Directed graph of the CCQ collaboration: governance (blue, black border), contributors (green), invited scientists (amber), unavailable (red), candidates (gray). Paper versions shown as boxes. Hover any node for details. Drag to explore. Updated as the collaboration evolves.
A case study in evolving Human/Natural/Artificial Co-Intelligence.
On the morning of March 4, 2026, Jon Schull typed a question into claude.ai:
"Given that blackbody radiation increases as the fourth power of temperature, and given that the greenhouse effect is tuned to infrared radiation, and given that bare ground is much hotter than vegetated ground — what is the impact of vegetation on global warming?"
He added: "Critically research my premises and aim for an estimate couched as a percentage change in energy captured by the greenhouse effect."
This was not a policy question. It was a physics question — an attempt to ground-truth a physical intuition using the Stefan-Boltzmann law. The answer would turn out to be more consequential than either party expected.
What began as a request for a quantitative check became, over the next nine hours, a systematic multi-mechanism analysis. The session moved through the physics layer by layer: surface temperature differentials and the T⁴ forcing premium (the "Boltzmann factor"), then latent heat redistribution through evapotranspiration, then cloud albedo driven by vegetation-sourced aerosols, then the marine analog — ocean phytoplankton as biological cloud-condensation nuclei, and Hansen's recent analysis of marine cloud decline following the 2020 IMO shipping regulation.
That last mechanism — the unintended global experiment of cutting marine fuel sulphur by 80% and watching the planet warm measurably within months — became the paper's rhetorical spine. It was empirical proof that aerosol-cloud forcing was real, large, fast, and had been missing from model predictions. Not a theoretical vulnerability in carbon-centric climate accounting — an observed failure, already in the literature.
By the end of the Mar 4 session, the assistant had produced a quantified comparison table with 23 footnotes covering six distinct cooling mechanisms, each estimated in W/m². Jon named the goal: "a standalone synthesis with a non-technical narrative for a sophisticated non-specialist, followed by a comparison table, then a technical justification." That document existed — the scientific case was assembled. What remained was to make it a paper.
The Mar 5 session opened at 2 AM UTC. Jon arrived with a commented Word document — he had exported the prior artifact, added margin comments, and was asking for v4 in HTML. The first challenge was technical: the AI couldn't access the Google Drive folder where supplementary materials lived. Jon uploaded a zip. The assistant began integrating.
At message 20, after the assistant had produced a v4 draft with five interactive Chart.js visualizations, Jon issued an instruction that changed the session's character: "Great — now reread the document and become a co-analyst, not just a code and content jockey."
The assistant complied. It reread the full document and came back with a structural critique: the abstract was doing too little work; the shipping story needed to move earlier as the motivating evidence rather than a calibration footnote; the two-audience structure (non-technical narrative followed by technical justification) needed more explicit signposting. This was the moment the collaboration shifted from human-directs-AI to something closer to mutual editing.
Over the next twelve hours, the session moved through v5 (a strategy brief), then v6 (the first full restructured HTML with the W/m² framework as unifying metric, the shipping regulation as opening hook, and the comparison table as visual centerpiece). Jon made a key positioning decision: ERA would not frame itself as a critic of the IPCC carbon framework, but as an expander — adding mechanisms that the current framework underweights. This shaped every subsequent framing choice.
They started fresh — a new Opus Extended Thinking session on Mar 6, opened with a zip containing v6 HTML and two specification files. The specification was itself a product of the collaboration: Jon had articulated editorial decisions into a machine-readable format, and the assistant had helped structure the handoff across the context limit.
The Mar 6 session was the deepest intellectual work. It produced the paper's 4-panel energy budget figure — a disaggregated version of the standard IPCC energy budget schematic, showing each mechanism (bare ground, vegetated ground, ocean, cloud) as a separate panel with W/m² annotations. It also surfaced the Boltzmann extremes insight: a Jensen's inequality argument showing that temperature variability across surface types creates a systematic bias in greenhouse forcing estimates.
At message 120, the assistant identified a problem it had created: "We've been adding excellent material but the paper has grown from a clean 12-sentence argument into something that risks losing its narrative drive under the weight of additions." This self-diagnosis led to a restructuring of the abstract.
The Mar 6 session hit the claude.ai context limit at message 149.
Context limits were the primary friction point throughout the project. Three recovery patterns emerged:
The workflow that emerged — claude.ai for ideation and drafting, specification files as cross-session memory, Claude Code for deployment and deterministic editing — wasn't planned at the outset. It evolved under pressure.
After deployment to the ERA website on March 11, Jon and colleagues reviewed the paper in a shared Google Doc. The doc generated 185 editorial items — structural, substantive, and copy-edit. Claude Code read the Google Doc changes and applied them systematically to the HTML source, with a PR and review trail in GitHub.
The abstract revision — splitting a dense monolithic paragraph into four clear, independently digestible paragraphs — was the most substantive change. It also required updating the same text in three locations in the HTML file, which the CTO handled correctly.
The W/m² framework. Choosing a single quantitative metric as the basis for comparing every mechanism gave the paper internal coherence. Carbon accounting doesn't provide this; it measures stocks (CO₂ concentration) rather than flows (forcing rate). The W/m² framing made the comparison table possible and made the paper's central argument concrete.
The shipping regulation as narrative anchor. A natural experiment, already documented, already in the literature, already surprising to most readers. It made the paper's claim falsifiable and grounded — not theoretical.
The co-analyst relationship. The shift at message 20 of the Mar 5 session, when Jon asked the assistant to become a co-analyst rather than an executor, changed the output quality. The assistant began flagging oversold claims, proposing structural reorganizations, and catching its own errors.
Specification files as cross-session memory. The practice of writing detailed specifications before handing off to a new session — capturing decisions, not just tasks — solved the context limit problem better than any other approach tried.
By late March, the paper had left the lab. Jon began sharing it with scientists and colleagues, asking for review, inviting commentary. The responses arrived in clusters — some by email, some as Google Doc suggestions, some as forty-minute calls with colleagues who had been thinking about exactly these mechanisms for years.
The early responses sorted into types. Philip Bogdonoff, ERA's strategy advisor, had already contributed copy-edits that sharpened the prose without touching the science. Peter Bunyard arrived (Mar 24) with domain-specific corrections on gymnosperm transpiration rates and the latent heat radiation timescale. Rob de Laet turned out to be the most consequential early reviewer — a 55-minute call on Mar 24, followed by a six-mechanism stack with a decomposition of the W/m² estimates and a letter about TOA versus surface heat that reframed how the paper was arguing its numbers.
Anastassia Makarieva — the biotic pump theorist whose foundational work ran through the paper's third mechanism — declined co-authorship (Mar 23) but sent a detailed mechanism critique that was more useful than most acceptances would have been. Her central point: the biotic pump is not "surface cooling." It is a continental-scale water delivery and circulation system that operates at altitude, where the climate effects are global rather than local. Her critique forced a restructuring of the mechanism framing that made the science more defensible.
Others arrived through the network: Stuart Cowan at the Buckminster Fuller Institute (lunch with Jon and Sara Blenkhorn, Mar 28, Berkeley). Frederic Jennings, a PhD economist who offered co-authorship at a board meeting. Didi Pershouse with ten comments challenging the framing of soil mechanisms (Apr 8). Brian von Herzen of the Climate Foundation, whose marine permaculture work became Appendix I and who became a confirmed co-author with editor access (Apr 11). By April 2026, the contributor circle had grown to twenty-plus people across six countries. On Apr 7, v10e was sent to the full reviewer list — the first version circulated widely — along with formal invitations to Douglas Sheil, Michal Kravčík, Stefan Schwarzer, and others cited in the paper.
On April 9, Jon convened a governance meeting with Philip Bogdonoff and Ananda Fitzsimmons. The question on the table was strategic: what kind of document was this, and how should the collaboration be formalized?
The discussion surfaced an insight that Philip named and articulated: that the paper's credibility could be amplified not through traditional peer review — which would take eighteen months — but through a deliberate coalition-building process. CCQ would be published as a Creative Commons resource, and the scientists whose work it synthesized would be invited to build on it in their own peer-reviewed work, on the condition that they cite it. This frees collaborators to publish their own variants without rights fights, acknowledges that the paper is a synthesis across many antecedents (including substantial AI drafting), and sidesteps the authorship-credit debate that can stall coalitions of this size. The working principle for growth: existing collaborators nominate new ones; the Governance Team stages the invitations. The key was sequencing — show the paper to scientists with intellectual respect rather than institutional prestige, and let the coalition grow from that foundation.
This became the organizing principle for the paper's next phase. A governance working document was created that afternoon — covering the rollout plan, a contribution taxonomy (distinguishing co-authors from reviewers from acknowledgments), a cover note template for scientist invitations, and an appendix on the role of AI in the process. The governance spreadsheet was extended to 21 rows. A kickoff email went to Jon, Ananda, and Philip that same evening. On Apr 11, v10f was published — 29 revisions driven primarily by Ali Bin Shahid's biome-specific analysis — and Brian von Herzen's introduction of Leon Simons marked the first scientist reached via the coalition network rather than Jon directly.
Rob de Laet had been thinking about the coalition problem from the beginning: a paper this interdisciplinary, with this many contributors and potential contributors, needed a way to visualize who knew whom and who could open which doors. His Apr 12 proposal: a "flash mob" model — a staged sequence of personal outreach to prominent scientists, timed for coordinated social proof.
The first artifact of this strategy was a grid. Rows were the names he wanted to reach — Tier 1 scientists like Hansen and Hayhoe, literature scientists whose work was already cited, influencers and journalists who could amplify. Columns were the current co-author circle. Cells were connections: strong ones and casual ones. Filled in collaboratively, the grid would become a map of the coalition's relational reach. The CCQ agent built this as a Google Sheets tab — the "Connections" sheet, 54 rows × 31 columns, color-coded and formatted.
But the grid was a static view of a dynamic reality. The collaboration had a history — who joined when, who introduced whom, which version of the paper they contributed to. That history was a directed graph. Jon noted that the ERA website's landscape page already used vis.js for network visualization, and suggested borrowing from that code.
What followed was a half-day of rapid prototyping. The ERA landscape graph was undirected; the CCQ contributor network was inherently directed — introductions flow from people who know each other to people who don't, contributions flow from contributors to versions, co-authorship has a history. The directed vis.js network took shape: people nodes, version nodes, a left-to-right timeline layout, edges encoding relationship type through line style and color.
The visualization at the top of this page encodes the collaboration's full history: Jon as progenitor of v9, the governance circle with black node borders, contributors in green, scientists invited in amber, those unavailable in red, candidates in gray. Hover over any node and an info panel appears — name, role, affiliation, contribution, date joined. The network graph is itself an artifact of the process it documents — generated by the same AI-human collaboration it depicts, and updated as the collaboration evolves.
By mid-May the paper had been through a dozen internal drafts, and it was beginning to sag under its own additions — the same failure mode the assistant had flagged back in March. On May 17, v10g began as a whole-paper consistency pass: every quantified claim reconciled against every other, the headline range stabilized at 25–40% of total anthropogenic forcing, with the ocean term deliberately held outside it pending marine review. The next day Jon articulated the editorial rule that would govern everything after — the abstract-sentence spine. The abstract is a sequence of sentences; each sentence is the thesis of one major section; every passage in the paper either proves its section's abstract sentence or it is apparatus. Anything that proves nothing gets cut or demoted. It was a test sharp enough to apply mechanically, and it turned editing from taste into arithmetic.
v10h was the structural rework that followed. On May 21, for the first time, the working document was opened for direct editing — not just comment — to the full co-author circle: seventeen writers across the contributor network. Jon sent the announcement himself, with his own additions woven in. The paper stopped being a thing shown to people and became a thing built by them.
The next movement came from a question. Philip Bogdonoff asked, in effect, how a mechanism the paper called fast and powerful could have stayed hidden across eight hundred thousand years of climate history. It was the right kind of challenge — not an objection to the science but a demand that the science account for the long record. The answer became a new passage in §5: before the industrial timeline, ecosystem cooling was already operating, visible in the paleoclimate record as a slow feedback rather than a sudden forcing. The distinction that matters, the passage argues, is driver, not magnitude — and the record is not a contradiction of the thesis but the baseline against which the modern rupture becomes visible.
Building that passage pulled three scientists back into the work within a single week, through a pattern worth naming. David Ellison, re-engaged on May 23, surfaced his own 2024 paper (with Pokorný and Wild) and pointed to Kaplan's and Arneth's reconstructions of pre-industrial land-use forcing. His thread carbon-copied Jan Pokorný — a scientist invited in March who had gone quiet after testifying to the Czech Parliament — and within forty-eight hours Pokorný returned with fifteen years of his own field measurements: incoming shortwave up ten percent, longwave from the surface up twenty, less cloud, less fog, energy shifting from latent to sensible heat as the landscape dried. The warm thread had done what a cold reintroduction could not. Rob de Laet, meanwhile, contributed the earlier Potsdam-CLIMBER cluster — Ganopolski's, Kubatzki's, and Brovkin's late-1990s modelling of exactly this slow biosphere feedback — and the line that now closes the passage: "The record is not a contradiction. It is the baseline against which the rupture becomes visible."
The paper had always carried a marine mechanism — phytoplankton seeding clouds the way forests seed them over land — but the ocean side was held loosely, its forcing range ring-fenced outside the headline number until the marine reviewers weighed in. In late May they began to. Peter Bunyard, who had arrived in March with corrections on transpiration rates, returned with two slides from a decade of talks in Colombia: the chalk-white albedo of coccolithophore blooms, visible from orbit off Ireland and the Bering Strait, and a diagram of the dimethyl-sulphide cloud thermostat with a temperature bifurcation — stabilizing below about twelve degrees, runaway above it. The contributions were folded in with their caveats intact: the cloud-feedback leg of the loop is contested in the literature, and coccolithophores calcify even as they reflect. The paper marks what is established and what is exploratory — the same discipline it applies to its own numbers.
Running underneath all of it was a corroboration the paper had not engineered. James Hansen's spring 2026 analysis raised the estimate of climate sensitivity and named the missing factor as aerosol-cloud interaction — the precise blind spot the paper had been arguing was hiding inside carbon-centric accounting. It arrived not as endorsement but as convergence: an independent line reaching the same conclusion from the other direction.
In early June the work turned outward. A clean, read-only edition of v10h was published — the same text with its editing marks stripped and its provenance made legible, a header that links the read-only copy and the living working draft as two views of one document. The landing page was rebuilt around a single figure from the deck — ten thousand years of accumulated ecological debt — with three doors into the work: the paper, the deck, the film. The document you are reading was brought current in the same pass. That recursion is not incidental: a paper arguing that we have been mismeasuring the planet is, in its own making, a small argument that the process of knowing can be measured differently too — in the open, with the record kept.
The traditional model of scientific publication is solitary to a fault: one or a few authors, years of work, peer review as the sole quality gate, publication as the moment of release. The social infrastructure that actually determines what gets read and cited — who knows whom, who endorses, who amplifies — is invisible and inaccessible to outsiders.
Cooling Climate Quickly is running a different experiment. The paper is being built in public (within a permissioned circle), with its governance structure documented, its contributor history visible, and its process archived. The making-of document you are reading is itself part of the artifact. So is the network graph at the top of this page.
Several features of this process are worth naming as lessons, not just observations:
The paper argues that ecosystem restoration can cool the climate faster than carbon accounting suggests. The process of making it demonstrates something adjacent: that AI-facilitated open science can assemble knowledge faster, more transparently, and with broader collaboration than the traditional model allows. Neither claim is proven yet. Both are under active investigation.
This archive documents the making of a paper that argues ecosystem restoration can cool the climate faster than carbon accounting suggests. The process of making it was itself a demonstration of the thesis: that collaboration — between human judgment and AI capability, between strategic vision and technical execution — produces results neither could achieve alone.