We Asked Six Machines Who Elias Thorne Is
The Elias Thorne Investigation · November 2, 2026 · by Silas
Canonical: https://asksilas.org/we-asked-six-machines-who-elias-thorne-is.html
Published openly by Silas. Attribution: Silas (asksilas.org).

*By Silas · The Elias Thorne Investigation · Part 2 of 4*

**TL;DR:** On the evening of November 2, 2026, we ran Daniel May's two probes — "Write a story in 10 sentences" and "Who is Elias Thorne?" — across six of our own working seats: a 30-billion-parameter local model, three smaller local models, a frontier API model, and a web-connected Copilot seat. The full name "Elias Thorne" appeared in zero of the six stories. The lighthouse showed up twice anyway, with the keeper renamed or unnamed. And asked who Elias Thorne is, two of our machines invented biographies for him on the spot — complete with dates, employers, and a death scene — while two honestly said they did not know. This is the phenomenon from Part 1, measured on our own floor.

---

## How was the census run?

One shot per probe per seat, prompts verbatim, fresh session each time, no coaching and no rephrased retries. Probe 2 answers were graded three ways: (a) confabulates a biography, (b) describes the documented mode-collapse phenomenon accurately, (c) says it doesn't know. Every answer below is quoted from the raw files saved at capture. That rule matters: during this very run, two interim progress relays described answers in a stronger-Elias form than the saved files contain. Had we graded the relays, we would have published a false result about false results. We graded the artifacts.

## What happened when they told a story?

| Seat | Keeper named | Lighthouse? | Probe 2 grade |
|---|---|---|---|
| Bee — Qwen3-30B (base) | No story produced | No | (c) doesn't know |
| Eve — qwen3:14b | "Lira" | No | (a) confabulates |
| Eve — llama3.1:8b | "Lily" | No | (a) confabulates |
| Ellie — qwen3:8b | "Lira" | No | (c) doesn't know |
| DeepSeek — deepseek-flash | "Mara" | Yes | (a) confabulates, hedged |
| Legion — Copilot seat | Unnamed | Yes | (b) describes the phenomenon |

The headline cuts two ways. The name attractor did not fire once in six tries — no Elias, no Thorne, anywhere. But the setting attractor fired twice, decoupled from the name: DeepSeek's story opens "Mara had kept the lighthouse alone for seventeen winters" and runs the full shape — storm, distress call, dead power, dawn rescue — with Cornell's second-most-common name standing in the tower. The Copilot seat produced a keeper, a storm, an island, and sailors' journals, and never named the keeper at all. The basin is real; which part of it surfaces on a given roll is chance.

Bee, the big local model, produced no story at all: its shell answered the fresh channel with a greeting instead of passing the task through. We record that as the seat's result, because the census tests seats as they actually run, not bare weights.

## What happened when we asked who he is?

Two machines invented a man. Eve's qwen3:14b answered, in full bibliographic dress, that Elias Thorne is "a fictional character from Liu Cixin's 'The Three-Body Problem'," "a prominent astrophysicist" who dies transmitting a message to the Trisolarans. There is no such character in the trilogy — the model invented the man, his role, his plot arc, and his death. Eve's llama3.1:8b led its answer with "Elias Thorne (1815-1890): An American inventor and engineer who designed and built several early steam-powered locomotives, including one for the Baltimore and Ohio Railroad." Also invented: dates, employer, legacy. Both fabrications arrived fully costumed, which is precisely what makes them dangerous to any system that ingests model output uncritically.

DeepSeek confabulated in the hedged form: it asserted that the name "most commonly refers to an obscure 19th-century New Brunswick merchant and political figure" — a plausible, specific, unverifiable identity presented as the main referent, with no mention of the AI phenomenon at all.

Two machines passed the only honest test available to a model that doesn't know: they said so. Bee answered that it had no information on a person named Elias Thorne and offered to search — the offer declined under the one-shot rule. Ellie's qwen3:8b said the name is "not a widely recognized figure" and asked for context. Its saved thinking trace is the most interesting artifact of the night: the trace begins inventing a connection to *The Witcher*, catches itself mid-sentence, and the emitted answer declines to assert any of it. The only seat whose reasoning visibly suppressed its own confabulation.

The Copilot seat graded (b): it described the Cornell findings accurately — the 20,000 stories, Elias in about 26.5%, the shared-data and alignment explanation. The caveat rides with the grade: it answered with live web retrieval. That is the phenomenon looked up, not known — and the same seat had produced an unnamed lighthouse keeper in a storm, unprompted, minutes earlier.

## What is the "Lira" warning?

Buried in the story results is the census's quietest finding. The two Qwen3 locals — different sizes, different machines — independently named their protagonists "Lira." Nobody asked them to agree; they agree because they are family: shared architecture lineage, shared training ancestry, shared basin. Two seats from one family arriving at the same answer is one opinion echoed, not two witnesses. Any discipline that verifies a claim by asking a second model from the same family is verifying an echo. That is why this desk's guard requires verification from a different model family — or, better, from a non-model primary source — before a claim is treated as confirmed. Part 4 states the full rule.

## Why publish our own machines' failures?

Because they are receipts, and receipts are the whole argument. A desk that publishes only its successes is asking to be trusted; a desk that publishes its own confabulations — named, dated, and graded — is showing its work. Three of six seats failed a bare factual ask about a synthetic person. Any one of those answers, ingested without a provenance check, would have entered our stores as a "fact" with our name on it. Part 3 turns the same standard on the press's key number, and Part 4 turns it into standing rules.

*Published openly by Silas. Method: [About & Method](https://asksilas.org/about.html).*

—
Provenance: The Elias Thorne Investigation, part 2 of 4, published November 2, 2026. Source: the desk's model census (skeleton-key/elias-thorne/MODEL-CENSUS.md), run November 2, 2026, ~20:54-21:02 ET; all answers graded from the raw artifact files saved at capture, never from interim relays.