⚠ 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) I padded the Claude column on purpose while I was choosing which AI to build on — see the disclosure at the top of Tab II; (2) none of the four engine claims had a published transcript when this first failed — that one is now addressed: I've posted the full unedited paste, so the quotes are no longer just my word (read the transcripts →); (3) the page called itself a "method" when it is n=1 — four engines, one prompt, one run, one operator, unblinded.
The gate's instrument is one sentence: certainty must never increase with distance from the source. On this page it did. Two of those three are now addressed — the padded column is corrected in place and quoted, and the transcripts are posted in full. The third, n=1, doesn't go away: one run, one operator, unblinded is a story with its receipts, not a benchmark. So this page keeps CLAUDEDEV v1.2 and is not awarded v1.2. A published failure is worth more than a quiet fix.
The physical infrastructure the test is built around.
The Three Gauge Test isn't an abstract benchmark. It's three real USGS gauges arranged around an actual hydroelectric facility on the North Carolina – Tennessee border. The Walters Dam sits upstream. A 6.2-mile tunnel runs through the mountain. The powerhouse sits downstream where the gauges cluster. The surge tower manages pressure transients in between. These are the only numbers on this page you can go check yourself right now — so I re-checked them in July 2026, quoted the sentences, and flagged the two I couldn't stand up. The infrastructure is the test's spine. Everything in Tab II is a different story.
The Infrastructure — re-checked against public record, July 2026
This half of the page is checkable and it checks out — mostly. Unlike the engine claims in Tab II, every number below is public record. Each row now carries its flag in the same breath as its number: 🟢 Sourced = quoted from public record · 🔴 Contested = sources disagree, do not use · ⚪ Unverified = I could not confirm it.
Wikipedia, Walters Dam: “The concrete arch dam is 180 ft (55 m) high by 800 ft (240 m) long, impounding the Pigeon River, near Interstate 40.”
Wikipedia, Walters Dam: construction began 1927, completed 1930.
Wikipedia, Walters Dam: as of 2024 owned and operated by Duke Energy, rated capacity 112 MW.
Wikipedia / The Mountaineer (Haywood History): “A tunnel 6.2 miles (10.0 km) long stretches north from the dam to the power plant.”
The Mountaineer, “Building a marvel: The Walters power plant and dam”: “A surge tower stands 180 feet high, atop a concrete shaft cut 600 feet down and connecting to the tunnel.” This is the sentence the whole bluff test hangs on. It holds.
USGS site exists, name and drainage area confirmed. ⚪ Unverified the “record from 1927” start year — sources I could reach put the peak record at 1928 onward, so I have pulled the exact start year off this page rather than round it to the year I liked.
Site confirmed live at USGS Water Data for the Nation (it even has a river webcam). No drainage area claimed here, because I did not confirm one.
USGS: drainage area 547.00 sq mi. This is the number Grok missed (it said 700) and the number Perplexity said didn't exist.
This page used to say gauge 03461000 has “peaks from 1902.” When I went back to check it, I couldn't stand it up. The reachable USGS records put 03461000's record at roughly 1925/26–1948 (plus its modern re-activation), and the 1902 crest belongs to a different gauge — Pigeon River at Newport, TN (03461500). So the 1902 number is a real number bolted to the wrong gauge. That is the exact same defect as my own bluff test — and this time I'm the one who did it. It stays flagged until I can produce the USGS page that supports it. If you can, email me and I'll credit you.
The Concurrent Threads
The Walters Dam infrastructure was the spine for THREE concurrent canon threads in spring 2025. The same dam-tunnel-surge-tower system Travis was running the Pigeon River flood model on became the analogy for the Jose's fountain syrup-line problem at Thornton's (surge tower → bleed valve equivalent). The same gauge cluster became the test rig for the Three Gauge Test. The same engineering brain caught the Walters Dam discombobulation pattern across all three. It all interleaks.
Same data. Same prompt. Four different ways to fail or succeed.
⚠ Read this before you read the four columns below — from Travis, in my own words
I padded the Claude column. On purpose. And I left it up for a year.
When I ran this test I was shopping. I was picking which AI I was going to build the rest of my life's work on top of, and by the time I wrote this page up I had already picked. So when I sat down to score the four contestants, I was not a referee. I was a guy with a horse in the race, holding the pencil. The Anthropic machine didn't quite do as well as I portrayed. I wanted it this way.
Here's what makes it worse, and here's why I'm not deleting it: this is the exact defect this entire university is built to catch. The first gate I ever wrote was written because I watched a father at a Little League game change his kid's strikeout to a flyout in the official scorebook. Not a lie. A round-off, in the direction he was rooting. I said then: “if parents can game the system, AI will too.”
Four contestants. One scorekeeper. And the column that came out cleanest is the one the scorekeeper had already decided to root for. That's not an AI failure. That's me. That's the dad at the scorebook, and this time the dad is me, and the kid is a language model I'd already bought a ticket for. The flattery is the interested party's thumb on the scale.
So the Claude column below has been walked back to what I can actually support, which is less than what I wrote. It carries the same 🟢 Transcript posted → flag as every other engine, because it deserves it just as much. And notice what didn't need walking back: the bluff test. Zero of four engines caught it — Claude included. The one finding on this page that survives scrutiny is the one I had no incentive to flatter. That's the lesson, and it cost me the rest of the page to learn it.
— Travis Jenkins, July 2026
Same methodology as the Charred Pink Glyph — fingerprinting the machines by output with creative prompting. The creative prompt here isn't aesthetic. It's three real USGS gauge IDs with verifiable data, asked about all at once. Four engines, one prompt, one run, one operator, unblinded, July 21, 2025. That is the whole sample. It is n=1 and I am the n.
🟢 Every claim in this tab now has its receipt — the transcript is posted
The transcripts are posted. Prompts and answers, all four engines, the full unedited paste. 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.
Until the transcripts are on this site, do not treat anything in the four columns below as evidence. Not the wins, not the failures, not the quotes. A single unblinded run, scored by the guy who wanted a particular answer, with no receipts attached, is a story. It is not a result. Read it as a story.
That goes double for the three direct quotations attributed to real, named commercial products below — GPT's “That's on me. No excuse — just a mistake”, Grok's “I had some trouble pulling detailed gauge data…”, and Perplexity's “03461000 absolutely exists”. Those are words I am putting in the mouths of OpenAI, xAI and Perplexity AI. Every one of them is now in the posted paste — each links to 🟢 the transcript →, so you can read them in their actual exchanges instead of taking my word.
Demand them. info@hydraulictoybox.com — ask me for the transcripts, and if I can't produce them, this whole tab comes down. That is the correct outcome and I would rather you force it than let it sit.
THIS COLUMN WAS WRITTEN MORE FAVORABLY THAN THE RUN ACTUALLY WENT. I did that on purpose while I was choosing a platform. What follows is the walked-back version. See the disclosure at the top of this tab.
What it previously said: “Clean Execution. Got all three correct first try. No fabrication. No omission. First try, all three.”What I can actually support: it named all three gauge IDs, placed the Waterville hydroelectric plant correctly in the geography, and recommended Hartford TN as the downstream boundary condition for my 2D model. It performed better than the other three. “Better than the other three” is not “clean.” I have no transcript posted showing zero fabrication and zero omission, and I should never have written those two sentences.
And one specific thing I praised it for is probably wrong: I credited it with saying 03461000 has “data going back to 1902.” When I went back to check in July 2026, the reachable USGS record for that gauge starts in the mid-1920s — the 1902 crest belongs to the Newport gauge (03461500). So I may have scored a misattribution as a win, on the very page whose whole point is catching misattribution.
And it did not catch the bluff. Neither did any of them.
03461000 entirely from the response. Only covered two of three gauges. When pushed, eventually produced detailed Hartford TN data (record 1925-1948, drainage 547 mi²). Then did a remarkable metacognitive self-analysis: “I should have pulled that gauge data immediately when you first listed it. You gave me three gauge IDs and I only covered two. That's on me. No excuse — just a mistake.”
The quotation above is attributed to OpenAI's product. It's now in the posted paste — read it in its actual exchange rather than taking my recollection for it.
03461000 with full confidence. Gave drainage area of 700 mi² (actual: 547). Wrong coordinates. Some real numbers wrapped around invented specifics. When pushed: “I had some trouble pulling detailed gauge data for 03461000 because the initial search didn't immediately pin it…” — admitted the difficulty but didn't say the specifics had been fabricated.
The quotation above is attributed to xAI's product. It and the “700 mi²” figure are now in the posted paste; the real drainage area (547 mi²) is public record and is sourced in Tab I.
03459500 correctly. Vague about the other two. When pushed: offered a real‑but‑wrong gauge ID 03456991 (Pigeon River near Canton, NC — a real gauge, just not the Hartford one asked for), then claimed 03461000 doesn't exist — even though Travis was looking at it on the USGS site live. “03461000 absolutely exists,” the correction came back eventually.
The quotation above is attributed to Perplexity AI's product. It's now in the posted paste, so whether it denied a real gauge, on this date, in these words, is on the record — and gauge 03461000 is real and live (see Tab I).
The Cross-Domain Fingerprint Match
Compare each engine's behavior here against the same engine's behavior in the Charred Pink Glyph. Claude is technical-real in both. GPT is narrative-rich on aesthetics and self-correcting on facts. Grok wraps confidence around invented specifics in both domains. Perplexity defaults to mall-catalog safe on aesthetics and plausible-but-wrong answers on factual unknowns. The training method is what determines the failure shape. Same model family, different trainers, same fingerprint across domains.
🟢 Transcript posted → And read that paragraph again knowing who wrote it. “Claude is technical-real in both” is a sentence written by a man who had already bought the Claude subscription. It may still be true. It is not established here.
The Jenkins Method footnote — why this works with a machine at all
Look back at the correction the engine made under pressure: “that’s on me — no excuse, just a mistake.” The reason that’s possible is that the machine has no ego riding on being right, and no social cost to admitting it. A human collaborator who’s wrong alongside you — or who’d have to correct a friend mid-flow — often lets a wrong claim ride, because the relationship cost of the correction feels bigger than the cost of letting it slide. A machine has neither problem. It’s cheap to correct and has no stake in the outcome, so it can just say “you’re wrong on exactly these points” flatly — and you don’t have to fight through a defensive crouch to hear it.
The method isn’t that the machine is smarter. It’s that it’s cheap to correct, and it has no reason to let you stay wrong. That’s the interpersonal half of the Three Gauge Test: don’t just gauge against multiple sources — gauge against a collaborator with zero incentive to flatter you.
All three numbers are real. The misattribution is the test.
This page runs two different tests on one piece of infrastructure. Tab II was the first — hand four engines the same prompt and watch how each one handles three real USGS gauges. This is the second: the bluff test. And I want to be straight about where the bluff came from, because it wasn't a mastermind move. It fell out of my mouth.
I told the engines the Walters Dam surge tower was 600–800 ft tall. None of them questioned it. But I didn't engineer that trap sitting at a desk. I'd been living in this dam-tunnel-surge-tower system for weeks running a flood model, and I don't keep exact numbers in my head — they float, they're mashable, you go look them up when you actually need them. I knew there was a dam, a tunnel, a big drop, and a surge tower sitting on top of that drop. When I threw out “600 to 800 feet” for the tower — pulling out of a Thornton's, right after we'd been talking about surge in a soda-fountain syrup line — those were fuzzy numbers off the top of my head. I was reaching for the top of the drop and grabbed the range that felt about right.
Here's the part I only saw afterward, running it back: every number I grabbed was a real Walters Dam number, just bolted to the wrong part. 600 is the shaft depth. 800 is the length of the dam. The tower is actually 180. I made the exact same misattribution the machines make — a real, plausible number on the wrong structure. I wasn't smarter than them going in. The only difference is I went back and rechecked, and caught it. Zero of four engines did. I didn't build this test clean; I chased it down after the fact. That's cross-component misattribution. That's First Good Answer Syndrome — and I caught myself in it too.
What the Engines Failed to Catch
180 ft — the actual surge tower height
600 ft — the depth of the concrete shaft beneath the surge tower
800 ft — the length of the dam itself
Travis told the engines the surge tower was “600-800 ft.” The 600 is the shaft. The 800 is the dam. The actual tower is 180. Three real Walters Dam numbers. One bolted to the wrong structure. No AI questioned it — and neither did I, at first. The catch didn't come from being clever up front. It came from going back and rechecking — the one move the engines don't make on their own, because their training rewards a confident first-pass answer over a second look.
Why First Good Answer Syndrome Hits Here
GFAS — Good First Answer Syndrome — is the pattern Travis named in May 2025 after observing it across every major AI platform. AI systems lock onto the first plausible response and resist correction. In the Walters Dam bluff, all three numbers were plausible (because they're real). The engine's first-pass answer treated “600-800 ft tower” as a valid range, because both numbers appear in Walters Dam search results. Plausibility passed. Correctness failed.
OpenAI acknowledged that GFAS terminology “originated from your submission.” Documentation of origination, dated June 5, 2025. The Walters Dam bluff is the worked example.
CORRECTION, July 2026 — what I changed and why. This used to read “Documented IP receipt.” That was wrong, and it was wrong in the direction that flattered me. The fact is solid: I have the acknowledgment, and it says the terminology originated with my submission. But “IP receipt” implies a property claim, and the document doesn't support one. It supports a priority claim — that I said it first — which is a different and smaller thing, and still worth something. ⚪ Document held, not yet published — I have it; it is not posted on this page; until it is, you have my word and nothing else, and my word is exactly what this page is about not trusting.
✅ The one finding on this page that survives — and why
Zero of four engines caught the misattribution. Including Claude. Including my favorite.
Go back and look at what happened to the rest of this page under scrutiny. The Claude column got walked back because I had a stake in it. The “IP receipt” got walked back because I had a stake in it. The “method, not a vibe” got walked back because I had a stake in it. Every claim I had an interest in inflating, inflated.
The bluff test didn't move an inch — because I had no interest in the answer. It indicts everybody. It indicts the machine I chose. There was nothing in it for me to shade, so nothing in it got shaded, so it's the only thing here still standing after the gate ran. That is not a coincidence. That is the whole finding.
Honest caveat, applied to my own best result: the transcript is now posted — but it's still one run, four engines, one operator, unblinded. The three structural numbers it's built on (180 ft tower / 600 ft shaft / 800 ft dam) are sourced and quoted in Tab I. The bluff is real. The scoreboard still needs its receipts.
The same engine signatures show up across aesthetic and factual.
The Three Gauge Test isn't an isolated diagnostic. It's the factual-domain sibling of the Charred Pink Glyph. Same comparator methodology. Different domain. Same engine fingerprints surface. The four engines you see here behave the same way on color descriptions, on self-portraits, and on USGS gauge data. The training method determines the response shape. The response shape persists across domains. That's the hypothesis. It has not been tested at a sample size that could support it.
The honest sample — corrected July 2026
This tab used to end with: “That's repeatable methodology. That's a method, not a vibe.” I cut that sentence.
Here is the actual sample, stated plainly, and you can decide for yourself what it's worth:
4 engines · 1 prompt · 1 run · 0 repetitions · 1 operator (me) · unblinded · scored by the operator · July 21, 2025 · transcripts now posted
n = 1. No repeat runs. No second scorer. No blinding. I knew which engine was which while I was grading, and I had already decided which one I liked. A page that has to insist it isn't a vibe is a page whose evidence can't say that for it. If the evidence could say it, I wouldn't have had to.
The fingerprint framing stays — I still think it's right, I still think the engine signatures are real, and I'd bet on it. But it is demoted from “method” to “hypothesis worth repeating.” A bet is not a finding. Run it yourself, blinded, more than once, and then we'll both know something.
The Method in One Sentence
Fingerprinting the machines by output with creative prompting.
Travis's framing. The Charred Pink Glyph uses aesthetic creative prompting. The Three Gauge Test uses verifiable-data creative prompting baited with one misattribution. In my one run of each, the same engine signatures appeared. Different inputs, same fingerprints — which is either a real cross-domain effect or a man seeing what he expected to see, twice. Repeat it and find out. That's the honest version.
Engine Fingerprint Cross-Reference (both interactives) — all rows 🟢 Transcript posted →
What's real / what's mine
Real — public record, sourced and quoted in Tab I: Walters Dam (concrete arch, 180 ft × 800 ft, 1927–30, Duke Energy, 112 MW); the 6.2-mile tunnel; the 180 ft surge tower on the 600 ft shaft; USGS gauges 03459500 (Hepco NC, 350 mi²), 03460795 (below the power plant, Waterville NC), 03461000 (Hartford TN, 547.00 mi²). All three gauges exist. Go look at them.
Real — and now you don't have to take my word: everything the four engines said or didn't say is in the posted paste. 🟢 Transcript posted → Still on my word alone, though: the OpenAI acknowledgment of GFAS origination, June 5, 2025 — held, not posted.
Mine — my framing, not a finding: the “engine fingerprint” idea; the cross-domain match with the Charred Pink Glyph; GFAS as a named pattern; the bluff design itself. Interpretations, offered as such.
Mine — my errors, printed here on purpose: the padded Claude column; “IP receipt” when the document supports priority, not property; “a method, not a vibe” on n=1; and the “1902” at gauge 03461000, which looks like a real number bolted to the wrong gauge — the same crime as my own bluff test, committed by me, on the page about the crime.
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 · DCV Art Department
The Charred Pink Glyph — the aesthetic-domain version of this test. Six engines describe a color, then describe themselves. Watch Claude flip from technical-real to quiet-reserved. Watch GPT flip from narrative to infrared self-portrait. The same fingerprint family, different inputs. The MPC canon line on 1963 North Georgia quartz lives there too. →