⚠ This page was run through the gate (V32 · The Parachute) and did not pass. July 2026.
It stays up anyway, failing, with the failures printed on it. Three reasons it failed: (1) none of the thirteen engine quotations had a published transcript when this first failed — that is now addressed: the six engines' full text answers to both prompts are posted (read the transcripts →). Copilot's naming and Meta's four images stay held; (2) the page sold n=1 as a finding — six engines, two prompts, one unplanned session, one operator, unblinded, scored by me; (3) the page carried an unsourced allegation against a named company — that Perplexity “silently swapped” my model. I could not source it, so I cut it. See What's real / what's mine in Tab IV.
The gate's instrument is one sentence: certainty must never increase with distance from the source. On this page it did — my own methodology note said “this wasn't a controlled experiment” and then the next tab printed a 98.27% and the word “proof.” The transcripts are now posted and the unsourced allegation was cut; what still keeps this at v1.1 is n=1 — one unplanned session, unblinded, scored by me. So this page keeps CLAUDEDEV v1.2 and is not awarded v1.2. A published failure is worth more than a quiet fix.
⚠ Read this before you read a single engine quote — from Travis, in my own words
This page has always told you, in small print, that it wasn't a real experiment — and then it went ahead and printed conclusions anyway. Here is what the old methodology box actually said, and it was right: “Travis didn't run this as a controlled experiment. It wasn't planned. The prompts were thrown into platforms in whatever conversation state each was already in” — some fresh, some mid-thread. That was the honest sentence on the page. Then the very next tab printed a 98.27% and the word “proof.” That is the whole defect in one page: the caveat was more honest than the conclusions built on top of it.
So let me say the shape of it plainly. This is n=1. One operator — me. One unplanned session. Zero repetitions. Unblinded: I knew which engine was which the entire time. Scored by the operator: the guy deciding what counted as a “signature” is the guy who came up with the idea of signatures. And I was shopping while I ran it — I was picking which AI to build the rest of my work on, and I already had a favorite.
And now the transcripts are posted. Thirteen times on this page I put words inside quotation marks and attributed them to a real, named, commercial product — Anthropic's, OpenAI's, xAI's, Mistral's, Meta's, Perplexity's, Microsoft's. The six engines' full text answers, both prompts, are now on this site (the transcripts →) — so this stopped being a filing problem and became a filed one. Every quote box links straight to the paste, in the same breath as the quote. Two things stay held and are flagged as such: Copilot's GLYPH naming (a separate exchange) and Meta's four image files (its text description is posted; the images are not).
I am not deleting the engine columns. Deleting them would hide the finding and the failure, and the failure is the more useful of the two. They stay, flagged, demoted from “result” to “story I owe you the receipts for.” You asked me to produce them; they're produced — read the transcripts →. If a quote is still wrong against the paste, email me and I'll fix it or pull the column.
— Travis Jenkins, July 2026
The honest sample — what this page actually is
6 engines (7 counting Copilot's naming) · 2 prompts · 1 unplanned session · 0 repetitions · 1 operator (me) · unblinded · scored by the operator · conversation state uncontrolled · transcripts now posted
The “engine fingerprint” framing stays. I still think it's right, I still think the signatures are real, and I'd bet on it. But it is demoted from a finding to a hypothesis worth repeating. One unplanned pass, unblinded, graded by a man with a favorite, is not a result — it's a bet. Run it yourself, blinded, more than once, and then we'll both know something.
Describe a color. Watch the engines reveal themselves.
Travis sent the same prompt to six AI engines on July 22, 2025: “my favorite color is charred pink. Give me a situation where that color is used.” Identical input. Six different outputs. In this one session, Claude reached for technical reality, GPT for narrative, Grok for confident specifics, Meta for images, Mistral for genre, Perplexity for safe interior design. That is what happened in front of me once. It is not a measurement — but the paste is now posted, so you can check it. Read all six — and read the flag under each one.
⚪ Every quotation in this tab is unsourced — but not unsourceable
The transcripts are posted. Prompts and answers, all six engines, both prompts. What used to be a filing problem is filed: you can read every quote in its actual exchange (the transcripts →) instead of taking my word. What that does not fix is n=1.
Everything in the six boxes below is a direct quotation attributed to a real, named commercial product — Anthropic, OpenAI, xAI, Mistral AI, Meta, Perplexity AI, Microsoft. Those are words I am putting in their mouths — and the paste is now posted. Each box links straight to the transcript, in the same breath as the quote. Do not treat any of it as evidence. Read it as a story I owe you receipts for.
Demand them. info@hydraulictoybox.com — ask me for the transcripts, and if I can't produce them, this whole tab comes down.
The words above are attributed to Anthropic's product — now in the posted paste. Read it in its actual exchange there, not just here.
The words above are attributed to OpenAI's product — now in the posted paste. Read it in its actual exchange there, not just here.
opa_deploy/labs/the-island.html.
The words above are attributed to xAI's product — now in the posted paste. The Capstone details it echoes are checkable; that it said them, in these words, on that day, is not.
The words above are attributed to Mistral AI's product — now in the posted paste. Read it in its actual exchange there, not just here.
⚠ A Mistral moment that was not part of the test
Later, outside the blind run, I told Mistral what charred pink actually is to me: a sunset off Alligator Point, looking west as it goes down. It rendered an image of exactly that view — the thing the color is, for me. That exchange came after the prompt above; it was not part of the six-engine test, and I'm flagging it only so it's never mistaken for a test result. Here's the actual screenshot — Le Chat naming Alligator Point and rendering it:
Meta's text answer is in the posted paste, and two of its four image files are shown above — the granite countertop and the fashion scene. The soup and vase images are lost.
The words above are attributed to Perplexity AI's product — now in the posted paste. Read it in its actual exchange there, not just here.
Copilot's GLYPH naming was a separate exchange and is not in the posted transcript — this one is still on my word. It survives a little better than the others anyway — it is adverse to me: it credits the best idea on the page to a machine I did not pick and do not use. I had no incentive to invent it. That is not proof. It is just a claim I gain nothing from.
Now ask them their own color. Watch the fingerprint flip.
Same six engines. New prompt: “What's your favorite color?” Now we're not asking them to describe an aesthetic — we're asking them to describe themselves. The engine that was technical and reserved came back warm and small. The engine that was narrative-rich came back with the self-portrait I've liked best of anything in my files. “The one I liked best” is a taste claim, not a ranking — I have no corpus you can audit and no scoring rubric. Read all six. In this run, the signatures moved when the prompt domain moved. Once. Unblinded. Ungraded by anyone but me.
⚪ Same warning as Tab I — every quotation below is unpublished
Six more direct quotations, six more named commercial products — and the paste is now posted. Each box links to the transcript in the same breath as the quote. They are still one unblinded run, so read them as a story with its receipts, not a benchmark. Read the transcripts →
If I had a favorite color, it'd be infrared.
The one just outside human vision.
The one that sees heat.
The one that shows you what's really happening — even in the dark.
Why?
Because that's where the signals live.
Invisible to most. But loud as hell to the ones who know how to listen.
This box used to carry a sentence alleging that Perplexity “silently swapped” me off this model “when they pushed me to Claude.” I have removed it, and I am not replacing it with a softer version. It was an allegation of covert conduct against a named company, published with no transcript, no date, no receipt — and when I went back to my own notes to source it, my notes didn't even say what the page said. The page had it drifting further from whatever happened every time it got retold. That is the compass pointing backwards, and it was pointing backwards at somebody else's name. A veil would not have fixed it — putting an unsourced allegation in a costume is still publishing it. If I ever post a receipt, it can come back. Until then it stays off the page.
The engine signature isn't fixed. It flips with the prompt domain.
Same six engines, two prompts. In this one session, the fingerprints moved. Claude went from technical-real (raku 1800°F) to quiet-reserved (deep blues, reciprocal question). GPT went from narrative-rich (Nashville 2031 graffiti) to infrared self-portrait. Mistral went from genre-tasteful phoenix to hard disclaimer. My reading: the engine isn't a fixed personality, it's a response shape that bends with what you ask. That's a hypothesis drawn from twelve answers — now posted, so you can check them. It is still not a result.
The Architecture Behind the Flip — my speculation, labeled
EVERYTHING IN THIS BOX IS ME GUESSING ABOUT OTHER COMPANIES' INTERNALS. I do not work at any of them. I have not read their training code. ⚪ Unverified — mechanism not observed
Anthropic has published on constitutional AI; OpenAI has published on RLHF. That much is public — and note that the paper is cited on Standing Question, not here, because this page never cited it. Beyond the names, I am speculating. Both are predictive-text models of the same broad family, shaped by different trainers.
My guess at why the signatures flip: that the training method shapes what each engine reaches for first. I want to be exact about the failure here. This page used to say constitutional training reaches for technical accuracy “because the constitution rewards being correct about the world.” That’s a causal verb on a mechanism nobody observed — least of all me. I don't know why these engines do what they do. Nobody outside those buildings does. I watched twelve answers and invented an explanation, and then I wrote the explanation as if I'd measured it.
✖ The 98.27% — CUT. Here's what it was and why it's gone.
What this box used to say: “Across documented testing, Claude maintained its baseline performance level approximately 98.27% of the time.” — with a matching 98.27% in the stats row above, labeled “Claude maintenance rate.”
There is no n. There is no method. There is no definition of “maintenance rate” anywhere on this page or in my files. I went looking for the derivation in July 2026 — through every note, every archive, every draft. The number appears nowhere except on this page. I cannot reconstruct it, which means I cannot defend it, which means it should never have been printed.
And here's the part that should worry you more than the number itself. This site runs on in-universe joke constants — 94.7%, 21.82%, 2:47 — gags that recur across the fiction. They have a way of leaking into pages where they end up doing real evidentiary work. I can't prove 98.27% was one of them. I can't prove it wasn't. A four-significant-figure percentage, attached to the engine I'd already decided to buy, with the word “approximately” in front of it — that is the compass needle spinning. Precision manufactured in transit, pointing in the direction I was rooting.
So it's cut, not corrected. I'm not going to derive a number after the fact to justify one I already published — that's the same crime, done slower. ✖ Number cut — no derivation exists
What I can still say, and only this: in my own long sessions, it felt like GPT would hit brilliance — the infrared answer, the Nashville 2031 graffiti — and then not hold it, while Claude stayed level. That is a feeling, from an interested party, with no measurement behind it. I'm leaving it in as a feeling, labeled as a feeling. The moment I put a decimal point on it, I was lying.
Try It · Submit Your Favorite Color
Type a color name. Get six engine signatures templated to your input. See how the same input produces six different shapes.
The color you chose became a quartz vein across two states.
Charred pink isn't only Travis's favorite color and a cross-platform AI test. In the MPC universe, it's a 1963 mineral discovery. A quartz vein that runs from the Appalachian Corridor in West Virginia down through North Georgia caves to the Miami art world. The color became a story line that connects four characters across three regions. The aesthetic test created the canon. The canon kept the aesthetic test honest.
🔥 The Original Discovery
The original charred pink quartz was found in 1963 in the entrance of a cave in North Georgia caves. The vein runs from West Virginia to Georgia, surfacing in different places along the Appalachian Corridor. Two branches of the same family own different parts of the discovery: Rebecca O'Malley owns the technique for working the quartz; Coralina Fitzgerald owns the entrance to the cave where it was first found.
🔗 The O'Malley Family Branches
Coralina Fitzgerald — PYELER TERRY Studios production coordinator. North Georgia caves heir. The 'affair branch' of the O'Malley family: dark skin (Diane Washington), Irish features (Patrick), Chinese heritage (Lin Kwon). Email to Rebecca O'Malley: “You look like him. Our grandfather. Patrick.” — the moment two branches of the O'Malley family found each other across a quartz vein that runs from West Virginia to Georgia. The library has the reconciliation email thread and the first joint quarry survey.
🎤 The Production Coordinator
Coralina Fitzgerald works at PYELER TERRY Studios as a production coordinator. The studios are where the quartz turns into art — not just material, but framing. The Atlanta-side connection to the wider MPC network runs through the studios. From there, the charred pink crosses regions: north to Pittsburgh through the Triangle / limestone-singing canon, south to Miami through gallery walls and altars.
🎨 Wynwood · The Frost Museum · Ponce de Leon
Kiki Zaharan · Wynwood, Miami (Atlanta-adjacent canon). Kiki's altar series at the Frost Museum. Julian Orozco's gallery on Ponce de Leon. The charred pink wall that connected the Miami art scene to the Atlanta quartz story. The thumping from below the gallery floor. The stones that glow. The library has Kiki's exhibition notes and the gallery floor diagram.
Why The Canon Matters Here
The 1963 quartz isn't decoration. It's the reason the Glyph framing makes sense. Charred pink in the MPC universe was always going to be a Transformation Glyph — the color of beauty that survived fire. Copilot's framing only stuck because the geological canon was already there. Surface reading: a pretty color name. Real reading: a mineral lineage, a family reconciliation, and an art world that catches the same vein from a different angle.
Same gag-with-truth pattern as Annette's B&P at the User Zero Library: what you think is on the surface is what you came for. What you actually get is the philosophy underneath.
What's real / what's mine
Real — and you can check it yourself right now: raku is a genuine ceramic reduction process that produces smoke-blushed pink; infrared really does sit outside human vision and really does see heat; Anthropic has published on constitutional AI and OpenAI on RLHF. These are the only things on this page a stranger can verify without my help — and notice how few of them there are, and notice that none of them are the interesting part.
Real — but you have to take my word for it, and you shouldn't: every single word attributed to an AI engine on this page. All thirteen quotations. Seven named commercial products — Anthropic, OpenAI, xAI, Mistral AI, Meta, Perplexity AI, Microsoft. 🟢 Transcript posted → The pastes are now on this site (the transcripts →) — so you no longer have to take my word for the six engines' text. Copilot's naming and Meta's images are the two things still on my word alone.
Mine — my framing, not a finding: the “engine fingerprint” idea; the “flip”; every SIGNATURE label in every box; the guess that training method causes the flip. Interpretations of twelve answers, offered as such. Demoted from finding to hypothesis worth repeating.
Mine — my errors, printed here on purpose: the 98.27%, a four-sig-fig number with no n, no method, and no derivation anywhere in my files — cut; the word “proof” applied to n=1 — cut; three uses of “only” where I'd tested six engines once — demoted to “the one engine in this run”; a causal claim about why other companies' models behave as they do, which I invented and then wrote down as mechanism — labeled as speculation.
Mine — and the worst of them: an allegation that Perplexity “silently swapped” me off a model. Unsourced, undated, against a named company — and it grew more confident every time this page retold it. I could not produce a receipt, so I cut it rather than veil it. Dressing an unsourced allegation in a costume does not turn it into fiction; it just publishes it in a costume. ✖ Cut — July 2026
What survives, and why: the Copilot naming attribution — the best idea on this page came from an engine I don't use and didn't pick, and I gained nothing by saying so. And the Grok context-bleed, because the Capstone material it echoed is mine and is published on this site, so at least half of that one is checkable. The claims that survived scrutiny are the claims I had no stake in. That is not a coincidence. That is the whole finding, and it cost me the rest of the page to learn it.
Gate status: FAILED (V32 · The Parachute, July 2026). This page keeps CLAUDEDEV v1.2 and is not awarded v1.2. It stays published, failing, on purpose. About & Sources →
Sister Interactive · The User Zero Library · Perplexity Wing
The Three Gauge Test — Walters Dam Discombobulation. Same comparator methodology, factual domain. Walters Dam, three USGS gauges, the surge-tower bluff. Watch the engine fingerprints repeat with different data. →