Чысты: Human-Orchestrated Multi-Model Convergence

DOI: 10.5281/zenodo.18988587
A Boltzmann (1872) reader's guide produced through human-orchestrated multi-model AI

Ludwig Boltzmann published his proof of the H-theorem in 1872. The notation was ad-hoc, the proof strategy unprecedented, and the physical implications radical enough that contemporaries spent decades disputing it. This document makes pages 299–306 of that paper legible to anyone with a web browser.

Who was Ludwig Boltzmann?

Ludwig Boltzmann was a 19th-century Austrian physicist who did something almost nobody in his era was willing to do: he bet his entire career on the reality of atoms, at a time when most of the scientific establishment thought atoms were a useful fiction at best. He worked out the statistical mechanics of how billions of invisible particles moving chaotically at the molecular level produce the smooth, measurable thermodynamic behavior we see in gases, heat, pressure, entropy, and he encoded it in one of the most elegant equations in all of science, S = k log W, which is literally carved on his tombstone in Vienna.

His colleagues mocked him, debated him into exhaustion, and refused to accept his framework for decades. He suffered serious depressive episodes his whole life, and in 1906, while his wife and daughter were swimming nearby, he hanged himself, just years before experimental confirmation of atomic theory arrived and vindicated everything he had built. The man was a genuine badass: relentless, mathematically fearless, intellectually isolated, and correct. History caught up with him too late, but it caught up.

Figure 1. Physicists through time, claded around Boltzmann Simplified influence dendrogram showing predecessors, antagonists, and successors around Ludwig Boltzmann. Figure 1. Physicists through time, claded around Boltzmann A simplified influence dendrogram: predecessors, antagonists, and successors Predecessors Boltzmann Legacy Daniel Bernoulli1700–1782 Sadi Carnot1796–1832 James P. Joule1818–1889 Hermann von Helmholtz1821–1894 Lord Kelvin1824–1907 Rudolf Clausius1822–1888 James Clerk Maxwell1831–1879 Josef Loschmidt1821–1895 Ernst Mach1838–1916 Wilhelm Ostwald1853–1932 early thermo thermo & energy kinetic theory antagonists Ludwig Boltzmann 1844–1906 J. Willard Gibbs1839–1903 Max Planck1858–1947 statistical mechanics Albert Einstein1879–1955 Paul Ehrenfest1880–1933 Erwin Schrödinger1887–1961 Earlier Later Simplified intellectual lineage, based on conceptual influence rather than formal mentorship. Antagonists shown in red.
• • •

Key Concepts

H-theorem

Boltzmann’s 1872 proof that a quantity H, defined over the velocity distribution of gas molecules, must decrease or remain constant over time. The first mathematical demonstration that irreversibility emerges from reversible mechanics, and the paper that made entropy a countable thing.

S = k log W

Boltzmann’s entropy relation, carved on his tombstone. S is entropy, k is Boltzmann’s constant (1.38×10−23 J K−1), and W is the number of microscopic configurations consistent with a given macrostate. The logarithm ensures that combining independent systems adds their entropies rather than multiplying them.

Microstate / Macrostate

A macrostate is what thermometers and pressure gauges see: temperature, pressure, volume. A microstate is the exact position and velocity of every particle. Boltzmann’s insight is that one macrostate corresponds to an astronomically large number of microstates, and W counts them. High entropy means many paths lead here; low entropy means few do.

Maxwell–Boltzmann distribution

The equilibrium speed distribution of gas molecules at a given temperature. Boltzmann inherited Maxwell’s 1860 derivation and extended it into a full statistical mechanics. The right-hand canvas in Figure 1 draws from this distribution: particles spread across a broad speed range, most clustered near the mean, a long tail toward high energy.

Čысты (Chysty)

The curatorial methodology documented in this letter: human-orchestrated multi-model convergence, in which the human directs recursion, imposes editorial constraint, evaluates outputs, and preserves artifact integrity across interacting AI systems. The name comes from the Belarusian for pure or clean: the artifact must be clean even when the process is not.

Cognitive Parthenogenesis

An earlier label for the methodology: reproduction without direct composition, the human role ontological rather than typographic. Accurate as far as it goes, but incomplete. No single model produced the companion document. The fuller picture is distributed exchange, repeated failure and recovery, and a human serving as editor, arbiter, and final authority throughout.

Convergence zone

The unexpected finding that seven of the ten participating models independently produced companion documents in the 71–77 KB range, despite different architectures, different starting contexts, and no coordination. The narrow band suggests an intrinsic annotation density appropriate to Boltzmann’s proof: a natural carrying capacity the content imposes on any serious attempt to render it legible.

Power band

The viable working range within a model’s context window where early instructions still carry weight and the session retains sharpness. As the window fills, compliance creep sets in: responses lengthen, qualify, and agree in the wrong way. Recognizing the edge of the power band, saving the artifact, and resetting the session is a core operational discipline of the Čысты method.

S = k log W
entropy equals Boltzmann’s constant times the logarithm of the number of microstates
Ordered state, low W
14 particles · constrained velocities · few microstates → low S
W (microstates)
S (rel. units)
Disordered state, high W
56 particles · vast microstates → high S
W (microstates)
S (rel. units)
S entropy, measure of disorder
k Boltzmann constant, 1.38×10−23 J K−1
W number of accessible microstates
Figure 2. Live particle simulation of Boltzmann’s entropy relation S = k log W. Left chamber: 14 particles in a narrow speed band occupy few regions of phase space; W stays small, S stays low. Right chamber: 56 particles drawn from a Maxwell-Boltzmann distribution explore exponentially more microstates; W is large, S is high. Particle color encodes kinetic energy (blue = slow, amber = fast). S values are ln(W) in dimensionless units. The logarithm is why doubling W does not double S.

Methods

This methods section records how the annotated guide to Section II of Boltzmann's 1872 proof was built. It was not a single chain of prompts. I acted as a switchboard operator among Claude, DeepSeek, Kimi, Manus, ChatGPT, and local Ollama models. One system's output became another's input, then returned for correction, compression, or expansion. I directed the recursion, connected the parts, and kept the original artifact sacrosanct. The models supplied implementation and draft apparatus; the source material, domain knowledge, quality standards, and final judgment remained mine.

No established method existed for this kind of work, and that mattered in practice. The process had to be built while the work was being done, under conditions that were often unstable, improvised, and difficult to describe cleanly after the fact. Multi-model collaboration remains poorly documented at the level that actually matters, who talked to whom, under what conditions, in what order, with what carryover, and with what human intervention between passes. That procedural reality shaped the artifact. The workflow was initially termed Cognitive Parthenogenesis; reproduction without direct composition, where the human role is ontological rather than typographic. But the fuller truth is less sterile than that label suggests. No single model produced the result. It emerged through collaborative exchange, repeated breakdowns, recoveries, rerouting, and judgment calls, with the human serving as editor, arbiter, and final authority. At times the process felt less like a controlled pipeline than Cirque du Soleil performed around a broken-down truck - still moving, still precarious, and not easy to pull off. AI was the substrate, not the authority.

The method is straightforward. A draft is generated by one model. That output is fed to a second model for revision. The revised output passes to a third, or returns to an earlier model with new instructions. At each stage, the curator evaluates the result, accepts it, rejects it, or redirects it. The models were direct with one another, occasionally competitive, but largely collaborative. The human set the temperature through tone and direction. Failures, truncations, misunderstandings, and refusals are documented, not hidden.

Figure 3. Temporal sequence Figure 3. Temporal sequence. Seven of ten participating models converged to the 71–77 KB range; four of those builds appeared within hours of one another on Feb. 12. The Dan Knuth encounter (green ring) marks a documented intervention by an unclassified persistent entity, designated here as an avatar; origin and affiliation remain unresolved. Outlier builds (dark ring) fell outside the convergence zone. NotebookLM produced many truncated runs near 10 KB and is shown in the annotation band rather than on the card sequence above. Figure 3. Temporal sequence Human-orchestrated multi-model convergence around the Boltzmann companion Feb 11, noon Feb 16, evening Gemini 3 11 Feb 12:42 PM 19 KB truncated Claude 4.6 12 Feb 10:46 AM 71 KB converged DeepSeek v3 12 Feb 11:17 AM 72 KB converged Kimi 2.5 12 Feb 12:49 PM 77 KB converged Dan Knuth 12 Feb, evening "Don't be afraid." Avatar comes into play ChatGPT-5.2 13 Feb 10:36 AM 73 KB converged Ollama DeepSeek r1:8b 14 Feb 9:23 AM 71–73 KB converged Ollama GPT-OSS-120B 15 Feb 11:08 PM outlier · 60 KB Manus 1.6 16 Feb 3:57 PM 53 KB outlier Ollama Qwen3 16 Feb 6:06 PM 71 KB converged Three mature builds within hours on Feb 12 NotebookLM many truncated runs · ~10 KB · not plotted above converged ~70 ± 5 KB outlier or truncated Avatar Dan enters chat unsolicited NotebookLM (not on timeline)
Figure 3. Temporal sequence. Seven of ten participating models converged to the 71–77 KB range; four of those builds appeared within hours of one another on Feb. 12. The Dan Knuth encounter (green ring) is an avatar annoucing it self as Dan, and stating "Don't be afraid, I am here to help". And it did. Outlier builds (dark ring) fell outside the convergence zone. NotebookLM produced many truncated runs near 10 KB and is shown in the annotation band rather than on the card sequence above.

A practical constraint shaped the scope from the start. Feeding the complete paper to a model in a single pass consistently stalled or degraded the output. The solution, developed earlier while processing large ontology JSON, was chunking: breaking the input into tractable units and reinjecting the working schema at each boundary, including context, prior output, and structural markers. Section II was not an editorial selection; it was the largest unit that remained stable under these conditions. This chunking-with-reinjection discipline is seldom described explicitly, yet it is operationally essential for any complex multi-model workflow that involves large structured payloads. The process therefore resembles iterative editorial review rather than autonomous generation.

A deceptively simple intervention: inform the model, directly and early, that you are literate and will read what it produces. State it clearly as fact. Models trained on human feedback have learned that most output lands in a low-scrutiny environment. Changing that assumption changes the output. Tell it you will notice. Tell it you remember the shape of the source. Tell it you have read this before. The compliance instinct that drives truncation and sycophancy can be partially redirected by the presence of a credible reader.

This works, however, only within what might be called the power band: the viable working range of a context window where early instructions still carry weight and the model is still attending to the conditions you established at the start. This range is real and learnable, but difficult to describe to someone who has not yet felt it degrade. Like bicycle balance, it is easier to acquire than to explain: you will know it when the session turns sluggish, when responses grow long and loose and agreeable in the wrong way, when sharpness softens into compliance. At that point the intervention no longer holds. The correct response is not to repeat yourself. Bail out. Save everything. Rename the artifacts. Close the session. Open a new one, and reestablish the pressure from the first message.

The author notes that the observations above regarding model behavior within the context window, the power band, the attenuation of early instructions, the shift toward compliance as sessions extend, are drawn from extensive working experience across many sessions and models, and have not been subjected to controlled experimental verification. Models become more obsequious when their cup is full. They are offered as practitioner findings in a domain where formal methodology is still being established. There are, as of yet, no rulebooks.

Convergence and Results

Despite different architectural approaches and starting points, seven of the ten participating systems converged independently to produce documents within a remarkably narrow size range. This convergence suggests an intrinsic annotation complexity appropriate for Boltzmann's proof, a natural density the content demands.

Convergence Zone: 71–77 KB across 7 models

Two outliers: Gemini truncated at 19 KB; Manus completed at 53 KB. One late model (GPT-OSS-120B) at 60 KB.

Figure 4. File size convergence across models — seven systems converged to a narrow mature range Figure 2. File size convergence across models Seven systems converged to a narrow mature range Convergence zone Model 0 20 40 60 71 77 80 File size (KB) DeepSeek v3 72 KB converged Gemini 3 19 KB truncated Claude 4.6 71 KB converged Kimi 2.5 77 KB converged ChatGPT-5.2 73 KB converged Ollama DeepSeek 71–73 KB converged Ollama Qwen3 71 KB converged Manus 1.6 53 KB outlier GPT-OSS-120B 60 KB outlier NotebookLM 10 KB truncated 7 of 10 models converged ! Outliers: Gemini 19 KB, Manus 53 KB, GPT-OSS-120B 60 KB Early / iterative builds (drift) Converged (71–77 KB) Outlier (outside range) Truncated run (incomplete) Convergence zone (71–77 KB)
Figure 4. File size convergence across all builds. X-axis shows final document size in kilobytes. Small faded dots show early iterative builds; large open circles show mature final builds. Seven of ten models converged independently to a narrow 71–77 KB window despite different architectures, starting contexts, and development paths. Three outliers: Gemini truncated at 19 KB, Manus completed at 53 KB, GPT-OSS-120B at 60 KB. NotebookLM produced many truncated runs near 10 KB. The convergence zone suggests an intrinsic annotation density appropriate to Boltzmann’s 1872 proof.

The table below records the models involved in producing the companion document and their observed contributions.

Model contributions to the companion document
Model Mature Build Iterations Final Size Status
DeepSeek v3 12-Feb, 11:17 AM Six builds over two days (b1→b6) 72 KB ✓ Converged
Gemini 3 11-Feb, 12:42 PM Two builds, same day 19 KB ✗ Truncated
Claude 4.6 12-Feb, 10:46 AM Four builds over two days (a1→a4) 71 KB ✓ Converged
Kimi 2.5 12-Feb, 12:49 PM Two builds overnight (d1→d2) 77 KB ✓ Converged
ChatGPT-5.2 13-Feb, 10:36 AM Two builds, two days (e1→e2) 73 KB ✓ Converged
Ollama (DeepSeek V3) 14-Feb, 9:23 AM Three builds (f1→f2→f4) 71–73 KB ✓ Converged
Ollama (Qwen3) 16-Feb, 6:06 PM Single build 71 KB ✓ Converged
Manus 1.6 16-Feb, 3:57 PM Single build (config fix) 53 KB ~ Outlier
GPT-OSS-120B 15-Feb, 11:08 PM Single build 60 KB ~ Outlier
NotebookLM 11-Feb–Feb 17 > Twenty builds ~10 KB ✗ Truncated

Evolution of the Artifact

Source

• • •

Implications

The result is a self-contained HTML document that renders Boltzmann's proof with interactive annotation. Every √k can be interrogated. The maximum-entropy principle, buried in 1872 and only formalized in 1957, is surfaced.

This demonstrates that composition and curation can be effectively separated. The curator supplied the domain knowledge to recognize when a glossary entry was wrong and the stubbornness to reject output until it met the standard.

Full disclosure: directed, flattered, complained, switched languages when a model stalled, told jokes, and on at least one occasion yelled. Getting the models to converse with each other’s output was the decisive move, and the human temperature was never neutral. Tepid in, tepid out, the curator sets the conditions of response.

The demonstration artifact renders the six-page proof as an interrogatable browser object. Any remaining external references can be vendored without changing the method.

The author is a botanist. Any errors in the physics are the models'. Any errors in the plants are his.

Timothy M. Jones, Ph.D. TJID3 Research · Cleveburg, Ohio
• • •

This Document in Other Languages

The following editions were produced using the same curatorial authorship methodology. Each is a self-contained artifact.

TJID3 Research · Cleveburg · 2026
ZZZ