User-observations log

Append-only log of observations the human author makes about the cooperative-writing process itself, separate from observations about F(AI)²R as a methodology. Both are valuable, but they have different fates:

• An observation about the methodology is a candidate fair2r:Claim and lives in paper/sections/*.tex if it survives the verification ladder.
• An observation about the cooperation is a candidate fair2r:MetaContribution and lives here. A subset graduates to paper content (typically in paper/sections/evolution.tex or paper/sections/authors-note.tex) with an explicit "graduated" marker; the rest stay here as context for future readers and successors.

Convention

## YYYY-MM-DD --- short title
*Source:* <session id or commit sha>
*Stage:* observation | reflection | hypothesis | finding
*Promoted-to-paper:* yes / pending / no — <reason>

The observation, in the user's own words where possible, with
short-quote attribution. AI commentary, where invited, in italics.

Newest entries at the bottom. Do not re-edit prior entries; if a prior observation needs revision, append a follow-up referencing the prior date.

2026-05-07 — Associative-thinking capacity as the cooperation multiplier

Source: session 2026-05-07 (researcher prompt at the end of the complete-loop audit pass). Stage: hypothesis Promoted-to-paper: pending — strong candidate paragraph for paper/sections/evolution.tex (process angle) or paper/sections/authors-note.tex (researcher-experience angle).

The researcher noted, in their own words: "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".

The observation, restated. The model's value in this cooperation has often not been in being right about a suggestion. It has been in producing an adjacent suggestion that the researcher then used as a springboard to a different idea the model had not proposed. The model's role in those moments is closer to a brainstorming partner or rubber duck than to an oracle: a generator of associatively-near-but-not-quite-right candidates whose job is to keep the researcher's own associative thinking primed.

Implication for the cooperation rule. The benefit a researcher extracts from an LLM cooperation may scale with the researcher's own associative-thinking capacity, not with model accuracy. A researcher who can hold five frames at once and pick the angle that fires gets more out of a noisy generator than a researcher who can only act on a confident answer. This reframes the failure mode from "the model was wrong" to "the human did not have a frame the model's near-miss was useful to".

Where this lands. This reframes the coupling-rule diagnostic (Author's Note: would the abstraction have been thinkable without the model?). The honest extension is: would the abstraction have been thinkable without the model and without the human's associative latitude to convert a near-miss into a usable angle? If yes, the cooperation was an accelerator on a frame the human already held; if no, the question genuinely opens whether the cooperation produced something neither party could have produced alone.

Risk. Naming this risks reducing the model's contribution to "useful noise". The fairer phrasing: the model's contribution is upstream of the suggestion that lands — it shapes the search space the human's associative leap then operates inside. Two associative generators working alternately may explore more space than one generator working alone, even when neither generator is more "creative" than the other in isolation.

For F(AI)²R as a published claim. The verification ladder grades whether a claim is sourced; it does not grade whether the claim's generation was solo or cooperative. The current schema captures this only via prov:wasAttributedTo (single agent) and prov:wasInformedBy (chain of activities). A future schema growth could mint fair2r:cooperativeOriginType with values {solo-author, accelerated-by-near-miss, co-generated, prompt-driven} so the cooperative-origin classification becomes queryable. Not for this paper — flag for the next case study.

2026-05-06 — Cooperation as conversation, decision-density as pace

Source: session 2026-05-06 (user message after the homepage CTA-row was wired) Stage: finding Promoted-to-paper: yes — paper/sections/authors-note.tex, What surprised me about the cooperation.

The user observed mid-session that "collaborative writing makes the user already think of next steps. Also real abstraction and logical reasoning seem to be the role of him. The LLM takes these ideas and turns them into something readable. Fascinating." 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 user followed up with "still I wonder how much of the work I'm actually doing. What's my contribution. I would love your perspective on that." The exchange became the seed for the What surprised me about the cooperation paragraph in the Author's Note, which now contains both the question and the answer.

2026-05-06 — Token budgets as schedule constraint

Source: session 2026-05-06 (user message after the position-paper Author's Note was drafted) Stage: observation Promoted-to-paper: pending — likely a footnote in paper/sections/sustainability.tex or a paragraph in paper/sections/evolution.tex.

The user noted: "when on a limited [plan] (Anthropic Max, the \$100 one) the question of focus becomes very important — token usage is very high and projects have to be prioritised — keeping track of usage windows. Schedule working sessions around them... weird stuff."

This is the operational analogue of the resource-cost paragraph in §. The sustainability section currently discusses inference cost from the perspective of the planet and the discipline; the user's observation is the same cost from the perspective of the individual researcher, who must batch sessions around a fixed monthly budget. Worth a one-paragraph addition that names the practical experience: an LLM-assisted writing pipeline imposes a schedule shape on the human, not just a compute cost on the provider.

2026-05-06 — Obscurity-Is-Dead as first try; F(AI)²R as second

Source: session 2026-05-06 (same message as above) Stage: finding Promoted-to-paper: implicit — paper/sections/evolution.tex and paper/sections/authors-note.tex already note the methodological ancestry.

"Obscurity paper was first try of collaborative writing. Now profiting from this experience (also a contribution: 'experience with the tool', meta-knowledge)." The implication is that the cooperation has its own learning curve, and that experience with the cooperation is itself a research contribution — one that the verification ladder does not currently capture, because it grades claims against sources, not authors against tools. A future revision of the methodology might add a fair2r:CooperationExperience property to the human-agent record.

2026-05-06 — Researcher role is becoming more technical / experimental

Source: session 2026-05-06 (same message) Stage: hypothesis Promoted-to-paper: pending — candidate paragraph for paper/sections/evolution.tex (third pattern observation).

"The researcher's role is becoming more technical expertise with these systems — more experimentation / conduction, which might be a good thing, as publication writing could be considered administrative overhead for experimentators. But what about disciplines that are more paper-driven — like philosophy?"

This is the strongest reframing-relevant observation in the log and deserves to graduate. The argument is that LLM cooperation shifts the researcher's centre of gravity from prose-production to experiment-design, which is a net good for empirical disciplines but possibly a loss for argument-driven disciplines (philosophy, parts of the humanities) where prose IS the experiment. The position paper should at least name this asymmetry rather than pretend F(AI)²R is discipline-neutral.

2026-05-06 — Prompts feel easier to write than the paper

Source: session 2026-05-06 (same message) Stage: finding Promoted-to-paper: pending — candidate paragraph for paper/sections/evolution.tex.

"Somehow writing prompts feels easier than writing the paper yourself. Why is that? Proof are these short cycles of conversation. As a computer scientist does collaborative writing feel more like 'writing a program' (at least interacting with new tech), which is closer to experiments than the formal writing process. AI as a motivation driver?"

Naming the psychological reframe: prompt-craft is closer to debugging a program than to drafting a paper. For a computer scientist, that maps to a more familiar and more rewarding cognitive loop. The paper should name this honestly — prompt-craft is not a substitute for argument-craft, but it lowers the activation energy of getting started, which is itself a contribution to throughput. The risk is that prompt-craft becomes a comfortable substitute for the harder parts of writing.

2026-05-06 — Researcher contribution: responsibility-uptake plus interpersonal sharing

Source: session 2026-05-06 (same message) Stage: finding Promoted-to-paper: pending — candidate addition to the Author's Note or to the Statement of authorship.

"Another contribution — the actual uptake of responsibility for the results and interpersonal sharing (e.g. conferences, discussions, panels...)."

This complements the third item in the model's perspective on the human contribution (staking reputation under a published name). The interpersonal-sharing leg is missing from that list and should be added. The Author's Note currently names "stakes professional reputation"; it could be extended to "stakes professional reputation and carries the work into the rooms where it has to be defended in person".

2026-05-06 — Gambling metaphor; bank-always-wins

Source: session 2026-05-06 (same message) Stage: observation Promoted-to-paper: no — too provocative for the paper's institutional register; lives here as honest context.

"Or is the whole thing more like gambling with the mindset 'just play long enough and you will score big' (although the bank always wins — in this case current AI model service providers, at least with the current business model)."

This is the dual-use surface of §, restated from the researcher's seat. F(AI)²R does not solve the business-model question. It makes the spend visible (the sustainability section now asks for an inference-budget disclosure) but the upstream economics are not within the paper's reach. Worth keeping in this log as a reminder that the methodology exists inside a market.

2026-05-06 — Afterthought crafting of prompts over hours

Source: session 2026-05-06 (user message after the position-paper Conclusion was sharpened) Stage: observation Promoted-to-paper: pending — candidate paragraph for paper/sections/evolution.tex or for paper/sections/authors-note.tex.

"Afterthought crafting of prompts over hours. Growing idea by idea."

The agent prompts under agents/ are not designed top-down; they accumulate. Each session adds a binding clause (chapter granularity, DLR voice rules, page budget, Consensus mandate, paywall escalation, contribution tracking) and the prompt grows by accretion. This is the same pattern as the verification ladder itself — the AI-CONFIRMED rung was added when the cost of single-tier verification was felt. Implication: the prompts are themselves a research artefact whose evolution is worth tracing. A future analysis pass could query the git log for agents/*.md and produce a histogram of "clauses added per session", giving an empirical curve for "experience with the tool". Candidate for the provenance_analysis.py extractor in a follow-up commit.

2026-05-06 — Coding with LLMs: structure becomes more important; visions, feature sets, dependencies as the human's load

Source: session 2026-05-06 (user message after the slide-decks PR #16 merged) Stage: hypothesis Promoted-to-paper: pending — candidate to flag in the Conclusion's The longer arc paragraph as a generalisation of the F(AI)²R partition beyond scholarly writing, OR as a footnote to the coupling-rule paragraph in paper/sections/authors-note.tex. We mark it pending so the paper takes one stand at a time.

The user observed: "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 (cooperation-as-conversation, accelerator-or-blender, latency-as- feature). Both surfaces reveal the same partition: human at the architectural layer, model at the executable layer. For writing, the architecture is the argument and the verification ladder; for coding, it is the dependency graph, the type system, and the test surface. The "river" metaphor is exact: with prose, the river is the prose-stream and the human shapes the argumentative direction; with code, the river is the code-stream and the human shapes the architectural direction.

Why structure becomes more important counterintuitively: when the model generates fluently, the bottleneck shifts from production to integration. A line of code (or a sentence) in isolation is cheap; a line that fits the surrounding system is not. Without an explicit structural commitment from the human (visions, feature sets, dependencies in coding; position, eight practices, verification ladder in writing), the model produces fragments that locally make sense and globally do not. The river floods. With explicit structure, the river irrigates.

Generalisation of the coupling rule. The accelerator-vs-blender coupling rule from the Author's Note transposes to coding: accelerator regime = human supplies architecture (interfaces, dependency graph, test surface, naming conventions), model fills implementations; blender regime = human supplies vague intent, model picks a locally-plausible architecture that does not compose, codebase entropy rises. The diagnostic transposes too: would the architecture have been thinkable without the model?

For F(AI)²R: the schema (PROV-O + verification ladder + per-claim graph + handback discipline + the chapter-per-file rule + the agents/ taxonomy) is the structural skeleton — the analogue, in cooperative writing, of an interface contract for cooperative coding. The eight integrated practices are essentially the architectural spec.

Recommendation: flag the generalisation as future work in the Conclusion's The longer arc paragraph (one sentence: "the partition and the coupling rule appear to apply to long-form cooperative coding under the same diagnostic, though we do not defend that claim here"), and leave the full observation in this log. The cleaner paper takes one stand at a time.

The user extended: "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."

Important addition. This is the vocabulary-fluency-as-collaboration-multiplier finding. The coding-side accelerator regime depends not just on the human supplying architecture, but on the human having the precise vocabulary of the architecture — interfaces, dependency graphs, data structures, design patterns, type-system idioms — so that prompts can refer to specific things rather than gestures at vague things. The model locks onto specific terms; the river bed widens or narrows in proportion to how specifically the human can describe what flows through it.

Two implications follow. First, cooperation quality scales with vocabulary fluency: a researcher cooperating in a domain whose formal vocabulary they have not internalised is more vulnerable to the cultural-blender pole than one cooperating in their home domain. The vocabulary itself is a kind of API to the model. Second, this adds a second layer to the equity discussion in §: it is not just access to frontier models that is asymmetric, it is access to the vocabulary that makes the cooperation effective. Researchers from disciplines with rich shared formal vocabularies (CS, mathematics, physics, well-formalised parts of biology) start with an advantage that does not transfer to disciplines whose vocabularies are still being built (much of the humanities, parts of the social sciences). The pattern is therefore even less "neutral" than the equity paragraph already concedes.

For F(AI)²R: the eight integrated practices are themselves a vocabulary for cooperative writing. The act of naming them — transcript preservation, verification ladder, handback discipline, mirror discipline, etc. — gives a future researcher a set of API terms to reach for when working with an LLM on a paper. The paper's contribution, on this reading, is partly the practices and partly the vocabulary that lets a researcher operate them precisely.

2026-05-06 — Latency as cognitive feature, not bug

Source: session 2026-05-06 (user message after the "The longer arc" paragraph was merged into the Conclusion) Stage: hypothesis Promoted-to-paper: pending — strong candidate for one or two sentences in paper/sections/authors-note.tex extending the coupling-rule paragraph, or for a new field-note in paper/sections/evolution.tex.

The user observed: "I think 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."

The model offered a perspective in chat. The observation is sharper than it looks: latency between prompt and response forces the user into formulation work — "what do I want to say next?" — which is the abstracting layer of the partition the Author's Note describes (§authors-note, "What surprised me about the cooperation"). If responses were near-instantaneous, the user-as- co-author would degrade toward user-as-consumer: read, approve, scroll, repeat. The decision points the deciding layer needs would compress out of existence. There is a slow-tools-beat-fast-tools literature on this in cognitive science — constraints, intermediate states, and unhurried loops often improve creative output rather than reduce it. Programming has the same shape: compile-and-run cycles create decision points an instantly- evaluating REPL collapses.

Counter-claim worth recording: instant response might be a different kind of beneficial — closer to typing assistance, where the human stays on the abstracting layer the whole time and the model handles low-level wordsmithing transparently. In that mode the latency-bonus disappears because the model has become part of the keyboard rather than part of the conversation. Whether that is better depends on whether the human can hold the abstracting layer without the periodic pause prompt-craft currently provides.

For F(AI)²R: the partition is sustainable in part because of the friction that lets the human stay on the deciding side. If we lose the friction, the partition needs a different defence than "the model takes a moment to respond". Recommendation: extend the coupling-rule paragraph in authors-note.tex with one sentence naming latency as part of the defence, and flag the counter-claim honestly so the paper does not pretend the question is settled.

The user extended: "and allows for short context switches — e.g. checking on a coding agent."

Important addition. The latency window is not idle waiting; it is productive parallel work on another track. Cooperative writing sessions are typically running alongside other slow loops — a coding agent compiling, a CI workflow finishing, a paper PDF rebuilding, a colleague's Slack thread waiting for a reply. The seconds-to-minutes wait between prompt and response is exactly the right shape for those context switches: too short to start something genuinely new, just long enough to glance at one of the other parallel tracks. The cognitive overhead of the switch itself is amortised across the wait that was happening anyway.

This is a generalisation of the slow-tools-beat-fast-tools finding: the slow loop is good for the work-currently-in-hand and good for the supervisory glance at the work-in-progress elsewhere. Instantaneous response would collapse both gains simultaneously — no formulation pause for the current task, no natural breakpoint to check on the parallel one. The cooperative- writing pattern, in this light, is part of a broader multi-tracked-research workflow that the human author is running across several slow loops, not a thing happening in isolation.

Implication for the paper: the eight integrated practices include "transcript preservation" but say nothing about the researcher's actual work-pattern (which loops they are running in parallel, how the cognitive switching cost is amortised). A candidate ninth practice — or a footnote to the existing eight — might name the supervisory-glance pattern as part of why the cooperation works. We mark this for a future revision.

2026-05-06 — Accelerator or cultural blender? — the cognition question

Source: session 2026-05-06 (user message after the position-paper completion was merged) Stage: hypothesis Promoted-to-paper: pending — strong candidate for an extension to paper/sections/authors-note.tex (sharpens What surprised me about the cooperation).

The user observed: "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. The honest answer is: probably both, conditional on where in the cognitive stack the model is allowed to operate. LLMs interpolate fluently within the manifold of prior text; for ideas that lie on or near that manifold — F(AI)²R itself is one — the model is a structuring accelerator. LLMs do not extrapolate well to ideas off the manifold; the failure mode there is exactly the cultural-blender one (smooth the unfamiliar into the nearest familiar shape). The diagnostic is a coupling rule: keep the model on the prose-craft side of the partition and the answer is closer to accelerator''; let it across to the abstraction side and the answer is closer toblender''. The Author's Note's partition — the human does the deciding, the model does the drafting — is therefore not just a description of how this paper was written; it is the rule that keeps the cooperation on the accelerator side of the question. The honest test for any given collaboration: ask whether the abstraction would have been thinkable without the model. If yes, the model was an accelerator on that abstraction; if no, the question genuinely opens.

Recommendation: extend the Author's Note's What surprised me about the cooperation paragraph with one sentence stating the coupling rule explicitly. The current paragraph already names the partition; this observation supplies the reason the partition matters.

2026-05-07 — Why audited defects are pedagogical, not embarrassing

Source: session 2026-05-07 (provenance-curator repair pass on the 8 claims surfaced by the topology page's Worked verification example) Stage: hypothesis Promoted-to-paper: pending — strong candidate for a sentence in paper/sections/provenance-analysis.tex once a second audit cycle confirms the pattern.

The provenance-verification pass surfaced 8 fair2r:Claim entities missing prov:wasGeneratedBy. The repair pass classified each one and a clear pattern fell out: claims that are graduated from researcher quotes into existing sections, or that ride alongside meta-decision PRs, tend to skip the section-authoring activity — the curator has no act:author-* parent to attach to. Five of the eight needed a synthetic act:meta-cooperation-<date>-<slug> activity minted at repair time; the other three had a real act:rev-* parent but the curator pass dropped the edge on the claim while keeping it on the section entity.

The model offered a perspective in chat. The cooperative-process implication is uncomfortable but useful: an audit that surfaces nothing has stopped looking. Defects that surface, get classified, and are repaired in the open are the audit doing its job; the embarrassment-vs-pedagogy framing is a design choice the project makes about its own posture. The substantive rule the pattern suggests is concrete: every claim-add step must mint or attach to some activity, even a synthetic act:meta-cooperation-<date>-<slug>. The provenance-curator agent prompt should grow a refusal condition: refuse a claim with no parent activity. The schema already supports this; the curator just needs to enforce it. The topology page itself was rewritten in this session as a two-state walkthrough (BEFORE table preserved, AFTER table added, work-in-progress disclaimer) to make the audit-then-repair cycle visible to a reader of the public site.

Recommendation: (a) add the refusal rule to agents/provenance-curator.md: no fair2r:Claim triple without a prov:wasGeneratedBy edge — if no natural act:author-* or act:rev-* parent exists, mint a synthetic act:meta-cooperation-<date>-<slug> activity at claim-add time; (b) consider promoting one sentence to paper/sections/provenance-analysis.tex after a second audit cycle confirms the pattern; (c) flag the next likely defect class (prov:wasDerivedFrom missing on verif:ai-confirmed / verif:source-vendored claims) in the same observation so the audit programme has a forward edge.