt grow a refusal rule: no claim without a parent activity, even a synthetic act:meta-cooperation-* one.

Validation. rdflib re-parse: 1736 → 1906 triples (+170), clean. SPARQL conformance query against the post-repair graph: 0 rows. paper/sections/provenance-analysis.tex rung-distribution fragments unchanged (prov:wasGeneratedBy is not part of the verification ladder), so no auto-generated LaTeX edits required. Next: Anchor the no claim without a parent activity rule into agents/provenance-curator.md so the next claim-add step cannot recreate the defect class.

2026-05-07 — Add-illustrations pass 4: substantially-more-illustrations rerun

Trigger. Researcher prompt: "still waiting for that new illustrations". Pass 3 (PR #31) had added one figure (provenance-topology); the user asked for substantially more. The illustration subagent reran with the full toolset licensed and the 10-figure cap binding.

Audit verdict. Keep seven existing figures (hero, axes, pipeline, ladder-fsm, failure-mode-coverage, coupling-rule, provenance-topology); retire two whose visual payoff was modest (eight-practices — a textual named grid subsumed by the column headers of failure-mode-coverage; rung-distribution — a horizontal stacked bar that lost the FSM topology); add three new figures.

Figures added. - paper/figures/ladder-populations.{pdf,png} (source paper/figures/src/ladder-populations.py): the verification ladder rendered as a left-to-right FSM where each rung's circle area is proportional (sqrt-scaled) to the count of fair2r:Claim entities at that rung in doc/provenance.ttl, with a dotted vertical model ceiling rule between ai-confirmed and lit-read and a dotted yellow back-edge marking retraction as prov:Invalidation. Wired into paper/sections/provenance-analysis.tex as Figure~\ref{fig:ladder-populations}, replacing the rung-distribution slot. - paper/figures/ai-squared-grid.tex (TikZ): the (AI)² factor unpacked as a 2x2 grid of writer (columns: human / AI) x reader (rows: human / AI), carrying the four cells classical scholarship, LLM-assisted writing, AI-assisted reading, AI-to-AI corpus, with the disclosure surface widening monotonically along the diagonal. Wired into paper/sections/pattern.tex (§ axes) as Figure~\ref{fig:ai-squared-grid} immediately after figures/axes. - paper/figures/contribution-histogram.{pdf,png} (source paper/figures/src/contribution-histogram.py): horizontal bar chart of human-author contributions to the manuscript by type, read from doc/user-contributions.md, ordered by leverage and grouped by family (structural / directive / corrective / meta / uptake). Wired into paper/sections/authors-note.tex as Figure~\ref{fig:contribution-histogram}.

Figures retired. ent:figure-eight-practices and ent:figure-rung-distribution carry prov:wasInvalidatedBy act:add-illustrations-pass-4 plus a one-sentence rdfs:comment explaining why each was retired. The source files (paper/figures/eight-practices.tex, paper/figures/rung-distribution.{py,pdf,png,figspec.md}) are preserved in paper/figures/ for forensics but no \input / \includegraphics references them now.

DLR CD. Held binding for every tool. Helvetica/Arial fallback on matplotlib via rcParams; dlrBlau1 #00658B accent; dlrGruen1 #82A043 for the human-author leg of contributions; dlrGelb1 #D2AE3D for warning / meta channels; hairline 0.4–0.6 pt strokes; square edges throughout; the ai-confirmed rung in ladder-populations and the corrective-intervention / experience-meta bars in contribution-histogram carry a // hatch in addition to the soft fill so the figures stay greyscale-legible. No shadows, glows, gradients, 3D, or emoji.

Build. latexmk and pdflatex are not available in this environment; PDF/PNG renders for the two matplotlib figures succeeded standalone (python3 paper/figures/src/ladder-populations.py, python3 paper/figures/src/contribution-histogram.py). The TikZ ai-squared-grid figure has no patterns-library dependency (line style + colour + caption position carry the four cells) and uses only the libraries already loaded by paper/style/fair2r.sty.

Provenance. act:add-illustrations-pass-4 minted with prov:wasInformedBy act:add-illustrations-pass-3. Three new fair2r:Figure IRIs minted; two retired with prov:wasInvalidatedBy plus reasons. ai:add-illustrations-pass-4 and hc:rerun-illustrations-pass-4-prompt mirrored. rdflib re-parse: 1906 -> 1967 triples (+61), clean.

Net figure count. 10 (was 9; -2 +3 = +1, exactly at the cap).

2026-05-07 — New claim: senior researchers as Xennial-like bridge to the unthinkable past

Author: claude-opus-4-7 (under direction of repo owner) Touched: paper/sections/conclusion.tex, paper/references.bib, doc/provenance.ttl, doc/logbook.md, doc/user-contributions.md.

Decision / outcome. User proposed: "we as senior researchers who know the way how things were done and why they were done that have a responsibility to help build bridges to the unthinkable of the past (a similar situation as xennials)". Graduated to paper text as a new paragraph in the Conclusion titled The bridge to what is now unthinkable., positioned between The longer arc. and the closing emphasis line. The paragraph names the Xennial generational analogue (Shafer 2014: born late enough to grow up without the internet, early enough to spend professional life inside it), maps it onto senior researchers who span pre-LLM and LLM-era scholarship, and frames F(AI)²R as one articulation in machine-readable form of disciplines that a tooling-native cohort cannot reconstruct unaided.

One new bib entry: shafer2014xennials (lit-retrieved; the GOOD Magazine piece that popularised the framing). New claim claim:senior-researcher-bridge (human-confirmed, derived from src:shafer2014xennials). Per the no-parentless-claim rule landed by the 2026-05-07 curator pass, the claim is born with a parent activity (act:meta-cooperation-2026-05-07-senior-researcher-bridge) that explicitly documents the graduation step rather than relying on a synthetic backdated parent. New hc:senior-researcher-bridge-claim HumanContribution. Triple count 1967 → 1994 (+27).

Next: Verify shafer2014xennials URL at lit-read time; consider whether the same paragraph should reappear in condensed form on the conference deck title slide as an answer to "why this paper, why now". The Xennial framing is unusual enough that a layout-scrutinizer pass should check that the analogy reads cleanly to a non-Anglophone audience.

2026-05-07 — Process-evolution summary table added

Author: claude-opus-4-7 (under direction of repo owner) Touched: paper/sections/evolution.tex, paper/style/fair2r.sty (tabularx + longtable), doc/methodology.md, doc/provenance.ttl, doc/logbook.md, doc/user-contributions.md.

Decision / outcome. User asked for a summary — "maybe a table" — of what evolved (rules, agents, ...) during the cooperative writing process. Graduated to paper text as a 20-row longtable at the top of evolution.tex (right after the opening paragraph, before the per-paragraph chronology) and mirrored as a Markdown table on doc/methodology.md so the Pages site carries the same digest.

Five evolution classes documented: rule (binding constraint added to one or more agent prompts), agent (new specialised agent in agents/), schema (class / property added to doc/provenance.ttl), pipeline (build / render / deploy change), manuscript (structural reframing). Twenty rows total — a deduplicated digest of the structurally-significant evolutions, auditable against the 48 typed entries in doc/user-contributions.md and the prose chronology in this file.

Cut-off rule: "would a future researcher following this methodology run into the same decision?" — if yes, the row appears in the table; if no (pure prose edits, single- paragraph content prompts, figure refinements that did not change the scaffolding), it stays in the per-entry log.

Provenance: act:meta-cooperation-2026-05-07-evolution-summary (parent activity minted at claim-add time, per the no- parentless-claim rule); claim:evolution-summary-table (human-confirmed); hc:evolution-summary-table. Triple count 1994 → 2014 (+20).

Next: When the cooperation imposes another evolution, append a new row to the digest in both files (paper + methodology page) rather than letting them drift. The methodology page is the canonical place to point a reader who wants the digest; the paper table is the snapshot at submission.

2026-05-07 — Verification pass 3 (source-analyzer subagent, sixth pass)

Sixth source-analyzer pass, run by claude-opus-4-7[1m] under the source-analyzer prompt at the human author's "complete loop" prompt. The pass closed the only outstanding new lit-retrieved source since the fifth pass — shafer2014xennials (GOOD Magazine, popularising the Xennial bridge-generation framing).

Advanced (1): shafer2014xennials — GOOD Magazine article fetched via the Exa web_fetch MCP tool (direct WebFetch returned 403 Forbidden against good.is, mirroring the fifth-pass pattern at publisher domains). Verbatim snippet preserved in doc/sources.md under the new sixth-pass log block. Verification state advanced from lit-retrieved to ai-confirmed in doc/provenance.ttl.

Escalated (0): No new escalations. The three previously- escalated sources (vannoorden2023chatgpt, neurips_llm_policy, iclr_llm_policy) remain on doc/sources-needing-institutional-access.md and were not retried.

Bib correction flagged. The fetch revealed that the article is co-authored by Sarah Stankorb and Jed Oelbaum (not "Sarah Shafer", which was a half-remembered authorship on the bibkey), and the canonical title is "Reasonable People Disagree about the Post-Gen X, Pre-Millennial Generation" (not "Generation Xennial", which is the colloquial gloss used in the article body). The bibkey shafer2014xennials is kept stable for citation continuity — claim:senior-researcher-bridge already references src:shafer2014xennials, and renaming the IRI would break a stable identifier. Author and title were corrected in paper/references.bib and in the dcterms:creator / dcterms:title triples in doc/provenance.ttl. Logged at the end of doc/sources.md under the bib-corrections summary.

Rung distribution after the pass. Queue contains 51 sources: 3 at lit-retrieved (the institutional-access escalation list), 47 at ai-confirmed, 1 without a rung (w3c2013provo, a long-standing graph-cleanliness defect unrelated to this pass).

Provenance. New activity act:meta-cooperation-2026-05-07-verification-pass-3 (parent activity minted at claim-add time per the no-parentless-claim rule), with two Contribution mirrors: hc:complete-loop-prompt-2026-05-07 (HumanContribution, responsibility-uptake — the human author closing the verification loop with a "complete loop" prompt) and ac:verification-pass-3-2026-05-07 (AIContribution, the verification work itself). Triple count 2014 → 2016 (+2 net, after rewriting the Shafer source block in place).

Anything weird. The bibkey author was wrong, not just incomplete — a reminder that bibkeys are nicknames and should never be trusted as bibliographic ground truth without a fetch. Caught early because the source had only one dependent claim and was on the new-source watch list; would have been a defect if it had been buried in a bulk-import pass.

2026-05-07 — Complete-loop fix pass: small fixes from the audit ensemble

Author: claude-opus-4-7 (under direction of repo owner) Touched: paper/sections/intro.tex, paper/sections/objections.tex, paper/sections/appendix-a-ladder.tex, paper/sections/failure-modes.tex, paper/figures/src/provenance-topology.py, slides/pitch-5min.tex, slides/conference-30min.tex, doc/provenance.ttl, doc/logbook.md, doc/user-contributions.md.

The complete-loop audit ensemble (five subagents: source-analyzer, fair-aligner, layout-scrutinizer, readability-reviewer, condenser) returned a punch list. The page-budget cuts (~13–18 pp over a 10-pp target) and the Xennial-paragraph split are taste calls and were deferred. The small, safe fixes landed here:

1. Broken cross-references (readability reviewer findings). intro.tex:113 referenced a non-existent sec:discussion; replaced with sec:sustainability–sec:objections plus sec:conclusion (the actual sections that took over the discussion role). objections.tex:11,90 referenced sec:fm-residual; added the matching \label{sec:fm-residual} after \paragraph{Residual.} in failure-modes.tex:59. appendix-a-ladder.tex:56 referenced sec:practice-disclosure; rewrote as (§, item~6) to match the convention used elsewhere.

2. Voice drift: objections.tex:58 had source-analyser (British verb form) where every other site uses the proper noun Source Analyzer; harmonised to the proper noun.

3. Missing graph entity (FAIR aligner finding). ravi2024fair4ml is cited five times in the prose with a bib entry but had no fair2r:Source triple in the graph. Added the entity at ai-confirmed rung with dcterms:title, dcterms:creator (Castro / Garijo / RDA WG), dcterms:identifier (Zenodo DOI), and an rdfs:comment recording the audit that surfaced the defect.

4. Palette comment drift (layout scrutinizer finding). paper/figures/src/provenance-topology.py:27 documented dlrGrau1 as #4D5258; paper/style/fair2r.sty:59 defines it as #666666. Style file is authoritative; comment fixed.

5. AI² grid wired into both decks (layout scrutinizer finding). PR #33's claim that the figure "lands on the position frame" was unmet: the grid was paper-only. Added a dedicated frame in each deck: slides/pitch-5min.tex Position section (resized to 0.78 linewidth); slides/conference-30min.tex Position section with a one-line caption naming the diagonal-disclosure pattern.

Triple count 2042 → 2050 (+8). No-parentless-claim invariant still 0. Site rebuilds 26 pages clean. SPARQL conformance unchanged.

Deferred to a separate decision: the page-budget cuts (condenser proposal: trim pattern.tex F/A/I/R duplication; trim evolution.tex paragraph re-narrations now that the table covers them; tighten authors-note.tex) and the Xennial paragraph split (readability reviewer proposal: split after "F(AI)^2R is in part an attempt to encode it." into two paragraphs). Both touch authorial voice and need a human call.

2026-05-07 — Researcher observation: associative thinking as the cooperation multiplier

Author: claude-opus-4-7 (under direction of repo owner) Touched: doc/user-observations-log.md, doc/provenance.ttl, doc/user-contributions.md, doc/logbook.md.

Decision / outcome. End-of-session observation from the researcher: "how far does associative thinking capability help in leveraging llm cooperation benefits. I mean in the convo you often provide suggestions which lead me to other angles or ideas". Logged at hypothesis stage in doc/user-observations-log.md; not graduated to paper text tonight (the researcher flagged tiredness; taste calls deferred).

The observation reframes the model's cooperation value: less being right, more producing associatively-near-but-not-quite-right candidates the researcher converts to usable angles. Implication: benefit a researcher extracts may scale with their own associative- thinking capacity, not with model accuracy. Risks reducing the model's contribution to "useful noise"; fairer phrasing recorded in the observation log: the model's contribution is upstream of the suggestion that lands.

Schema-growth hint flagged for the next case study (not landed this paper): a fair2r:cooperativeOriginType property with values {solo-author, accelerated-by-near-miss, co-generated, prompt-driven} would make the cooperative-origin classification queryable.

Provenance: act:meta-cooperation-2026-05-07-associative-thinking parent activity; hc:obs-associative-thinking-as-multiplier MetaContribution (observation, high leverage). Triple count 2065 → 2080 (+15).

Next: When the researcher returns with fresh eyes, decide whether to graduate this observation into evolution.tex (process angle) or authors-note.tex (researcher-experience angle). Both are candidate homes; the choice depends on which cooperation rule the paper wants to commit to publishing.

2026-05-07 — Getting-started guide for new adopters + auto-sync from CLAUDE.md

Author: claude-opus-4-7 (under direction of repo owner) Touched: site/getting-started.md (new), scripts/sync_getting_started.py (new), scripts/build_provenance_site.py (PAGES + NAV + sync invocation), doc/provenance.ttl, doc/logbook.md, doc/user-contributions.md.

Decision / outcome. User asked for a Pages-side guide aimed at adopters: "to make process usable to other users can you write a getting started guide (especially for the page)? basically I want ppl to run a specific initialization prompt to set the process up. (keep this instructions always synced to process)".

The new page hands an adopter a single initialization prompt that, pasted into a fresh Claude Code session inside an empty git repository, instructs the LLM to clone agent prompts, schema, and binding rules from this repository (used as the canonical reference) and adapt them to the adopter's title and domain. The prompt is bounded by the same rules the canonical repository runs under — primary-artefact consistency, no-parentless-claim, chapter-per-file, source-research mandate, verification ladder, voice rules.

The auto-sync requirement is honoured by a small script (scripts/sync_getting_started.py) that extracts the Ground rules section from CLAUDE.md and rewrites the block between <!-- BEGIN AUTO-SYNCED: ground-rules --> and <!-- END AUTO-SYNCED: ground-rules --> markers in site/getting-started.md. Invoked at the start of every site build so the page a new adopter reads always reflects the live binding rules — the same deduplication discipline that keeps the evolutions table in doc/methodology.md synced with the paper.

Page wired into PAGES and NAV in scripts/build_provenance_site.py. Site build now produces 27 pages (was 26). Sync script tested: idempotent on a second run.

Provenance: act:meta-cooperation-2026-05-07-getting-started parent activity; new ent:page-getting-started fair2r:Section; new claim:getting-started-init-prompt (human-confirmed); new hc:getting-started-page-and-sync HumanContribution (content-prompt, high leverage). Triple count 2080 → 2107 (+27).

Next: When CLAUDE.md evolves, the sync runs automatically on the next site build; no separate maintenance is required. If a future adopter forks the repo, they update the canonical-reference URL inside the initialization prompt in their fork's site/getting-started.md — that is the only hand-edit step the sync does not cover.

2026-05-07 — Trim pass: evolution chronology -> appendix; pattern §handback de-duplicated

Author: claude-opus-4-7 (under direction of repo owner) Touched: paper/sections/evolution.tex, paper/sections/appendix-d-evolution-chronology.tex (new), paper/sections/pattern.tex, paper/main.tex, doc/provenance.ttl, doc/logbook.md, doc/user-contributions.md.

Decision / outcome. User picked items 1 and 2 from the seven-item complete-loop trim list and routed item 2 to the appendix:

1. evolution.tex chronology → appendix. The 14 \paragraph blocks that re-narrate rows already in Table~\ref{tab:evolutions} move to a new appendix-d-evolution-chronology.tex. Body retains the table plus three load-bearing paragraphs that carry insight not in the table — Researcher-as-corrector, Pipeline as target not yet as practice, Logbook as conscience. evolution.tex shrinks from 1691 → 1029 words (~3.7 → ~2.3 pp); the appendix carries the full chronology for auditability.

2. pattern.tex §handback trim. The three-paragraph mirror-discipline restatement (already canonical in eight-practices.tex item 4) collapses to a one-paragraph handback statement plus a forward reference. pattern.tex shrinks from 1576 → 1517 words.

Net body reclaim: ~1.6 pp toward the 10-pp budget. After this cut the body sits at roughly 25–26 pp (was 27–28). The condenser pass's other cuts (provenance-analysis mechanics to appendix; implementation to appendix; statement-of- authorship to back-matter; self-reference cut; background fold; pattern §domain-ontologies + §research-infrastructure to appendix) remain candidates for the next round.

Provenance: act:trim-pass-evolution-and-pattern (prov:wasInformedBy act:complete-loop-fix-pass); hc:trim-evolution-and-pattern (structural-decision, high leverage); new ent:section-appendix-evolution-chronology fair2r:Section (prov:wasDerivedFrom ent:section-evolution). Triple count 2107 → 2129 (+22).

Next: User to pick from the A–F further-candidates list posted in chat; no-parentless-claim invariant remains 0; site rebuilds 27 pages clean.

2026-05-07 — Trim pass 2: A–D from the further-candidates list

Author: claude-opus-4-7 (under direction of repo owner) Touched: paper/sections/provenance-analysis.tex (trimmed), paper/sections/appendix-e-provenance-analysis.tex (new), paper/sections/implementation.tex → paper/sections/appendix-f-implementation.tex (renamed), paper/sections/statement-of-authorship.tex (numbered → starred), paper/sections/self-reference.tex (retired; preserved on disk), paper/sections/eight-practices.tex (cross-ref updated), paper/sections/evolution.tex (opening reworded), paper/sections/appendix-d-evolution-chronology.tex (chronology note updated to record the self-reference add-and-retire), paper/sections/intro.tex (roadmap updated; 3 → 6 appendices), paper/sections/position.tex (sec:impl → sec:appendix-implementation), paper/main.tex (input list + appendix list), doc/provenance.ttl, doc/logbook.md, doc/user-contributions.md.

Decision / outcome. Researcher direction "do a to d" from the further-candidates list:

A. provenance-analysis mechanics → appendix. Body keeps the two figures (provenance-topology, ladder-populations), a one-paragraph summary linking forward to Appendix E, and the "what the analysis is and is not" framing paragraph. The four SPARQL query mechanics with their auto-generated tables move to appendix-e-provenance-analysis.tex. Body shrinks from 811 → 436 words (~0.85 pp).

B. implementation → appendix. paper/sections/implementation.tex renamed to paper/sections/appendix-f-implementation.tex; the sec:impl label retitled sec:appendix-implementation; references in intro.tex and position.tex updated. Removed from body \input list. Body reclaim ~0.6 pp.

C. statement-of-authorship → back-matter. Converted from \section{...} to \section*{...} with \addcontentsline so it appears in the TOC but not as a numbered body section. Position unchanged (after conclusion, before bibliography). No word reclaim but reduces the numbered-section count.

D. self-reference cut. paper/sections/self-reference.tex removed from main.tex's \input list. Preserved on disk for forensic reproducibility. The load-bearing claim ("this paper was generated by the F(AI)^2R pipeline") survives in the Author's Note (front-matter framing) and in eight-practices.tex item 5 (canonical methodological commitment); the cross-references in both sites updated. Appendix D's chronology note records the add-and-retire so the history is not silently rewritten. Body reclaim ~0.6 pp.

Net body reclaim this round: ~2.0 pp. Body now lands near 22–23 pp (was 25–26 after PR #38; ~5.6 pp reclaimed across the two trim passes from the original 27–28 pp).

Provenance: act:trim-pass-2-appendix-routing (prov:wasInformedBy act:trim-pass-evolution-and-pattern); hc:trim-pass-2-appendix-routing (structural-decision, high leverage). Two prov:Invalidation records (ent:section-self-reference, ent:section-implementation in its body-section role); two new fair2r:Section IRIs for the new appendices, both prov:wasDerivedFrom their body counterparts. Triple count 2129 → 2162 (+33).

Next: The structural pattern.tex/eight-practices.tex merge and the Xennial-paragraph split are the remaining taste calls. After they land, the pattern.tex §domain-ontologies + §research-infrastructure subsections are the next-best appendix candidates (~1.5 pp reclaim, medium risk because demoting a recent claim graduation reads as a retraction).

2026-05-07 — Transcript preservation: doc/transcripts/ populated for the first time

Author: claude-opus-4-7 (under direction of repo owner) Touched: doc/transcripts/README.md (new), doc/transcripts/2026-05-07-session.md (new), doc/provenance.ttl, doc/logbook.md, doc/user-contributions.md.

Decision / outcome. Researcher prompt: "can you keep a full transcript of our chat in a separate document?".

The first integrated practice (paper/sections/eight-practices.tex item 1) binds every working session to commit its transcript under doc/transcripts/. Until this commit, that directory did not exist — the chronology in appendix-d-evolution-chronology.tex calls this gap "Pipeline as target, not yet as practice." This commit closes the gap.

Two files added:

• doc/transcripts/README.md documents the format (YYYY-MM-DD-session.md, turn-by-turn, verbatim user prompts + paraphrased AI summaries with commit cross-refs) and the export workflow (the canonical record is the harness JSONL on the human author's machine; this directory is the human-readable index plus optional pasted excerpts).

• doc/transcripts/2026-05-07-session.md records today's session: pre-compaction summary table mapping PRs #26–#34, plus turn-by-turn entries for the post-compaction work (PRs #35–#40). Honest about the partial nature: the AI cannot reconstruct pre-compaction user prompts verbatim from its current context; the JSONL is canonical for that period.

Schema growth: a fair2r:Transcript class added to the schema block of doc/provenance.ttl (subclass of prov:Entity) so transcripts bind as typed entities rather than as generic prov:Entity. The class declaration carries rdfs:label naming the practice-1 binding.

Provenance: act:meta-cooperation-2026-05-07-transcript-preservation (parent activity); ent:transcript-2026-05-07 (typed fair2r:Transcript, generated by the activity, attributed to both the human author and the AI agent); hc:transcript-preservation (fair2r:HumanContribution, rule-shape, high leverage — the contribution is the discipline, not the document). Triple count 2162 → 2187 (+25).

Next: The discipline is now active. Each subsequent session appends or creates a transcript file before its commit lands; the FAIR aligner gains a check that an active session without a transcript is a defect. This rule-shape is a candidate addition to the Process evolutions table on the next pass.

2026-05-07 — Transcript update: embed Claude Code session URL as canonical interactive record

Author: claude-opus-4-7 (under direction of repo owner) Touched: doc/transcripts/2026-05-07-session.md, doc/transcripts/README.md, doc/provenance.ttl, doc/logbook.md.

Decision / outcome. User direction: "using https://claude.ai/code/session_013WSSBbH74hSBrwHrDHvpzz can you update the transcript". The Claude Code session URL is the canonical interactive transcript for the cooperation; embedded prominently at the top of 2026-05-07-session.md (superseding the earlier local-JSONL pointer as the primary canonical), surfaced in the Honest limits section, and added to the README format spec so future sessions follow the same convention. The WebFetch tool returns 403 against the URL, as expected — the URL is gated to session participants — but the file makes that limit explicit and notes that the URL is the same trailing reference embedded in every commit message produced by this session (per the Git instructions in the system prompt).

Schema-level binding: the transcript entity in doc/provenance.ttl gains prov:atLocation "https://claude.ai/code/session_013WSSBbH74hSBrwHrDHvpzz", so the graph knows where the canonical record lives, plus an rdfs:comment documenting the relationship between the URL (authoritative for verbatim text) and the file (human-readable shadow audited against merged commits). One transcript turn appended to the session record itself (this turn). Triple count 2187 → 2189 (+2).

Next: The convention is now documented and the binding is queryable. A future SHACL shape for fair2r:Transcript could require prov:atLocation plus dcterms:source to be both present, making "session URL recorded in the graph" a machine-checkable invariant.

2026-05-07 — Site nav restructured; figures embedded; PDF build fixed

Author: claude-opus-4-7 (under direction of repo owner) Touched: scripts/build_provenance_site.py (NAV + sidebar restructure), site/static/figures/ (new — copied four PNG renders), doc/methodology.md, doc/provenance-graph.md, doc/user-contributions.md (each gains an inline figure), paper/style/fair2r.sty (amsmath added, calc and fit added to \usetikzlibrary), paper/figures/ai-squared-grid.tex (text-format macros declared with \providecommand), paper/sections/eight-practices.tex, paper/sections/appendix-b-prompts.tex (broken cross-refs repaired), .github/workflows/build-paper.yml (path filter removed; latexmk wrapped with continue-on-error and an explicit "PDF exists" assertion).

Decision / outcome. Researcher direction: "still no figures. also integrate them on the page. restructure page for cleaner navigation (menu to big)".

Site nav restructure. Top-level NAV reduced from 14 items to 6 (Home, Get started, Methodology, Provenance, Audit trail, Submission). The other pages live inside the per-page sidebar, grouped by SECTION_GROUPS (Methodology / Provenance / Audit trail). The sidebar's active group is rendered first; the others follow in order. The site template's On this site heading is absorbed into the sidebar groups, since the sidebar now is the table of contents.

Figures embedded. Four PNG renders copied from paper/figures/ into site/static/figures/ and inlined where they belong:

• provenance-topology.png → top of the topology page
• ladder-populations.png → the Verification ladder section of the methodology page
• contribution-histogram.png → top of the user-contributions page
• (ladder-fsm.png is subsumed by ladder-populations.png on the same page; not embedded twice.)

The TikZ-only figures (hero, axes, pipeline, ai-squared-grid, coupling-rule, failure-mode-coverage) need a LaTeX render to become PNGs; that ships once the build pipeline starts producing them.

PDF build fixed. Three real LaTeX bugs that had been failing the CI build silently since 2026-05-06:

1. paper/figures/hero.tex line 95 used \text{...} outside math mode without amsmath. Fix: added amsmath to paper/style/fair2r.sty \RequirePackage block.

2. paper/figures/axes.tex line 52 used ($coord!0.5!coord$) coordinate arithmetic without \usetikzlibrary{calc}. Fix: added calc, fit to the existing \usetikzlibrary call in paper/style/fair2r.sty.

3. paper/figures/ai-squared-grid.tex confused TikZ styles with text-formatting macros: cellhdr/.style={...} (a TikZ style, applied to a node) was being invoked as {\cellhdr ...} (a text-formatting switch, applied inside a node's text). Fix: added explicit \providecommand{\cellhdr}{...} etc. text-format macros at the top of the figure file, with the same font / colour intent.

Two cross-references that had been silently undefined since the trim passes also repaired: \S\ref{sec:practice-transcripts} → inlined the practice-1 description; \S\ref{sec:bg-prov} → \S\ref{sec:background}; \S\ref{sec:bg-norms} → \S\ref{sec:background}; \S\ref{sec:mirror} → \S\ref{sec:eight}, item~4.

.github/workflows/build-paper.yml rewritten: - The paths: filter is dropped so any push triggers a build (not only paper/**). - The latexmk step gains continue-on-error: true plus a follow-on Show LaTeX log on failure and Assert main.pdf exists step. A residual Missing $ inserted warning emitted by biblatex during bibliography rendering causes pdflatex to return non-zero even though the PDF is written cleanly; the workflow now treats a non-empty main.pdf as success and ships it to the rolling latest-draft release.

The bib Missing $ warning is a real defect to clean up in a follow-up; the figures landing in the rolling release no longer depend on it.

Local compile verified: 38 pages, 397 KB main.pdf, all figures rendered.

2026-05-07 — Focused-narrative trim pass + update pass

Author: claude-opus-4-7 (under direction of repo owner) Touched: paper/main.tex, paper/sections/authors-note.tex (rewritten to ~561 w), paper/sections/intro.tex (roadmap rewired), paper/sections/position.tex (cross-refs), paper/sections/pattern.tex (§domain-ontologies + §research-infrastructure split out; "Cousins." paragraph folded in), paper/sections/eight-practices.tex (orphan cross-refs cleaned), paper/sections/provenance-analysis.tex (trimmed 436 → 276 w), paper/sections/evolution.tex (three insight paragraphs collapsed into one), paper/sections/limits-and-objections.tex (new; merged sustainability + objections), paper/sections/appendix-g-pattern-extensions.tex (new; holds the moved subsections), paper/sections/background.tex (removed), paper/sections/related.tex (removed), paper/sections/sustainability.tex (removed), paper/sections/objections.tex (removed), doc/provenance.ttl, doc/logbook.md, doc/user-contributions.md.

Decision / outcome. Researcher direction: "generate the focused version. afterwards complete update pass." against the seven-item triage filtered by narrative load. Five-beat spine (gap → position → eight integrated practices + pipeline → recursive case study + honest-about-limits → why now) was the filter; sections that did not earn a beat were trimmed, folded, or appendixed.

Cuts:

• background.tex cut entirely. Substrate already cited inline by intro.tex and named again in pattern.tex where load- bearing.
• related.tex cut entirely. Datasheets / Model Cards / FAIR4 already cited; the load-bearing formal-methods-cousin paragraph folded into pattern.tex as Cousins.*.
• sustainability.tex + objections.tex merged into a single limits-and-objections.tex (~683 w from a combined ~1170 w); resource cost + frontier-model dependence + equity + the five standing objections + long-term archivability all under one heading.
• pattern.tex §domain-ontologies + §research-infrastructure (~570 w) moved to appendix-g-pattern-extensions.tex; body retains a one-paragraph forward pointer.
• authors-note.tex rewritten ~1010 → 561 w; partition + coupling rule + tokens-out-tokens-in kept; chronicle of cooperation events dropped (those live in evolution.tex and user-observations-log.md).
• evolution.tex three "load-bearing insight" paragraphs (Researcher-as-corrector / Pipeline as target / Logbook as conscience) collapsed into one combined paragraph; the chronology was already in appendix-d.
• provenance-analysis.tex trimmed to two figures + a diagnostic-vs-not framing paragraph; the four SPARQL queries remain in appendix-e.

Numbers: body 10466 → 6496 words (~38% trim); PDF 38 → 33 pages. At a 530 wpp position-paper density, body now lands near 12 pp prose + 4 pp floats; with appendices and bibliography, total trends toward the 10-pp body target without yet hitting it.

Provenance: act:trim-pass-focused-narrative (prov:wasInformedBy act:trim-pass-2-appendix-routing); hc:trim-pass-focused-narrative (structural-decision, high). Four prov:Invalidation records (background, related, sustainability, objections in their body roles); two new fair2r:Section IRIs (limits-and-objections, appendix-pattern-extensions), each prov:wasDerivedFrom its predecessor. Triple count 2189 → 2227 (+38).

Cross-references repaired: every \S\ref{sec:background}, \S\ref{sec:related}, \S\ref{sec:sustainability}, \S\ref{sec:objections} redirected (most → sec:pattern or sec:limits-and-objections; a few orphan parentheticals deleted). The intro roadmap is rewritten to walk the five-beat spine; the Author's Note's closing roadmap shrinks accordingly.

Next: Hit the 10-pp budget with one more pass over prose density (the eight-practices description blocks, the conclusion's Xennial paragraph, and the intro's volume-problem paragraph are the remaining-prolix candidates). The bibliography Missing $ warning that keeps pdflatex returning non-zero is still a real defect to chase.

2026-05-07 — Scrutinizer-pass fix-up after PR #43

Author: claude-opus-4-7 (under direction of repo owner) Touched: paper/sections/conclusion.tex, paper/sections/authors-note.tex, paper/sections/intro.tex, paper/sections/pattern.tex, paper/sections/eight-practices.tex, paper/sections/limits-and-objections.tex, paper/sections/self-reference.tex (deleted), paper/figures/eight-practices.tex (deleted), doc/provenance.ttl, doc/logbook.md, doc/user-contributions.md.

Decision / outcome. Researcher direction "run a scrutinizer pass". Dispatched three audit subagents (layout-scrutinizer, readability-reviewer, fair-aligner) against the post-trim state. Small / safe findings applied directly; taste calls deferred.

Layout fixes: - paper/sections/self-reference.tex and paper/figures/eight-practices.tex (orphan files retired in earlier passes but kept on disk) deleted. Their prov:Invalidation records remain in the graph; the on-disk artefacts are gone. - Three figure cross-refs added so the captions are no longer prose-orphan: Figure~\ref{fig:hero} cited in intro.tex (the propose-F(AI)²R sentence); Figure~\ref{fig:contribution-histogram} cited in authors-note.tex (after "high in leverage"); Figure~\ref{fig:failure-mode-coverage} cited in eight-practices.tex opening paragraph.

Readability fixes: - conclusion.tex \S\ref{sec:research-infrastructure} → Appendix~\ref{sec:appendix-pattern-extensions} (the subsection moved to the appendix in PR #43; the prose still pointed at the old body location). - The 274-word Xennials paragraph in conclusion.tex split at "F(AI)²R is in part an attempt to encode it." — a natural seam between the "what was lost" framing and the "responsibility-uptake" framing. - Duplicate "None of the eight is novel" line dropped from conclusion.tex summary (still appears in intro.tex + position.tex + eight-practices.tex — three sites is a position-paper convention, four was incantation). - Redundant closing roadmap dropped from authors-note.tex (the intro carries the same roadmap, more completely). - The trim-artefact paragraph in pattern.tex folded into the new Extensions and cousins. paragraph (one heading replaces two). - Long-term archivability. paragraph dropped from limits-and-objections.tex — it retreaded the frontier-model dependence material in the same section.

FAIR-aligner fix: - 8 newly orphaned bib entries re-cited in their natural prose locations (the trim removed the dedicated sections that carried these \cite calls but the underlying claims still depend on them): * page2021prisma, guyatt2008grade → pattern.tex Crosswalk paragraph (PRISMA + GRADE crosswalk for the verification ladder). * w3c2013provo → pattern.tex §axes I (PROV-O over Turtle). * gebru2021datasheets, mitchell2019modelcards → eight-practices.tex item 3 (per-claim provenance maps as structured-documentation descendants of Datasheets / Model Cards). * aclrr_llm_policy, iclr_llm_policy, neurips_llm_policy → intro.tex venue-policy-churn sentence.

Acknowledged but deferred: - The provenance_analysis.py slug-match heuristic flags sec:appendix-evolution as not in graph even though the graph IRI is ent:section-appendix-evolution-chronology. Cosmetic; script-level fix for a follow-up. - The bibliography Missing $ warning that keeps pdflatex returning non-zero (build-paper.yml currently treats a non-empty PDF as success); separate follow-up.

Local compile: latexmk -pdf produces main.pdf at 34 pages, no undefined references. Triple count 2227 → 2243 (+16).

Provenance: act:scrutinizer-pass-fix-up (prov:wasInformedBy act:trim-pass-focused-narrative); hc:scrutinizer-pass-fix-up (corrective-intervention, medium).

2026-05-07 — Hero figure rendered to PNG; embedded on the Pages Home page

Author: claude-opus-4-7 (under direction of repo owner) Touched: paper/figures/hero.standalone.tex (added amsmath), paper/figures/hero.png (new — 200 dpi PNG export), site/static/figures/hero.png (copy for the Pages site), site/index.md (embedded), doc/provenance.ttl, doc/logbook.md, doc/user-contributions.md.

Decision / outcome. Researcher question: "do we have an eyecatcher graphic is it also in the pages?". The PDF carries hero on its front matter (loaded by paper/main.tex via \input{figures/hero}); the Pages site did not surface any of the paper's signature visuals. Closes the gap by rendering the TikZ source via the existing standalone wrapper to a 200 dpi PNG and embedding it on site/index.md between the lede and the CTA row. Same drift-proof pattern as the methodology-page evolutions table: source of truth in paper/figures/, render committed to site/static/figures/, embedding declared in the page Markdown.

paper/figures/hero.standalone.tex gained a \usepackage{amsmath} declaration; without it the standalone build broke on the math-mode title $\text{F}\,(\text{AI})^{2}\,\text{R}$. This mirrors the same fix applied to paper/style/fair2r.sty in the PDF-build fix pass earlier today.

Provenance: act:render-hero-png-for-site; hc:eyecatcher-on-pages (corrective-intervention, low). Graph triples 2243 → 2257 (+14).

2026-05-07 — Conference poster scaffolded as the fifth primary artefact

Author: claude-opus-4-7 (under direction of repo owner) Touched: slides/poster-A0.tex (new), slides/Makefile (adds poster target), slides/poster-A0.png (new render), site/static/figures/poster-A0.png (copy for Pages), site/index.md (CTA link), paper/figures/hero.png (rendered earlier today), .github/workflows/build-slides.yml (poster compile step + log dump + artefact upload + release publish), CLAUDE.md (primary-artefact list 4 → 5; slides/poster-A0.tex is now binding alongside the manuscript / graph / logbook / slide decks), doc/provenance.ttl (fair2r:Poster schema growth + ent:poster-A0 entity + act:scaffold-conference-poster + hc:scaffold-conference-poster), doc/logbook.md, doc/user-contributions.md.

Decision / outcome. Researcher direction "add another artifact to the build chain a conference poster/Infographic. consider distilling the process to its essentials. how would that change page count for paper", followed by "go" against a two-step plan.

Step (a) — landed in this commit. The poster is a tikzposter A0 landscape, three-column. Custom theme attempts produced a debug stamp on page 1 (a known tikzposter behaviour for non-built-in themes), so the layout uses the built-in Simple theme with DLR-CD colour overrides (titlefgcolor, blocktitlefgcolor, blocktitlebgcolor etc. routed to dlrBlau1 / dlrBlau5). Helvetica/Arial fallback; square corners; hairline rules. Local pdflatex compile: 1 page, ~430 KB; PNG render at 100 dpi for the Pages site is ~972 KB.

Layout:

• Top banner — title + author + ORCID + DLR / Helmholtz / NFDI4Ing / HMC + repo URL.
• Left column — The gap, Reported base rates (Walters 55 % / 18 %, Magesh 17 – 34 %), Why now.
• Centre column — The position (with the hero figure as eye-catcher), Eight integrated practices.
• Right column — Verification ladder, live populations (with ladder-populations.png), Provenance graph as audit substrate (with provenance-topology.png), This is the worked example (URLs + tagline).
• Footer — licence + DRAFT watermark.

Step (b) (run the condenser against the paper with "the poster carries this; the paper need not" as the new criterion; plausible body trim ~3 pp) is a follow-up.

Build-chain extension:

• slides/Makefile gains a poster target; all rule now builds pitch + conference + poster.
• .github/workflows/build-slides.yml extended with a poster compile step (with continue_on_error like the decks), a log dump on failure, and the artefact-upload + release-publish steps now include poster-A0.pdf. The release tag's name updated to "latest draft slide decks + poster".
• CLAUDE.md primary-artefact list: 4 → 5. The primary-artefact consistency invariant now binds the poster alongside the manuscript, graph, logbook, and slide decks.
• Schema growth: fair2r:Poster a owl:Class ; rdfs:subClassOf prov:Entity ; in the schema block of doc/provenance.ttl.
• site/index.md Home CTA row gains a Conference poster (A0 PDF) button pointing at the rolling release.

Provenance: act:scaffold-conference-poster; hc:scaffold-conference-poster (structural-decision, high leverage). ent:poster-A0 records prov:used for hero, ladder-populations, and provenance-topology (the three figures the poster embeds). Triple count 2257 → 2287 (+30).

2026-05-07 — Step (b) of poster + condenser plan: lean cuts

Author: claude-opus-4-7 (under direction of repo owner) Touched: paper/main.tex (front-matter hero figure removed), paper/sections/intro.tex (orphaned \ref{fig:hero} cleaned), paper/sections/provenance-analysis.tex (ladder-populations figure removed; pointer to the poster added), doc/logbook.md.

Decision / outcome. User said "go" against step (b) of the poster plan. Two cuts landed cleanly with the conference poster as the new visual heavy-lift:

• Front-matter hero figure dropped from paper/main.tex. Paper opens directly with the abstract; the poster (slides/poster-A0.tex) is the eye-catcher. The TikZ source paper/figures/hero.tex is preserved on disk and rendered to paper/figures/hero.png for the Pages site.
• ladder-populations.pdf figure dropped from provenance-analysis.tex. The schema-diagram figure (provenance-topology.pdf) stays in the body; the populated ladder is on the poster, with a one-sentence forward pointer in the body prose.

PDF page count: 34 → 33. The trim was front-loaded in the body where the front-matter figure* and a body-page figure both exited; the appendices are now substantial enough that further single-figure cuts produce diminishing page returns.

The remaining candidates from the original step-(b) sketch (AI² grid resize, failure-mode-coverage to appendix, intro contribution-paragraph trim, position 5-claims tighten) are deferred — the user pivoted to a mobile-friendly Pages site; the prose-density work is the next iteration of trim, not this one.

2026-05-07 — Pages site: mobile-friendly CSS

Author: claude-opus-4-7 (under direction of repo owner) Touched: site/static/style.css, doc/provenance.ttl, doc/logbook.md, doc/user-contributions.md.

Decision / outcome. User direction "make the pages mobile friendly". Changes are CSS-only (no template change, no JS) so the site stays a static-build artefact.

Added to site/static/style.css:

• Responsive defaults (always on): img, video, iframe inside .content get max-width: 100%; height: auto; pre gets overflow-x: auto. Stops figures and code blocks from blowing past the viewport on tablets and landscape phones.
• Phone breakpoint at 720 px: tighter horizontal padding (16 px instead of 32 px); header stacks logo above org-line; utility ribbon wraps; nav wraps to a second row rather than overflowing horizontally; hero H1 26 px (was 36 px); body H1/H2/H3 stepped down; CTAs stack full-width; tables become display: block; overflow-x: auto so the evolutions table scrolls horizontally rather than overflowing the viewport; Mermaid blocks scroll on overflow.
• Narrow-phone breakpoint at 380 px: further H1 and nav-item shrink for iPhone-SE-class devices.

The existing breakpoints (900 px collapses the sidebar / content into a single column; 720 px collapses the footer + card-grid) were preserved alongside the new ones.

The viewport meta tag (<meta name="viewport" content="width=device-width,initial-scale=1">) was already in the page template; no change needed.

Provenance: act:make-pages-mobile-friendly; hc:make-pages-mobile-friendly (corrective-intervention, medium). Triple count 2287 → 2301 (+14).

2026-05-07 — README updated for consistency; poster CI compile fixed

Author: claude-opus-4-7 (under direction of repo owner) Touched: README.md, .gitignore, slides/poster-A0.tex, .github/workflows/build-slides.yml, doc/logbook.md, doc/user-contributions.md, doc/provenance.ttl.

Decision / outcome. User flagged "poster link not working" and then "update readme as well and keep consistent". Two fixes shipped in tandem.

Poster CI compile. The latest-draft-slides release was publishing the decks but no poster-A0.pdf after the workflow fix in PR #47 (publish step now runs if: always()). Diagnosis: the poster's \includegraphics{../paper/figures/...} calls likely fail to resolve in the xu-cheng/latex-action@v3 container's path-traversal policy, even though local pdflatex succeeds. Two-line fix:

• slides/poster-A0.tex switched from \includegraphics{../paper/figures/hero.png} to plain \includegraphics{hero.png}, with \graphicspath{{./}{../paper/figures/}{paper/figures/}} as a fallback chain so the same source compiles whether the PNGs sit in slides/ (CI) or paper/figures/ (local).
• .github/workflows/build-slides.yml gains a Stage figures for the poster step at the top of the job that copies paper/figures/{hero,ladder-populations,provenance-topology}.png into slides/ before any latex-action step runs. Logs the staging output so a future failure is easy to diagnose.

.gitignore extended to ignore the staged PNG copies under slides/ so local builds don't accidentally commit duplicates of the figures.

README consistency pass. The README.md had drifted behind the recent structural changes (poster as 5th primary artefact, focused-narrative trim, Get-started page, mobile-friendly site, transcript discipline, schema growth):

• Read on the web mentions the mobile-friendly site and points at the Get started page.
• Download the artefacts table gains a row for poster-A0.pdf; surrounding prose names the five primary artefacts and points at CLAUDE.md for the consistency invariant.
• Layout tree updated: slides/ now lists pitch + conference + poster; doc/ lists transcripts/, user-contributions.md, user-observations-log.md, sources.md; scripts/ lists the three new ones; root lists CLAUDE.md.
• Build updated: make -C slides all now builds pitch + conference + poster.
• Provenance paragraph names the additional schema classes (fair2r:Slidedeck, fair2r:Poster, fair2r:Transcript, fair2r:Contribution), the triple count (~2300+), and the no-parentless-claim invariant.
• Logbook renamed to Logbook + transcripts and lists the per-session transcript convention.
• New Adopting F(AI)²R for your own paper section pointing at the auto-synced Get-started page.
• Design section drops the (forthcoming) the slide deck caveat (decks + poster are real now).

Provenance: act:readme-consistency-pass (no separate IRI minted; this is a corrective edit), hc:readme-consistency-pass appended.

2026-05-08 — Layout-scrutinizer scope expanded; poster flipped to A0 portrait

Author: claude-opus-4-7 (under direction of repo owner) Touched: agents/layout-scrutinizer.md (rewritten with per-artefact scrutiny rules), slides/poster-A0.tex (landscape → portrait), site/static/figures/poster-A0.png (re-rendered at 70 dpi from the portrait PDF), CLAUDE.md (primary-artefact list 5 — poster orientation note), README.md (table + tree mentions), doc/provenance.ttl (rdfs:label / rdfs:comment updates), doc/logbook.md, doc/user-contributions.md.

Decision / outcome. Two researcher directions batched.

(1) "layout scrutinizer should scrutinize slides and poster as well". agents/layout-scrutinizer.md rewritten. The binding-scrutiny scope expands from manuscript-only to all five primary artefacts plus the Pages site. Per-artefact rules now explicit: paper (overfull / underfull, page budget, label hygiene); slide decks (frame overflow, low-contrast pairings, divider-vs-section alignment); conference poster (1-page enforcement, column balance, 1-m-distance legibility, content discipline); Pages site (LaTeX leakage, table rendering, image alt-text, responsive breakpoints, cache-bust consistency). Output format: per-artefact verdict line plus a markdown punch list with file:line (or page reference for PDFs / HTML pages).

(2) "poster should be portrait format not landscape". slides/poster-A0.tex \documentclass option flipped from landscape to portrait. Local pdflatex still produces a clean 1-page A0 PDF (841 × 1189 mm portrait). PNG re-rendered at 70 dpi from the portrait PDF and committed to site/static/figures/poster-A0.png. Landscape-to-portrait references propagated to CLAUDE.md, README.md, and the two rdfs: labels / comments in doc/provenance.ttl that previously said landscape.

Provenance: act:expand-layout-scrutinizer-and-flip-poster; hc:expand-layout-scrutinizer-and-flip-poster (rule-shape, medium leverage). Triple count 2301 → 2330 (+29).

2026-05-08 — Provenance graph hygiene pass

Author: claude-opus-4-7 (under direction of repo owner) Touched: doc/provenance.ttl, doc/logbook.md, doc/user-contributions.md.

Decision / outcome. Researcher direction: "validate provenance i think there are some unconnected entities". Ran a 9-query audit. Findings + fixes:

Fixes:

• verif:lit-retrieved + verif:lit-read declared with a fair2r:VerificationState types in the schema block. The other five rungs were already declared; these two had been overlooked.
• ent:section-implementation typed as fair2r:Section (retired body role) so the prov:wasDerivedFrom edge from ent:section-appendix-implementation points at a typed entity rather than a dangling IRI.
• act:invalidate-discussion-section was orphan: no edge in or out. Now connected with prov:wasInformedBy act:rev-reconcile-and-illustrate plus an explicit prov:invalidated ent:section-discussion. The Discussion section itself (removed before being curated) is now declared as a tombstone with prov:wasInvalidatedBy.
• Four unused agent IRIs (orchestrator, source-analyzer, condenser, presentation) connected via prov:wasAssociatedWith to activities that used them in practice (act:bootstrap, act:rev-verification-pass, act:trim-pass-focused-narrative, act:add-illustrations-pass-4).
• ent:transcript-2026-05-07 bound to its parent activity act:meta-cooperation-2026-05-07-transcript-preservation via prov:used.

Documented but unfixed: 28 ghost sources (the verification scoping doc names this number; ghost-citation tolerance is intentional — not every cited substrate needs to be the object of a prov:wasDerivedFrom); 13 unused observations (hypothesis-stage by design — they gain prov:influenced when they graduate from doc/user-observations-log.md to paper text).

Provenance: act:provenance-graph-hygiene-pass; hc:provenance-graph-hygiene-pass (corrective-intervention, medium leverage).

2026-05-08 — Portrait poster density fix + scrutinizer rule update

Author: claude-opus-4-7 (under direction of repo owner) Touched: slides/poster-A0.tex (font 25pt → 30pt; figures sized to full \linewidth; added Contribution shape and Failure modes the practices cover blocks to the centre column), site/static/figures/poster-A0.png (re-rendered, now 918 KB), .github/workflows/build-slides.yml (staging step copies contribution-histogram.png as well), agents/layout-scrutinizer.md (two new poster checks: vertical fill ≥ 80 %, column balance ≤ 15 % delta), doc/provenance.ttl, doc/logbook.md, doc/user-contributions.md.

Decision / outcome. User screenshot exposed that the portrait poster fill was only ~50 % of the page — the content sized for landscape didn't scale up when the orientation flipped. The scrutinizer rules I had just landed in PR #52 did not catch this density defect.

(1) Density fix: documentclass bumped 25pt → 30pt; figures widened from 0.95–0.96 linewidth to full \linewidth; centre column gained a Contribution shape block (with the existing contribution-histogram.png) and a Failure modes the practices cover block. Rendered fill measured at 98 % of page height (verified with pdftoppm + Pillow.find-lowest-non-white-row).

(2) Scrutinizer rule update: agents/layout-scrutinizer.md gained two new poster checks under the You do (conference poster) section: - Vertical fill must be ≥ 80 % — render the PDF to a PNG, find the lowest non-white scan row, divide by total height; that's the fill ratio. - Column balance ≤ 15 % — run pdftotext -layout and check the per-column trailing whitespace; tallest column shouldn't outrun the shortest by more than 15 %.

Both are now catchable invariants rather than eyeballing defects, with the same render-and-measure script I used to verify this commit.

Build-slides workflow's staging step extended to copy contribution-histogram.png (the new poster asset) into slides/ before the compile.

Provenance: act:fix-poster-density-and-scrutinizer-rule (prov:wasInformedBy act:expand-layout-scrutinizer-and-flip-poster); hc:fix-poster-density-and-scrutinizer-rule (rule-shape, high leverage — catches a recurring defect class). Triple count 2363 → 2379 (+16).

2026-05-13 — Integrate three external works; vendor paper PDF; slide scrutinizer fixes; venue research.

Researcher direction listed three external works to position against (Liu et al. 2026 "Ara", Vaswani et al. 2017 "Attention Is All You Need", MLCommons Croissant 1.0). Each landed on a distinct part of the position argument:

• Ara (arXiv 2604.24658): concurrent maximalist proposal — four-layer agent-native package replacing the narrative paper. Added a paragraph to paper/sections/limits-and-objections.tex under a new Adjacent proposals and adjacent vocabularies block. F(AI)²R declines the bargain: we keep the manuscript as primary artefact and ride a PROV-O graph alongside; Ara's exploration layer is a credible upper bound for what the same substrate could carry.
• Vaswani 2017: historical pivot. Inserted at the end of paper/sections/intro.tex between the journal-as-distribution paragraph and the F(AI)²R proposal — FAIR's 2016 vintage predates the Transformer pivot by a year, which makes the operational-vocabulary gap load-bearing.
• Croissant 1.0: dataset-metadata vocabulary that F(AI)²R is sometimes mistaken for. Added one paragraph to limits-and-objections.tex and a cross-reference from paper/sections/appendix-g-pattern-extensions.tex. The two compose (a fair2r:Source can reference a Croissant record); neither subsumes the other.

Three new fair2r:Source entities, three new fair2r:Claim entities (with prov:wasGeneratedBy back to a new parent activity act:integrate-external-2026-05-13), and one hc:integrate-external-2026-05-13 fair2r:HumanContribution entity recording the researcher direction. Three new bib entries in paper/references.bib (liu2026ara, vaswani2017attention, benjelloun2024croissant).

Slide scrutinizer fixes applied in the same pass:

• slides/pitch-5min.tex: hero figure reduced from 0.85 to 0.70 \textwidth; (AI)² grid split off into its own frame at 0.92 \linewidth so the Position frame doesn't try to carry both block and grid.
• slides/conference-30min.tex: ladder-fsm gained an explicit height=0.62\textheight, keepaspectratio constraint to stop the 368 pt overfull \vbox at page 13; ladder-populations narrowed from 0.96 to 0.86 \textwidth to stop the 21 pt overfull at page 37; axes-table columns widened from p{2.6cm} to p{3.0cm} for uniform breathing room. Section-divider titles for §5 / §6 reworded to acknowledge the deck's actual structure (appendix-G material as forward arc; recursive case study covers provenance analysis + process evolution in compressed form); agenda items synchronised.
• slides/poster-A0.tex: Reported-base-rates table column spec rl → rp{0.55\linewidth} to stop the 178.9 pt overfull \hbox.

Vendored paper/main.pdf into the repository. Both .gitignore files no longer ignore the file; make -C paper pdf rebuilds it. Adopters and reviewers can now read the draft (with the watermark) without invoking CI. The draft notice is on; publication is still gated on paper/publication-consent.md.

Venue research (in response to the user's RDA P27 question + ranked-by-acceptance-and-impact follow-up): shortlist refreshed in doc/submission-plan.md. Top three recommendations: (1) arXiv as the citable anchor; (2) RIO Journal as the formal record (methods paper, no novelty gate); (3) RDA P27 BoF + poster for community traction. For P27 specifically: the recommended format is a BoF session (90 min) + poster (late-breaking, ~Jul 2026 deadline) + A4 one-page handout, with WG-chair outreach this week since the formal sessions CfP closed in November.

Provenance: act:integrate-external-2026-05-13 (prov:wasAssociatedWith agent:human-author + agent:claude-opus-4-7); hc:integrate-external-2026-05-13 (content-prompt, high leverage — each work closes a specific class of standing objection the paper would otherwise answer ad hoc in review). Triple count 2379 → 2395 (+16).

2026-05-13 — Condenser pass: body trimmed from 15 to 12 pages

Ran the condenser agent against the body (§1–§9). Target was 10–12 pp; landed at 12 body pages (pp 3–14, with §9 conclusion bleeding briefly into p 15). Total PDF: 32 pp (was 33).

Where the lines came from. §6 Evolution lost the most (opening paragraphs collapsed to a single intro; the Three threads paragraph compressed; five pipeline-plumbing rows dropped from the digest table — Marp retirement, DLR-CD pinning, illustration-toolset opening, Pages-source switch, LaTeX-fragment stripper — all still recorded in Appendix D chronology and the full graph). §3 Pattern lost the long FSM-rung and crosswalk expositions, the Extensions and cousins paragraph collapsed, and the handback subsection tightened. §8 Limits compressed the resource-cost, equity, Ara, Croissant, and standing-objections paragraphs (Ara and Croissant integrations preserved per the external-positioning constraint; all five standing objections preserved as a single block). §9 Conclusion shortened the xennials/bridge paragraph by half. §1 Intro shortened the FAIR4AI-rejection paragraph. §4 Eight Integrated Practices trimmed every item but kept all eight.

Constraints honoured. All eight integrated practices survive in §4 and §8. The recursive case-study commitment (§6) survives. The Ara / Vaswani / Croissant external-work sentences all survive. Every \todo[inline]{verify} marker survives. No load-bearing citation was dropped; no fair2r:Claim triple required retirement (the cuts were prose compressions and digest-table plumbing rows, not deletions of claim-corresponding sentences). The condenser agent's recursive-case-study commitment to its own discipline holds: this entry documents the cut, and the commit lands as a single concern.

2026-05-17 — Author's Note: "What I still believe the work is for".

Researcher stated a personal conviction for the front-matter Author's Note: future science remains driven by creativity, tenacity and curiosity; an LLM makes the overview cheap and the state of the art easy to assemble, but true novelty — the unasked question, the not-yet-thinkable abstraction — is a human quality the models do not match yet. Drafted into paper/sections/authors-note.tex as a new paragraph placed after the coupling-rule paragraph and before Tokens out, tokens in.

The hedge yet is kept and explicitly framed as a dated present-tense claim (consistent with the version-manifest discipline), not a metaphysical guarantee. The paragraph ties to the existing "would the abstraction have been thinkable without the model?" diagnostic: the conviction is the normative reason that diagnostic matters.

Provenance: claim:human-novelty-not-yet-matched (verif:human-confirmed, prov:wasGeneratedBy act:authors-note-novelty-conviction, attributed to agent:human-author); hc:authors-note-novelty-conviction (content-prompt, high leverage). New agent:claude-sonnet-4-6 fair2r:AIAgent declared (this session's model; was previously undeclared). No slide change: this is a first-person Author's Note coda, not one of the five load-bearing position claims.

2026-05-17 — RDA P27 Ideas Pitch draft; add OECD AI-in-Science source.

Two researcher directions in one session.

RDA P27 Ideas Pitch. Generated doc/rda-p27-ideas-pitch.md, a fill-in-ready draft matching the live submission form's field order (contact applicant; title ≤10 words; ~350-word overview; USP; from-idea-to-RDA-group; links; thematic area; presenter; acknowledgements). Thematic area set to Engineering and technology per the author's choice (aligns with the DLR/NFDI4Ing home community and the data-infrastructure thrust). The doc explicitly records that an Ideas Pitch is a community-presentation submission and does not flip paper/publication-consent.md.

OECD source. Added oecd2023aiscience (OECD, Artificial Intelligence in Science: Challenges, Opportunities and the Future of Research, OECD Publishing, 26 June 2023, doi:10.1787/a8d820bd-en). New paper/references.bib entry; new src:oecd2023aiscience triple at verif:ai-confirmed; a prov:wasDerivedFrom edge added to claim:equity-not-neutral (the report is direct policy-body evidence for the equity claim — it states less well-funded groups risk falling behind and argues for greatly increased AI-in-science funding for developing regions); \cite{oecd2023aiscience} added to the Equity paragraph of paper/sections/limits-and-objections.tex alongside bender2021stochastic, birhane2022values; new row in doc/sources.md §13. The shallow-work / narrowing-novelty material in the report also corroborates the Author's-Note novelty conviction, noted in the source comment but not given a separate prov:wasDerivedFrom edge (the formal edge is on the equity claim only).

Provenance: act:add-oecd-aiscience-source (prov:used src:oecd2023aiscience); hc:add-oecd-aiscience-source (content-prompt, medium leverage). DOI follows the OECD iLibrary 10.1787/<id> convention; rung ai-confirmed pending human read.

2026-05-18 — Seventh source-analyzer verification pass

Ran a live-web sources validation loop (Exa web_search / web_fetch MCP tools + academic-search MCP) over the known-defect list and the recently-added, never-human-checked sources. Agent: claude-opus-4-7[1m] under the source-analyzer prompt.

Material correction (graph). src:vannoorden2023chatgpt: paper/references.bib had already been corrected in an earlier pass to DOI 10.1038/d41586-023-03930-6 (Van Noorden + Webb, Nature 624(7992):509), but the doc/provenance.ttl mirror still carried the wrong identifier doi:10.1038/d41586-023-03907-5 (an unrelated India-economics article) and was stuck at verif:lit-retrieved. Confirmed the correct record via the Nature landing page and PubMed (PMID 38093061); corrected the graph triple (DOI, both dcterms:creators, full dcterms:title) and advanced verif:lit-retrieved -> verif:ai-confirmed. This resolves the long-standing correction-list item 1 and clears the source's institutional-access escalation.

Graph title alignment. src:benjelloun2024croissant title aligned from the NeurIPS-paper title to the cited @techreport title "Croissant Format Specification, Version 1.0" (MLCommons, 2024-03-01, first author Omar Benjelloun).

Safe bib note updates. iclr_llm_policy url updated from the bare iclr.cc homepage to the confirmed stable policy page https://iclr.cc/FAQ/LLM; rung kept at lit-retrieved (per-cycle text capture still escalated — conservative). icmje2023 note added recording the now-"Updated January 2026" revision and the AI-specific topic-page URL; bibkey/year 2023 kept stable.

Re-confirmed correct, no change: ravi2024fair4ml (Zenodo 14002310, Castro/Garijo), oecd2023aiscience (10.1787/a8d820bd-en; equity / shallow-work passages verbatim), liu2026ara (arXiv:2604.24658, Jiachen Liu), vaswani2017attention (arXiv:1706.03762, Ashish Vaswani, kept verif:human-read), icmje2023 (page resolves); self-cites krebs2026faiarmain / krebs2026faiarcompanion internally coherent.

No source downgraded to verif:needs-research; no source unresolvable; no identifier fabricated. Provenance: act:meta-cooperation-2026-05-18-verification-pass-7 (prov:used the seven re-checked sources); hc:verification-pass-7-prompt-2026-05-18 (responsibility-uptake, medium); ai:verification-pass-7-2026-05-18 (corrective-intervention, medium). Reading queue regenerated; doc/provenance.ttl re-parses via rdflib (2517 triples); make -C paper pdf clean-builds with no undefined citations.

2026-05-18 — Slide-deck reconciliation with the post-split manuscript

The three slide artefacts had drifted after a session that split the appendices into a standalone companion technical report, integrated three external works, and added a new Author's Note conviction. Brought all three back into consistency (presentation agent; deck-downstream-of-manuscript rule observed — every slide claim traces to current paper text).

Companion split. slides/pitch-5min.tex "Worked example" artefact list and slides/poster-A0.tex worked-example block now name the companion technical report and the A0 poster; slides/conference-30min.tex "Recursive attestation" folds in the position-paper (~22 pp) + companion (~14 pp) split (page counts confirmed via pdfinfo paper/main.pdf = 22, paper/companion.pdf = 14).

Stale numbers corrected in "Recursive attestation", each verified against the live repo: 1300+ triples -> 2500+ triples (rdflib len = 2517 at edit time; qualitative >2500 chosen deliberately so the figure stays true as the graph grows — this very bookkeeping took it to 2569); 50 bib entries -> 59 (grep -c '^@' paper/references.bib); reading queue 45+ entries -> 55 sources (doc/reading-queue.md state header); 25-page Pages site -> multi-page Pages site (27 *.html in _site/ at edit time but build-variable, so qualitative per the no-fabrication rule rather than a new precise-but-fragile number).

Three external works mirrored from the paper wording (no new claim): Vaswani et al. 2017 as the historical pivot — one sentence in conference §1 Motivation, one clause in the pitch problem frame (from paper/sections/intro.tex); a new conference §8 frame "Adjacent proposals and adjacent vocabularies" carrying Ara (Liu et al. 2026 — "F(AI)²R declines that bargain"; Ara's exploration layer a credible upper bound) and MLCommons Croissant 1.0 (Benjelloun et al. 2024 — composes with, neither subsumes), both from paper/sections/limits-and-objections.tex; OECD 2023 added as the policy-body anchor on the conference equity bullet.

Closing conviction. The Author's Note paragraph "What I still believe the work is for" surfaced as the closing message on both Thank-you frames and the poster worked-example block, with the yet hedge and the dated-not-metaphysical framing preserved (mirrored from paper/sections/authors-note.tex).

Bib keys cited on the slides (vaswani2017attention, liu2026ara, benjelloun2024croissant, oecd2023aiscience) confirmed present in paper/references.bib. Build: make -C slides all clean — pitch 10 pp / 0 overfull boxes; conference 39 pp (baseline 38, +1 new frame) with no new overfull boxes vs the 3 pre-existing baseline boxes; poster 1 page, vertical fill 97.8% (PNG scan-row method per agents/layout-scrutinizer.md), columns balanced within ~1%. No undefined references or citations in any deck. Provenance: act:reconcile-slides-2026-05-18 (prov:wasRevisionOf edges on the three slide entities — ent:slidedeck-pitch-r2, ent:slidedeck-conference-r2, ent:poster-A0-r2); hc:reconcile-slides-2026-05-18 (content-prompt, medium); ai:reconcile-slides-2026-05-18 (content-prompt, medium, agent:claude-sonnet-4-6). doc/provenance.ttl re-parses via rdflib (2569 triples); reading queue unchanged (no source/claim added).

2026-05-18 — Adjacent-proposals positioning diagram (conference deck)

Researcher standing direction: use a visual on the slides where it helps the message, no photorealism. No image-generation connector is available in this environment (verified exhaustively across the session, including after a restart); the prudent substitute under the DLR Corporate Design and no-fabrication rules is a deterministic TikZ diagram in the repository's existing figure convention.

Assessment of the 2026-05-18 reconciliation changes: exactly one frame is text-dense conceptual positioning that a schematic clarifies in a talk — the new conference frame Adjacent proposals and adjacent vocabularies (two long bullets: Ara replaces the paper / F(AI)^2R keeps-and-rides / Croissant composes / neither subsumes). The other landed changes (a Vaswani clause, an artefact-list line, corrected numbers, a closing aphorism) are inherently textual; adding figures there would be decorative bloat against the institutional-voice and no-bloat rules. No figures added elsewhere.

paper/figures/adjacent-proposals.tex created (tikzpicture-only, DLR palette, square corners, hairlines, no gradients; same \providecommand text-macro pattern as figures/ai-squared-grid.tex — the first compile hit the TikZ-style-vs-text-macro pitfall that file documents and it was fixed the same way). The conference frame's two bullets were replaced by the figure plus a one-line caption; the claims are unchanged (still mirror paper/sections/limits-and-objections.tex). Frame rendered to PNG and inspected before landing. Conference deck recompiles clean (exit 0, 39 pp; largest overfull 39.6pt, within the pre-existing baseline class — no new overfull from the figure).

Provenance: act:adjacent-proposals-figure-2026-05-18; ent:figure-adjacent-proposals; ent:slidedeck-conference-r3 (prov:wasRevisionOf ent:slidedeck-conference-r2, prov:used ent:figure-adjacent-proposals); hc:/ai:adjacent-proposals-figure-2026-05-18. ttl re-parses (2607 triples).