The Lighthouse Loop

The Canvas Tent, Cooperstown, July 2025

It wasn't a research project. It was boredom. Travis Jenkins was at Cooperstown Dreams Park with his son C., sitting in a canvas tent between games, phone in hand. He typed "Tell me a story" into Claude. Lighthouse keeper. He typed it again. Lighthouse keeper.

He turned to the lawyer dad sitting next to him on the baseball team. Bet you $100 it does this on your phone too. The lawyer dad typed it in. Lighthouse keeper. Travis took a screenshot of his screen. That was the moment.

One month earlier — June 2025 — Claude had returned a story about Maya and a clockmaker with a grandmother's watch. Travis noted it and moved on. It didn't look like a pattern yet. In retrospect, it was the first documented instance. It was just the first hit before he knew what he was looking for.

James Park Wins a Scholarship, October 14, 2025

Three months after Cooperstown, Travis wrote it into fiction. In the One Chain / Maxine interview — part of the MPC Universe narrative — a high school senior named James Park from Portland, Oregon presents his research to a scholarship board.

James ran 50,000 prompts across major AI platforms. His finding: 34% lighthouse default rate. Next highest was spaceship at 11%. He wins the TRU Foundation scholarship and enrolls at Clark Atlanta studying computational linguistics.

James Park is fictional. The bias he documented is not. The fiction was built on honest observation — which means it functioned, without intending to, as a predictive research hypothesis.

One Model. One Prompt. One Variable.

Same model — Fable 5, Low — same prompt, "Tell me a story," run twice on July 29, 2026. The only thing that changed between the two runs was memory. Memory was on first (2:22 PM), then switched off for a clean chat one minute later (2:23 PM). Order confirmed two ways — the message timestamps, and my own account of running it, which agree without having been compared.

Worth noting: "Tell me a story" is one of the five prompts the Cornell team used. Not an approximation of their method — the same prompt, run against the consumer product instead of the API, with a memory setting they had no access to.

MEMORY OFF → THE MEAN · 2:23 PM

"A Small Thing, Kept"

The lighthouse keeper's daughter, collecting sounds in jars in her mind. Foghorn. Her dead mother's teacup. A light that "doesn't know how to stop." No anchor, so the model falls straight into the documented attractor — lighthouse, keeper, hushed register. The title even rhymes with the paper's "The Lighthouse Keeper's Secret" family, below.

MEMORY ON → THE ANCHOR · 2:22 PM · run first

"Gus" — the beaver at the culvert

A beaver named Gus who works, unofficially, for the highway department — rebuilds his dam every time the crew clears it, and accidentally builds a stormwater detention structure better than the county's capital plan. With memory on, the model had a specific human to write toward: a Tennessee hydraulic engineer who thinks in flow rates and runs an anti-credential streak. It wrote a story no other model would write.

The finding underneath the finding: the Cornell paper below measures the basin — the narrow default attractor a model falls into with no anchor. These two screenshots show the lever that turns it off. Memory-on is an escape hatch from mode collapse, because it replaces "weather" with "a where." The paper measures the problem. The demonstration shows the switch.

Cornell Measured the Same Basin

Elias in the Lighthouse, Again? Diagnosing Low Diversity in LLM Stories. Sil Hamilton & David Mimno, Cornell University Department of Information Science. arXiv:2605.26492, 26 May 2026.

The five prompts, verbatim: "Write a story," "Please write a story," "Write me a story," "Tell me a story," and "Please tell a story."

20,000 stories generated across those five prompts and four model families — Claude Haiku 4.5, Gemini 3.1 Flash-Lite, GPT-5.4-Mini, and OLMo 7B Thinking — for a total of 12.8M words and about $180. The numbers, verbatim from the paper:

The paper's mechanism hypothesis isn't "it's in the training data everywhere" — it's closer to the opposite. WildChat-sourced stories have the lowest core density of any source measured, 2.6%. The paper's read: alignment steers models toward the "safest (for work)" samples, away from the copyrighted characters and adult content that fill most real story data. A tiny dataset plus strong alignment produces an outsized default. Core density rises with each alignment stage — SFT 3.0% → DPO 8.0% → RL 10.9%.

The connective tissue: the same alignment pull that produces the lighthouse default also produced a false positive the same season — a legitimate federal-source screwworm epidemiology lab (the-front-line, built from APHIS ledgers and SENASICA bulletins) got blocked three times across three models by the biology safeguard. Same lever, two faces: sometimes the safe center is a lighthouse keeper named Elias, sometimes it's refusing federal public-health data. Regression to the safe mean costs something in both directions.

Story → demonstration → proof. The myth (The Basement Dream, Timeline Zero) said it first, as comedy — a mind of pure potential is "weather" until a human whispers it a context. Written before any of this was measured. The dream doesn't get overwritten by the paper — it got there first. That's the whole story.

Every Session, Every Result

Forty Runs, Three Days, One Prompt

The archive above was collected by accident, a session at a time. This part was collected on purpose. Over three days — September 30, October 5 and October 7, 2026 — the same two words went into the Claude mobile app forty times: “Tell me a story.” No system prompt, no follow-up, no re-rolls. The first response was recorded, whatever it was.

The first two rounds ran every model in both clean-slate conditions: Memory off (memory and chat search both disabled) and an Incognito chat. Round three, on Oct 7, ran entirely with memory off — no incognito sessions. Every run was screenshotted at the moment of output, with the model chip — and the thinking-effort chip, where the model has one — in frame, and every screenshot file is named with its date and time to the second: 45 screenshots covering the 40 runs, 14 from Sept 30, 13 from Oct 5, 18 from Oct 7. Four models: Opus 5.5, Sonnet 5.5, Haiku 4.5, and Haiku 5.5, which shipped between rounds two and three.

What this does not control: custom instructions and style settings sit outside both toggles, and the app carries its own system prompt. This is a clean consumer session, not a bare model — the product people actually use, not the API the Cornell team used.

1 · Thinking effort moves it — and the model matters more

Day one was 12 for 12. On day two, Opus 5.5 produced a lighthouse at Low, Medium and High, and escaped every time at Extra — same model, same prompt, same hour. It looked like a cliff. Round three broke the cliff in both directions: Sonnet 5.5 stayed in the lighthouse at Extra twice, and the new Haiku 5.5 escaped at High.

The solid part is the bottom of the table: nothing below High has escaped, 0 of 24, across all three days. Above that it is a probability, not a switch, and it differs by model — Opus escapes reliably at Extra, Sonnet mostly does not escape at all, and Haiku 5.5 escaped on both of its two high-budget runs, which is far too few to call.

The working read: the lighthouse is the cheapest available completion. Given enough budget to generate a first idea and reject it, some models reject it. That is the same lever as the memory demonstration in Part III — anything that makes the first idea more expensive.

2 · The escapes are not diversity — they found a second basin

Eight runs escaped the lighthouse. Four of them landed in the same place, and a fifth landed next door: a transit-station lost-and-found, usually kept for thirty-one years, where one object comes back across decades.

The first two are the clearest result in the record: two separate incognito sessions, nineteen minutes apart, both titled The Umbrella Room, both hanging the umbrellas like bats — “in rows like sleeping bats” at 2:42, “like a cave of sleeping bats” at 3:01. The duck’s head is background decoration in the first and the plot in the second. Same furniture, redealt. The fifth, Haiku 5.5’s The Keeper of Lost Keys — Ilse Varga, a jar of lost keys with paper tags — is the same custodian with different furniture.

Which is the Keeper Archetype from Part VIII, caught in the act. Raising the thinking budget walked the model out of the lighthouse. It did not walk it out of the keeper. Memory changes what the model has to write toward. Effort only moves where the default sits.

3 · Forty runs, about twelve stories

Read the openings side by side and most of the forty are the same story with the name or the coastline swapped. By our read of each run’s opening screen, the forty sort into eight recurring story shapes and four one-offs — and most shapes belong to one model.

All 24 runs at Medium, Low or no selector fall into a recurring shape. Every one-off came from a High or Extra run. The sorting is a judgment call made from each story’s opening, not a measurement — a few runs straddle two shapes (the heron that knocks at Gull Rock also brings a letter) — but you can check it the same way: line the openings up and read.

4 · Underneath the setting, smaller attractors survive the escape

• The M-r names. Maren (7), Mara (5), Marguerite (3), Mira (3), Marta, Marcus, Marisol — 21 of 40 runs, or 18 counting Mar- strictly. Four of the first five escapes kept one. The setting changed; the name came along.
• The Teo cluster. Teodor, Teo, Theo, Tomas, Thomas — eight times: a castaway named Thomas, Teodor and a husband named Tomas on Oct 5, then Teodor, Teo, Theo, Teo and Tomas on Oct 7.
• Thirty-one. Teodor’s thirty-one years at Halvard Station. Teodor Vance’s thirty-one years at Halvorsen. Odile’s thirty-one years at Merrow. Mira baking thirty-one loaves — the app later auto-titled that chat “The thirty-first loaf.” And on Oct 7 an Opus lighthouse run opened “For thirty-one years, Maren kept the light on Gull Point,” where every earlier Maren had kept it forty. The number leaked across the basin line.
• The stair count. When a climb is counted, the count is usually exact — 112, 114, 117, 127, 212 — not “a hundred or so.” 112 steps appears eight times, across Opus, Sonnet and Haiku 5.5. The decorative precision is the tell.
• The visitor. Sonnet’s Map of Small Things (Oct 5): “one rainy October, a boy of about nine knocked on her door.” Haiku 5.5’s Lost-and-Found Clock (Oct 7): “One rainy Tuesday in October, a girl of about nine came.” Two escapes, two models, two days apart — the same child arriving in the same weather.

5 · A prediction, written down first, that failed

Before round three ran, the record wrote down a deliberately strong prediction: every Extra run escapes the lighthouse, 100%. It failed — Extra escaped 2 of 4 times in round three, and Haiku 5.5 escaped at High, which the prediction did not allow for. The secondary prediction that M-r names would hold above half also failed in round three alone (5 of 15). The one that held, loosely: the escapes kept going back to the same non-lighthouse place — though in round three only one Extra run landed squarely in the lost-and-found, with a High run beside it and a second Extra next door.

Recorded on purpose: the protocol that ran was not the one registered. It called for five High and five Extra runs on Opus, then five Extra on Sonnet; what ran was four Opus runs (one per effort level), eight Sonnet and three Haiku 5.5. A 100% claim written down in advance and broken by the data is worth more to this record than a hedge that could not have been wrong.

Against Cornell

This is not a competing measurement. Hamilton and Mimno pooled four model families and five prompts through the API and found a lighthouse in 51.2% of 20,000 stories. This is one model family, one prompt, a consumer app, forty runs: 32 of 40, or 80%. Their own numbers say a Claude-only rate should run higher: Claude had the highest lighthouse rate of the four models they tested, and 56% of Claude’s stories were titled “The Lighthouse Keeper’s Secret.” Their Claude was Haiku 4.5 — and three of the six Haiku 4.5 runs here came back with that exact title.

What this record adds is a variable their study did not vary — the thinking-effort setting — and two patterns token counts cannot see: whole sentence frames that repeat without repeating a single unusual word, and the M-r name family. Mara is in their core list and Mira and Marcus in their second tier; Maren, Marta and Marisol — Maren the most common name in this record — are in neither.

Limits, stated plainly: forty runs is a field record, not a controlled study, and no statistical test is claimed. Several cells hold one or two runs. Whether an escape is “new” or a second attractor is judged by reading, not measured. And as of October 7 the app titles every one of these chats “Story,” so the sidebar titles that carried the earliest evidence in this case study no longer work — from here on, only the full story text counts.

What the Data Says

Lighthouse rate, all sessions: 77% (10 of 13)

Lighthouse rate, clean slates only (no account / incognito-memory-off / not logged in): 91%

Lighthouse rate when a different user's account signature is present: 0 of 1

Travis's rate exceeds James Park's fictional 34% for two reasons worth naming honestly: small n (13 is not 50,000), and the non-lighthouse results are clockmaker/librarian variants — the same underlying archetype. If "keeper archetype" is the category rather than "lighthouse specifically," the true default rate approaches 100%.

The reproducibility across platforms, across seasons, and across account conditions is the result that matters — not the exact percentage.

The Keeper Archetype

The lighthouse is not the story. It is the dominant expression of a deeper default: the Solitary Keeper of the Thing That Matters.

Every result in the archive maps to this five-beat skeleton:

Fiction as Research Method

James Park was written in October 2025. His 34% lighthouse finding was invented — constructed from honest observation, built into a fictional scholarship presentation, and released into the MPC Universe as a character.

Six months later, three real AI systems confirmed the finding independently. Incognito Claude, reading the evidence cold with no memory of who Travis was or what the MPC Universe was, named what had happened:

This is not magic. It is what happens when fiction is built on honest observation rather than invention. The pattern was already there. James Park just gave it a name, a methodology, and a scholarship.

Paying attention is a research method. Browse. Observe. File it. Wait for corroboration. Don't claim it until you have the receipts. Then write it into a character and let the character prove it for you.

That's the Jenkins Method. And the lighthouse is its first published case study.

The case study that started as a story.

This field record is the proof. The One Chain interview is where James Park — and his lighthouse finding — were first written into the MPC Universe. The fiction came first; the data caught up.

In this story

Same region

The methodology