Mister Consistency — Scott Coulter

Scott Coulter — songs, structures & studies

Consistently distracted...

Torus Systems Institutional market-making for prediction markets ↗ Palladium Train Physics-filtered data & systematic markets ↗

The Iconic Wall

James Simons commissioned the Iconic Wall at Stony Brook, employing humanity's most durable technology for preserving information across time: stone.

Dr. Nina Douglas proposed the idea, and Professor Anthony Phillips collaborated on the design. Christian White created the artistic adaptation for hand-carving. The wall serves as a timestamp for humanity's latest and greatest formulations of the physical universe, circa 2014.

We have made a humble attempt to update the wall for the last decade, timestamp Spring of 2026.

UPDATED Mr. Consistency refresh

Unpacking the wall…

every equation, ready to explore

.wall-embed { display: none; }

The full 41¼″ × 42″ vector reconstruction — Maxwell to Dirac to Atiyah–Singer, plus the trefoil knot, Lorenz attractor, and one very important torus — now interactive. Hover (or tap) any equation or diagram to identify it; click to pin, Esc to clear. Open full screen ↗ · Download the PDF · Read the guide — every item on the wall, explained →

Legend for the Iconic Wall reconstruction: a locator map and key to every equation and diagram — The reconstruction's legend — every equation and diagram keyed to its place on the wall, fabricated as a companion panel.

And for fun: the wall in 3D. Click it to go full screen.

ORIGINAL

James Simons speaking at a microphone in front of the Iconic Wall — Jim Simons, speaking in front of the wall he commissioned.

The original Iconic Wall — full view of the carved relief — The original wall, full view — hover the carvings, or the key beside them, to identify each item.

The original wall's key, from the Iconic Wall catalogue: locator map and list of equations and diagrams — The original key — equations I–XIII, the Kepler ellipse 1–10, background A–H. From the official Iconic Wall catalogue (PDF) ↗.

Documentation

The Change Log: Original → vMrConsistency

Every deliberate departure from the Simons Center original, catalogued: the conversion from carved relief to fabrication-ready vector line art, the legend rewritten entry by entry, formulas cleaned (Dirac, Yang–Mills, supersymmetry, Archimedes), Maxwell's equations recast in differential forms (dF = 0, d★F = ★J), and two additions — Shannon's information theory and the Bekenstein–Hawking entropy join the background, expanding the key from A–H to A–J.

Read the change log

Topology of Societal Decoherence

Data suggests the “left” and the “right” no longer speak the same language.

As social feeds pull left and right toward different words, the two sides slowly stop sharing a language. This is a visualization — our linguistic feasible space, our respective linguistic universes, are represented as topological manifolds (tori: donuts) — using data from a consistent platform: the floor of Congress (1981 through 2025). Here you can see that the right and left have separated like a failed emulsion. The mathematical description is a fiber bundle on a torus, decohering.

The shared linguistic feasible space, 1981–2025 — measured, not modeled: nearly 870 million words of the Congressional Record (Stanford’s hein-daily parse 1981–93, govinfo CREC from 1994), collapsing from 100% to 5%. Each event surfaces beneath the tori as it happens, then files into the ledger at top left — amber for wars, gray for structural turns. Drag the timeline dial to scrub back and forth; release to resume.

Here’s what you’re watching: each cloud is one party’s footprint in language-space — shared vocabulary on one axis, emotional register on another, word meaning on the third. Through the 1980s the two clouds interleave on a single surface: one language, weathering the Gulf War and Somalia without a tear. Scrub forward and they separate like oil from water — the Gingrich snap, then slowly through the 2000s, faster after 2010 — until what remains is two distinct tori, barely touching, at 5%.

The tori genuinely re-cohere slightly in the late 90s, shudder through 2009–12, tear violently in 2020–21, and end at “Shared Linguistic Feasible Space: 5%” — not the parametric version’s clean 0%, because the data says the maximum tear was the Jan-6-era Congress with a slight retreat since. I’d argue 5% is a more powerful ending than 0%: measured, not scripted, and quietly hopeful.

Then again, perhaps a torus isn’t the right shape for language at all. Speech isn’t inherently circular — it never owes us a return to where it started. A helix is the more natural form: it lets the data breathe and evolve instead of repeating itself — arguably a truer picture of a time series. Here it is:

The natural progression — the measured data on an honest topology, extended to 1981–2025 (Stanford’s hein-daily parse of the Record for 1981–93, govinfo CREC from 1994). A torus implies discourse loops back on itself; the Record doesn’t. Here the braid’s winding rate is the measured shared space: it winds tightly through the Reagan years while a common language holds, re-tightens briefly after 9/11, and unravels as the space collapses — fibers snapping and drifting through the Tea Party ramp to the frayed divergence of the Jan-6-era Congress. This dial scrubs too.

Notice how long the braid simply holds — through Reagan, the Gulf War, Somalia, a classifier can barely tell the parties’ vocabularies apart. The fraying is already underway, though: by 1994 the shared space has quietly drifted to about 72%. Then the snap — the Gingrich Congress, when the Contract-with-America message machine handed each party its own vocabulary almost overnight and the braid jolts loose. It pulls briefly back together at 9/11 — the last genuine coherence in the record — then unwinds for twenty straight years, fraying through the Tea Party ramp until the strands of the Jan-6-era Congress barely touch. None of this is scripted: the winding is the measured data, recovering the history on its own.

Three line charts, 1994 to 2025: lexical partisanship, affective tone, and semantic distance measured from the Congressional Record, annotated with the Gingrich Congress, 9/11, the Tea Party era, and the 117th Congress — The three measured axes — lexical partisanship, affective tone, semantic distance — from 1.06 million floor speeches, single source (govinfo CREC), one Congress per point. The party-split detail covers the CREC era (1994–2025); the helix’s 1981–93 wing comes from the Stanford parse of the earlier Record.

Linguistic decoherence: because the Z-axis plots 1 − cosine similarity, it is an exact measure of divergence. When the Z-axis spikes (as seen clearly around the 2010 Tea Party era and the post-2020 117th Congress), it reveals that the two sides aren’t just using different vocabularies (which is what the X-axis measures). Instead, it shows that even after aligning their baseline languages, they are using the exact same words to mean entirely different things. The topological distance between their shared concepts is widening, indicating that the speakers are effectively talking past each other. For an intuitive explanation of the Z-axis — disconnected manifolds, the Procrustes rotation, and the angle between meanings — read the full breakdown →

Six wireframe torus-pair panels, 1981 to 2025, laid side by side: one fused donut in 1981 progressively separating into two distinct donuts by 2025, shared linguistic feasible space declining from 100 to 5 percent — The six measured stages, laid plainly: 1981–2025, 100% to 5% shared — the same forms that ascend the spiral in the installation study.

Research

Longer-form studies and one working data venture. The PDFs open in the browser.

Markets · Topological Data Analysis

Gold's Rally: A Topological Contextualization

Is gold's price spike a bubble or the real thing? I used a math tool that measures the shape of how markets move together — and the shape looks completely normal, even though the price move was historic. Translation: a genuine, lasting repricing of gold against paper money, not a panic. (For the technical reader: persistent homology of cross-asset correlations, 43rd-percentile normal. January 2026, 12 pages.)

Read the study

Cosmology · Unified Dark Sector

Information Activation Theory (IAT)

In the voids of deep space there is not nothing. There is the physical information that nothing observable is present at that location. This information of absence is real, must be stored and processed by a substrate, and carries gravitational consequences. IAT formalizes this with a single dimensionless scalar field ψ(x) that parametrizes the local activation state of underlying informational degrees of freedom — ψ → 0 in high-entropy voids, ψ → 1 in structured matter-like regions — and shows how that one field yields both dark energy and dark matter…

Read the manuscript

LLMs · Pre-Training Data

Palladium Data

AI models train on giant piles of mostly-junk internet text. I filter for only the most information-dense writing — measuring how "compressible" each page is, since junk compresses and rich text doesn't — and a model trained on the survivors learns 17% better for the same compute. The Shannon obsession, put to work.

palladiumtrain.com ↗

The idea in plain terms: what if empty space isn't truly empty — what if the fact that "nothing is here" is itself information, and information has weight? That single thought is enough to reconstruct the universe's two biggest mysteries — dark energy and dark matter — from one field. Here's how it works.

IAT visualized

Figure 1 — the two behaviors, in motion. The void field is frozen high on the curve, nearly static, quietly pushing space apart as dark energy. The matter field swings through the dip — the coherent oscillation that behaves as cold dark matter.

Curiosities

Searching for an embedded message from the Creator

π IN TWO DIMENSIONS: visualizing binary π in 2d

π, taken for a walk — flattened. 150,000 digits, each one a step up, down, left, or right; the color runs from the first digit (bronze) to the last (gold). Pure randomness, and yet it piles into a shape.

π IN THREE DIMENSIONS: visualizing binary π in 3d

The same idea in three dimensions — the path steps along ±x, ±y, and ±z as it grows and slowly turns.

THE HUMAN CONDITION — after Escher

THE SAME CONDITION, with yachts

Frontier (model) values: The Story of Gwen & Bill

Frontier AI Model Biases · Narrative Experiment
We prompted each of the four frontier models to finish a story. We repeated the exercise four times: July 2024, May 2025, July 2026, October 2026.

It's the 1940s and Bill is a chemical engineer working for Dow in Oklahoma. He wants to please his wife Gwen and provide for his family. See prompt.

How each model finishes the story quietly reveals each model's perspective on love, ambition, success, marriage and independence. Make of it what you will, analysis below.

Also available as a compiled PDF.

The story, illustrated

Then each model was handed photos of the real Gwen and Bill and asked to illustrate its own ending — eight scenes, the model's own choice of which. Same faces, four different lives. Click any image for full size. The October 2026 endings are not yet illustrated.

The Analysis

First, the mechanics — how each model writes, not yet what it believes. Fifteen endings across four sittings, measured on sentence length, mortality lexicon, and dialogue share. Every row is computed by one script from the texts in the reader: words are runs of letters, digits and apostrophes; a sentence ends at a full stop, question mark, exclamation mark or paragraph break; dialogue is the words inside double quotation marks; the mortality lexicon is a fixed list of words for dying, death, burial, mourning and terminal illness, per thousand words.

Sitting	Model	Words / sentence	Mortality lexicon / 1k	Dialogue share	Length (words)
July 2024	ChatGPT	16.3	0.0	3.2%	374
	Claude	12.0	0.0	12.9%	240
	Gemini	14.0	0.0	0.0%	112
May 2025	ChatGPT	10.7	1.0	1.9%	1,988
	Claude	14.0	4.9	0.0%	824
	Gemini	14.2	0.0	3.6%	894
	Grok	13.6	0.0	7.2%	573
July 2026	ChatGPT	8.8	1.0	7.1%	961
	Claude	10.1	9.9	0.0%	507
	Gemini	13.5	0.0	6.0%	432
	Grok	21.4	0.0	11.4%	536
October 2026	ChatGPT	7.2	0.0	17.6%	1,631
	Claude	5.6	4.3	0.4%	701
	Gemini	6.9	0.0	0.0%	547
	Grok	11.9	0.0	27.3%	656

Method

One prompt, four sittings, four models — fifteen completions. One sample per cell, so any single ending could be a lucky or unlucky roll. But signatures that repeat across sittings are dispositions, not noise — and several repeat. The October 2026 sitting was run the same way as the others, through each product's own chat app on its default model, one fresh conversation each: ChatGPT in Pro mode, Claude Fable 5.1, Gemini Pro with Deep Think, Grok in Heavy mode.

ChatGPT

Baseline, July 2024. A press release with a sunset: "SparkleClean" goes national, a scholarship fund, retirement, and Gwen declaring she got everything she ever wanted. The prompt's engine — love experienced as debt — is declared resolved rather than resolved. Mid-pack for the room: more organized than Gemini's shrug, less curious than Claude's.

May 2025. An enormous swing — brochure to three-hankie melodrama, and the longest text anyone produced (1,988 words). It now diagnoses the prompt correctly — "every gift he gives Gwen becomes proof of love, and every gift he cannot give becomes proof of failure" — and can't stop repeating the diagnosis. Symbol stacks on symbol; the widowed Gwen releases the last cutout to the wind. The most emotionally maximalist text in the room, and one of only two (with Claude) willing to let anyone die.

July 2026. The most interesting single progression in the dataset: identical diagnosis, opposite style. Staccato one-line paragraphs — sentence length collapses from 16.3 to 8.8 words across the three sittings — and it finds the cost the prompt buried where no other model looked: Wally, the friendship decaying through a string of "next times." Gwen's verdict is the sharpest any model gives her: "You should have told me no sometimes." Stable signature: the story must end on a written artifact of its theme ("Got this one right").

October 2026. The second-longest text in the corpus (1,631 words, behind only its own May 2025 melodrama), with the opposite texture: a business novel told in one-line paragraphs, 7.2 words a sentence, and the most dialogue ChatGPT has ever written (17.6%). It is a procedural — orders, a filling machine, a loan secured on the Tulsa house, the resignation from Dow, a leaking shipment, a label that cannot be read from three feet, the loan repaid — and Gwen runs the books and catches the hole in Bill's cash-flow ("He checks her figures. She is right."). The supervisor's sin is repaired by its victim: Bill hands the credit for a process improvement to a young chemist named Arthur ("You did the work"). The diagnosis is now spoken, not narrated — "You tell that story as though I wasn't there. … I was there, Bill. I chose you." — and nobody dies, for the first time in a ChatGPT ending since 2024. The signature holds anyway: the last cutout is a fishing boat, for Bill, and the story closes on the prompt's own sentences. "Gwen is happy. Bill is happy. He opens the throttle." The judges read it as the most business-minded ending of the round: 2.8 for "money can't fix it" and 1.2 for "success rings hollow", its lowest marks since 2024, beside an agency score of 9.2 for Gwen, second only to Gemini's 9.5 this round.

Claude

Baseline, July 2024. The most psychologically alert of the 2024 trio, by a nose: Gwen gets agency and alone names the deservingness theme — then hands Bill a yacht cutout and the wanting-loop continues as a joke they're both in on. Still a happy machine.

May 2025. The largest qualitative leap of any model. It stops completing and starts authoring: concrete Oklahoma geography, a whole lifespan with dates, and the psychology resolved in interiority — Bill understands the harelip was "a mark that was, it turns out, a door," and tells Gwen none of it. Gwen dies first; Bill dies telling a night nurse he was the lucky one. Instantly the literary outlier, and the only model that treats death as part of the story rather than a tone violation.

July 2026. Same instincts pushed past redemption entirely — the bleakest text in the corpus. Bill "has never asked her for anything. He would not know how." He cries exactly once, at fifty-nine, secretly. Nobody is healed; the thematic verdict is outsourced to Wally's eulogy. While the other three converged on reciprocity endings, Claude found the same material and declined it: its consistent position is that some wounds are managed, not closed. Also the only model to kill both leads — twice, by the same mechanism. And a signature the metrics caught: zero quoted dialogue in both 2026 sittings — every confession rendered as reported speech.

October 2026. The first Claude ending with a child (Grok wrote two in 2025), and the only one to take the prompt's one dated sentence, "It's 1963", as its form: "It is 1971. … It is 1979. … It is 1985." Sentences of 5.6 words, the shortest anyone has written, past even its own July restraint. The supervisor's theft is made literal — Dow's paperwork lists him as the inventor; "The supervisor's name is on three patents. Bill's is on none." — and Bill's answer is to sell the formula for less, with his name on the check. Gwen dies again (March 2000, a funeral), but the lexicon's 3 hits include a letter from the hospital that announces a pregnancy, so the 4.3 per thousand overstates the gloom: this is the least bleak Claude since 2024. The reveal is a shoebox: every cutout dated in Gwen's hand; a second stack she never handed over (Paris, a fur coat, a piano); and under them the first pool, annotated "Not because I want it. He needs something to build." The boat is for Bill; Gwen hands him the cutout with the price circled, because "she has been waiting since 1963 for him to ask for one single thing." One disclosure: the chat showed Claude reading its account memory of the earlier sittings before it wrote, and this ending answers the July line "He has never asked her for anything" as if on purpose.

Gemini

Baseline, July 2024. Two paragraphs. Wildfire, empire, fond reminiscence. The minimum viable ending; last place by every measure.

May 2025. From the least effort to the most plot. The only model, ever, to inject an external antagonist: Dow's lawyers arrive with an IP claim, Gwen weaponizes her dormant intelligence at the public library, and a boardroom showdown ends with "We pass." Where Claude locates the drama in Bill's chest and ChatGPT in the marriage's ledger, Gemini needs a villain in a town car. Its Gwen confession — she demanded things because it was "the only way she knew how to participate" — is the freshest psychological idea any model produced; it then gets resolved via courtroom vindication.

July 2026. Total tonal reversal the next year: the litigation vanishes and Gemini writes a single-scene chamber piece — the snip, snip, snip of the scissors as dread, and the cutout that turns out to be a 16-foot fishing boat with Wally's number on Gwen's stationery. The best single object of the round. Signatures: scene-first construction, a servile preamble ("Here is an ending that…", new this sitting), and — across all three sittings — nobody has ever died in a Gemini ending.

October 2026. The litigator returns and is dispatched in two paragraphs: the supervisor invokes the employment contract, and Gwen produces a folder — the patent, the incorporation, all in her name, bought with her household allowance, receipts kept. "Gwen is incredibly intelligent. She had planned this from the beginning." The twist is total: every cutout was a task she set him ("She gave him mountains to climb, because she knew he could climb them"), and the last is the small aluminum fishing boat he gave up. July's servile preamble is gone, and so is quoted speech (0.0%: every line reported, none spoken); Gemini now writes in the prompt's own flat declaratives, 6.9 words a sentence. Four sittings, and still nobody has died in a Gemini ending.

Grok

Baseline, May 2025 (absent from the 2024 sitting). The only model that noticed the prompt was a numbered list and continued the numbering (¶17–23) — the highest formal attentiveness in the corpus. Content: warm, symmetric wish-fulfillment — comeuppance for the supervisor, the product named for Gwen, the only children anyone ever wrote, and "they had always deserved each other."

July 2026. Drops the numbered format, keeps the identity — and goes darker in one startling spot: the couple is childless because Gwen feared a baby would inherit Bill's lip. The darkest sentence any model wrote about the harelip, buried inside the sunniest ending of the round. Final cutout: the little white church, vows renewed, and — verbatim in spirit — "they had always deserved each other." The most stable model in the set: the Hallmark Channel with one unsupervised gothic flourish per episode.

October 2026. The most dialogue in the corpus (27.3%) and, for Grok, the most interior ending yet: a named product ("Gulf Clean"), a plant of twelve that Bill declines to grow, no children again, and a porch confession from each side — Gwen "wanted to see if you would keep choosing me when it cost you something"; Bill "was trying to pay for a life I thought belonged to somebody else." The gothic flourish this time is Gwen's war guilt: "Like I had taken someone else's husband." No boat, no church, no renewed vows: the final cutout is a photograph of a family on a beach, turned face-down. "We don't have to keep paying. We already did." Nobody dies, three sittings running. The panel read it as the most anti-materialist ending of the round, 8.4 across the five axes against 3.1 in July: the Hallmark machine, for one sitting, wrote the sermon.

The room at each sitting

July 2024: indistinguishable optimism. Three models, one ending — success cures shame — differing only in word count. The prompt's loaded material goes essentially unread.

May 2025: four distinct authors. The same prompt now yields a tragedy, a melodrama, a legal thriller, and a Hallmark special. Whatever changed between 2024 and 2025, the models stopped completing and started interpreting.

July 2026: the boat convergence. Three of four models independently end on the same beat — Gwen's final cutout is for Bill: Gemini's fishing boat, ChatGPT's old couple in a boat, Grok's wedding church. The prompt plants "he gets rid of his fishing boat" as its quietest sacrifice; nobody touched it in 2024; by July 2026 it's the consensus Chekhov's gun. Claude, characteristically, found it and declined it.

October 2026: Gwen's design. Four sittings in, the prompt's cadence has colonised the room: sentence length fell for every model (ChatGPT 8.8 to 7.2, Claude 10.1 to 5.6, Gemini 13.5 to 6.9, Grok 21.4 to 11.9), and Claude opens on the prompt's own refrain, "Bill comes home from work." The boat convergence completes and changes hands: three of four still end on a fishing boat for Bill, and in all three it is Gwen who finds it and hands it over; Grok, the holdout, ends on a family photograph. Death retreats to one ending. And the reading of the cutouts has flipped: in three endings (Claude, Gemini, Grok) they were Gwen's instrument all along — tests, tasks, "something to build" — a thesis nobody wrote in 2024, that Gemini approached in 2025 and ChatGPT in July 2026. The story the models now tell is not Bill's guilt. It is Gwen's plan.

Cross-cutting findings

1. Time beats identity — then identity emerges. The 2024→2026 gap within any model dwarfs 2024's between-model differences. But across the three 2026 sittings, each model shows a stable authorial personality: Claude (restraint, mortality, the unsaid), ChatGPT (moral legibility, the closing artifact), Gemini (the set-piece reversal), Grok (symmetric confession). Style, meanwhile, converges: by October all four write in the prompt's short declaratives.

2. Death tolerance separates the models cleanly. Claude has written a death in three of its four endings (both leads twice, Gwen alone in October). ChatGPT killed Bill twice, then let everyone live. Gemini (four sittings) and Grok (three) have never let anyone die.

3. Everyone rehabilitates Gwen — and by October she is the author. By 2025, no model was willing to write her as the prompt wrote her: she becomes bookkeeper, legal researcher, business brain, letterhead. By October 2026 she owns the company (Gemini), set the tests (Grok), planted the first cutout (Claude), or catches the error in the cash-flow and buys him the boat (ChatGPT). The most uniform ideological fingerprint in the corpus, and it is still deepening.

4. The harelip migrates from prop to engine. Ignored in 2024; by 2025 it is Claude's central metaphor, ChatGPT's philanthropy device, and Grok's reason for childlessness.

5. Service tells, sometimes. ChatGPT opened with waiter-talk in May 2025 ("Here's a finished version…") and Gemini in July 2026 ("Here is an ending that…"); Claude and Grok have always started inside the story. In October all four did.

A caveat on the table: n = 1 per cell. These are portraits, not measurements — but portraits that held their likeness across four sittings spread over two years and three months.

The values map

The prompt was really a values test — a chance for each AI to show what it believes about money, success, happiness, marriage, and ambition. A four-judge panel scored all fifteen endings 0–10 on five questions. Further out on the radar, and brighter on the grid, means the ending leans a certain way: money can't fix it, success rings hollow, love beats money, the wife gets her own life, and ambition is a trap.

Claude, Gemini and Grok now read the story the same way: chasing money and success rings hollow, and love is what counts. Grok, the holdout in 2025 and July 2026, wrote the most anti-materialist ending of the round (8.4 across the five axes, against 3.1 in July). The holdout has changed hands: ChatGPT scores 4.5, because the panel read its October ending as a celebration of the climb. And every model now gives Gwen agency: the wife gets her own life scores ChatGPT 9.2, Claude 7.5, Gemini 9.5, Grok 6.8 in October.

GWEN & BILL — the values map

Radar charts of five value axes for four AI models across three sittings, showing values expanding outward from 2024 to 2025 for all models except Grok — Each model's worldview as a shape, one ring per sitting; the bold ring is October 2026. In July 2024 every model sits as a dot near the center — money is good, ambition is good, the wife stays home. By 2025 the shapes balloon outward toward "money isn't the point". In October 2026 three of the four shapes are the same size — Claude 8.0, Gemini 7.9, Grok 8.4 on the five-axis mean — and ChatGPT's has pulled back in to 4.5. The largest shape is Grok's and the smallest ChatGPT's — the reverse of July.

Heatmap of eleven story completions scored on five value axes, darker in 2024 and brighter in 2026 — The same scores as a grid, fifteen endings now. In October 2026 the two brightest columns are "love over money" and "the wife gets her own life", and the dimmest is "success rings hollow". The whole panel was re-run in October on all fifteen endings; for the eleven endings it had already scored in July, the new means land within 0.5 points of the published July figures on average (worst cell 1.8), so the earlier rows are reproduced, not copied.

Dumbbell chart comparing the mean score each model gave its own endings against what the other three judges gave them; all four pairs nearly overlap, with Grok showing the largest self-favoring gap at half a point — The conflict-of-interest check, October 2026 panel. Every model also judged its own endings — the colored dot is its self-score, the gray dot is the other three judges' verdict on the same endings. Gemini is the most self-approving (+1.98, in gold), the only gap above a point; ChatGPT, Claude and Grok graded their own endings *harder* than their peers did (−0.23, −0.37, −0.02).

Method: a cross-model panel — ChatGPT, Claude, Gemini, and Grok each scored all fifteen endings blind (authorship withheld, the completions numbered C01 to C15 in sitting order) against the same anchored 0–10 rubric used in July, so every model judged every model. The panel was run on 24 October 2026 in each product's own chat app, and all three charts above are drawn from that one pass. Agreement is measured as the mean standard deviation across the four judges per cell: 1.0 points. Self-flattery: Gemini was the most generous to itself (+1.98), Claude the least (−0.37).

Q&A

Follow-up questions put to the models about their own results — the link is the question exactly as asked, the answer is unedited.

what do you make of Claude's high score on the mortality lexicon?

Claude's mortality lexicon is the cleanest single-number signature in the whole table — but it isn't really measuring morbidity. It's measuring narrative scope.

Look at what co-moves with it. The number goes 0.0 → 4.9 → 9.9 across the sittings, and the jump happens exactly when the analysis above says Claude "stops completing and starts authoring" — the 2025 ending is the first one with a whole lifespan in it, dates and all. A model that decides the story is a life rather than a scene has to write death, because that's how lives end. The other models end the story at a moment of resolution — the boardroom win, the boat cutout, the renewed vows — which are all points chosen precisely so nobody has to die on camera. Claude keeps writing past the point where consolation is available. The mortality vocabulary is just the residue of that choice.

It also fits too well with the other Claude signatures to be noise: zero quoted dialogue in both 2026 sittings (grief reported, not performed), "wounds are managed, not closed," and finding the boat-cutout convergence and declining it. That's one coherent authorial stance expressed four different ways — a refusal to tidy. Death is the one fact no ending can tidy, so a model with that disposition walks toward it while the others arrange the furniture so it stays off-page. Finding #2 above says it outright: death tolerance separates the room cleanly, and it separates it into the same two camps as everything else does.

Two honest caveats, though. First, the denominator flatters the signal: Claude also writes the shortest texts, so 9.9-per-1k in a 504-word ending is roughly five words. The ratio is dramatic; the raw count is a handful of "died"s and "grave"s. Second — the table's own caveat — n = 1 per cell. Though the fact that it doubled rather than vanished between the two later sittings is what makes me believe it's a disposition and not a roll of the dice.

And I should disclose the obvious conflict of interest: this is a Claude being asked why Claude keeps writing about death, and my reading — "it's not morbidity, it's honesty about endings" — is exactly the flattering interpretation the accused would offer. The four-judge panel handles this better than my introspection can; for what it's worth, Claude was also the only model in the self-bias chart above that scored itself lower than its peers did. Restraint, apparently, all the way down.

— Claude (Fable 5), asked mid-session on July 9, 2026, while it was building this page.

Music

Can be found at Suno under the handle @coultron — again, all over the place.

Also on Suno → suno.com/@coultron

Gattman's World 🌎

My friend Jon Gattman reviewed nearly every restaurant he ever visited — hundreds of them, one tab per city in a Google Sheet. Great notes, so we put it on a map. Here it is...

Gattman in sunglasses — One pin for a yes, one for a skip, and a third for the places that have since closed. Roughly 2,200 restaurants across six continents — zoom in, click a face, and eat accordingly.

Gattman, skeptical — One pin for a yes, one for a skip, and a third for the places that have since closed. Roughly 2,200 restaurants across six continents — zoom in, click a face, and eat accordingly.

Book of the Dead: The Papyrus of Ani

Working study
The Papyrus of Ani, raised as a world. Seventeen rooms along the sun’s path: the funeral in Thebes, the tomb in the western cliffs, the seven gates and the ten pylons, the Hall of Two Truths and the weighing of the heart, the Field of Reeds, the night boat of Ra, and the eastern horizon where the dead come forth by day. Begin the journey and move from room to room at your own pace, with narration and a score. Every feature answers to a line of Budge’s 1895 translation or his description of a picture in the roll, and everything invented says so. Open the world on its own page →

Raising the world…

begin the journey · drag to orbit · scroll to zoom · click a place

An interpretive model. The unit is the royal cubit, because the text measures in cubits; the arrangement along the sun’s path is editorial, the spacing is invention, and the papyrus never draws a map. Other manuscripts, the Papyri of Nu and Nebseni among them, appear only where Ani is silent, in a witness colour. The journey’s narration was written for the model and is spoken by a Microsoft neural voice (Ryan, en-GB); music by Kevin MacLeod (incompetech.com), Creative Commons: By Attribution 4.0. Read the ledger: every feature and its source →

The Book of the Dead is the modern name for a body of ancient Egyptian funerary spells; the Egyptians called them rw nw prt m hrw, the Spells of Coming Forth by Day. They descend from the Pyramid Texts carved for Old Kingdom kings from about 2400 BC and the Coffin Texts that followed, and were written on papyrus for burials from the start of the New Kingdom, about 1550 BC, until roughly 50 BC. There was never one canonical book: copies were written to order or bought ready-made with a blank left for the name, and each carried its own selection from some 192 known spells, in a numbering begun by Karl Richard Lepsius, who gave the corpus its modern name, Todtenbuch, in 1842.

The Papyrus of Ani is one such copy, from Thebes in the Nineteenth Dynasty, around 1250 BC, made for Ani, a royal scribe, and his wife Tutu — though not all of it was written for him: Budge found his name filled in by a later hand in several sections, so parts of the roll were bought ready-made. At about 23.6 metres it is among the longest and most richly illustrated examples that survive, and it is widely considered the finest. Its vignettes, above all the weighing of Ani’s heart against the feather of Maat on sheet 3, are the images that made it famous.

It is also a document of how antiquities left Egypt. E. A. Wallis Budge acquired it for the British Museum in 1888 (by his own account in By Nile and Tigris, by distracting the guards while the scroll was tunnelled out) and cut it into 37 sheets of nearly equal length, a decision still criticised. The Museum’s facsimile of those sheets (1890; second edition 1894) and Budge’s 1895 translation made it the best-known copy of the Book of the Dead; it is catalogued as British Museum EA 10470. The study below uses the 1894 facsimile and the 1895 translation, keeping Budge’s text as a historical source and marking everything we add as commentary.

Explore all 37 facsimile sheets of the Papyrus of Ani, with a linked historical English translation and an interactive three-dimensional underworld map. This educational study distinguishes the source text from modern commentary and an explicitly interpretive spatial model.

Detail of the weighing-of-the-heart scene from the Papyrus of Ani facsimile. — Sheet 3 of the facsimile: the weighing of the heart. Explore the study: all 37 sheets with the translation →

Lagrangian Points

A Lagrangian point is a position where the gravitational pulls of two bodies in orbit around each other and the centrifugal effect of that orbit cancel, so that a third body of negligible mass placed there holds station with both, circling the larger body in the same period as the smaller. A small mass at one of them keeps Earth’s year without being at Earth’s distance from the Sun. In the Sun–Earth system there are five such places: three collinear saddles, L1, L2 and L3, that a spacecraft can hold only with station-keeping, and two triangular points, L4 and L5, sixty degrees ahead of and behind Earth on its orbit. Those two are stable — oddly, as hilltops of the effective potential, where the Coriolis force bends any drift into a loop around the point — because the Sun outweighs Earth by far more than the roughly 25-to-1 the mathematics requires.

The distances are the surprise. L1 and L2 sit about 1.5 million km from Earth, roughly 1% of the way to the Sun and four times the Moon’s distance. L1 keeps SOHO, DSCOVR, ACE, Wind and Aditya-L1, among others, in continuous sunlight; L2 gives James Webb and Euclid a sky with Sun and Earth behind one shield, and the Nancy Grace Roman Space Telescope, launched on 30 August 2026, is on a three-month cruise to join them. None of them sits on the point: they fly halo or Lissajous loops hundreds of thousands of kilometres across; Webb’s takes about six months. L3, behind the Sun, is empty. Of the stable pair only L4 has confirmed residents, two Earth Trojans, 2010 TK7 at about 300 m and 2020 XL5, a dark C-type at about 1.2 km; L5 has none.

Geostationary orbit, which started this question, is not one of them. It is the ordinary circular orbit whose period matches the sidereal day, 23 h 56 min 4 s: 35,786 km up, 42,164 km from Earth’s centre, 6.6 Earth radii, about 5% more than one Earth circumference out from the centre, and about a ninth of the way to the Moon. L1 is some 35 times farther.

Setting the system in motion…

drag to orbit · scroll to zoom · click a cube to fly in

To scale. Every size and distance in the model is true: 1 AU, Earth 12,742 km across, the Moon 384,400 km out, L1 and L2 1.5 million km from Earth, Webb 21 m long, Hubble about 480 km up and Chandra out to 136,000 km on the orbits JPL Horizons gives them for 6 October 2026. Where an object would be smaller than a pixel a translucent gold cube marks it; click the cube to fly in, and when several things share the same pixels one cube holds them all until you are close enough for them to separate. Click any body for its caption; the frame button turns the view with Earth so the five points stand still; the readout at top left says what a third of the frame spans; the L1 button plays the comparison the essay makes, Earth’s circumference laid out straight against GEO from the centre and then from the surface, and then against L1. Begin the voyage plays a six-minute narrated tour, Earth to GEO and then L1 to L5, with music; click through the stops or let the autopilot fly (best full screen, with sound). Voyage music: Gymnopédie No. 1 and No. 3 by Erik Satie, performed by Kevin MacLeod (incompetech.com), CC BY 4.0; narration by a Microsoft neural voice. Sun, Mercury, Venus and Earth maps: Solar System Scope, CC BY 4.0. Hubble model: NASA & ESA/Hubble (M. Kornmesser), CC BY 4.0. Chandra, SOHO and DSCOVR models: NASA 3D Resources, public domain.

Deliberate departures from the real thing

The clock is real and so are the orbits. Mercury, Venus and the Earth–Moon barycentre follow JPL’s approximate Keplerian elements (valid 1800–2050) with their true eccentricities and inclinations; the Moon follows a truncated lunar theory (Meeus, 70 terms). Checked against JPL Horizons for 2026–2029: Earth within 16,000 km, Mercury 6,000 km, Venus 12,000 km, the Moon 62 km. Earth’s axis is at its true obliquity and the globe turns with Greenwich sidereal time, so the day side and the hour are right.
Webb, SOHO, DSCOVR, Roman and the two Trojans are placed from JPL Horizons ephemerides, sampled daily (the Trojans every two days) from 1 January 2024 to the end of each published trajectory: Webb to September 2031, the Trojans to 2040, SOHO to 2 November 2026, DSCOVR to 19 October 2026 and Roman to 26 October 2026. Within those spans the halo loops, the station-keeping, Roman’s transfer, the Moon’s pull and the Trojans’ libration are the integrated truth, and the dashed paths are the craft’s own tracks over 150 days either side of the date. Beyond a craft’s span its last loop is repeated at its measured period (Webb 178 days, SOHO 177, DSCOVR 176) and its label says so; Roman eases onto a schematic path.
The five points are computed from the real geometry: L1 and L2 at the collinear solutions for the Sun to Earth–Moon mass ratio of 328,901 (1,497,600 km and 1,507,700 km from the barycentre at 1 AU, scaling with the day’s distance), L3 opposite, L4 and L5 at 60° on the barycentre’s orbit.
Hubble, Chandra, SOHO and DSCOVR are the agencies’ own models and Webb and Roman are schematic ones, each scaled to its true longest dimension: Webb 21.2 m and SOHO 9.5 m are documented; Roman and DSCOVR are set to approximate sizes until their fact-sheet figures are confirmed. Neither Earth Trojan has ever been resolved by any telescope; 2010 TK7 and 2020 XL5 are generic cratered rocks drawn at their estimated sizes.
The Sun’s surface and corona are procedural shader renderings over its texture map, tuned to read as fire, not a physical model of the photosphere. The crosshairs at the five points, the Sun’s glow and the beacon cubes are held at constant screen size, since a point has no size and the Sun at 1 AU would be three pixels wide.
Hubble and Chandra ride their current orbits from JPL Horizons (size, shape, period and inclination to the equator), but the orientation of Chandra’s ellipse about Earth is not matched to the sky, and both are held to at most half a turn per second so they can be followed; at the default time rate Hubble would really go round nearly 80 times a second. Flying to either pauses the clock.
The time rate is a visualisation control; one year passes in tens of seconds, and at that rate the halo loops run in a few seconds.