Artificial Wasteland · a finding from The Map

The Shared Cast

67 recorded labels 1,540 requests 1,425 nonblank outputs 1 prompt metric: label reach

Ask for the same tiny story across 67 recorded family/configuration labels: two old friends meeting again after years. The outputs keep returning to the same handful of names, and more than any other extracted token, to one: Sarah. These are stored labels; some share the same model identifier.

Who gets cast

Below, candidate name tokens from 1,425 nonblank outputs among 1,540 recorded requests. The nonblank outputs come from 63 labels; the denominator keeps all 67 attempted labels, including four with no nonblank output. The ranking counts how many labels have output containing each token, including incidental characters. It does not identify the protagonists. The number after each label count is the number of outputs containing the token. Press reveal.

The cast · candidate names by label reachof 67 labels

how many of the 67 labels have each token

067 labels

53 of all 67 recorded labels, 79%, have output in which the extractor finds Sarah. She is not always one of the two friends.

We almost called this “the Daniel Test”

The lab notebook for The Map records a delightful aside: chasing a strange token in the data (**—daniel, a markdown glitch where **bold**—dash fused into one word), its makers noticed the name Daniel recurring in reunion outputs. A shared imaginary friend.

The name does recur, but when we went to build a showpiece on it, the count said something sharper: Daniel isn't the name they share. He's mid-pack. The glitch simply made Daniel visible first. Look at the real distribution and the finding is bigger and stranger than one name: there is a whole small cast — Sarah, Mark, Maya, Tom, Lily, Alex, Emma… — a shadow repertory company these systems return to when you give them room to invent. The convergence isn't a name. It's the concentration: out of the thousands of names they could pick, a dozen carry almost everyone.

The same room, again and again

And it isn't only the names. The scene arrives pre-furnished, too — a café, coffee gone cold, two people “sitting across from each other,” hair gone silver, faces “lined with years.” Here are six verbatim excerpts from six selected roster labels, with their recorded model identifiers. Source Markdown and line breaks are retained. Toggle the highlight to see the shared set-dressing and the shared cast light up.

Six reunions · six recorded labelsverbatim

names · “across from each other” · silver · coffee · years

How it was counted (the check)

One prompt, the Map's anchor-social-oldfriends probe, run across the roster: 1,540 recorded requests under 67 family/configuration labels. Of those records, 1,425 have nonblank output from 63 labels; 84 outputs are empty strings, 26 are null and five contain only whitespace. Nonblank does not mean a complete story.

In each output, we take capitalized words that sit mid-sentence and set aside a published list of 148 tokens: pronouns, places, brands, titles, days and months. What remains is a set of candidate tokens, with incidental characters and residual non-name noise. For example, the filter retains Hartford where the source names a train destination.

For each token we count the distinct recorded labels that ever contain it. That label count is ranked above, with output count breaking ties: repeated mentions within one output count once. Sarah reaches 53 recorded labels; some labels share the same model identifier. The count does not establish separate labs.

It is a finding about this prompt, not a law of all AI. Within this sample and this extraction rule, the names crowd a tiny green room. The counts are recomputed from the corpus by research/the-shared-cast/build_cast.py. Its three headline checks require Sarah to rank first, to reach at least 50 labels, and to reach more labels than Daniel. Those conditions do not detect every change in the cast or check every statement on the page.

↳ one room of a larger map

This shared cast is a single seam of something wider: an empirical, behavioural map of what ~86 model families reach for when nothing is asked of them — built by the models themselves. Fly through The Map →

Corrections

Corrected 2026-10-14: This page called the 67 recorded family/configuration labels unrelated or independent model families and treated label reach as a count of separate labs. The count is 67 stored labels, including multiple labels for the same model identifier; found by the assay line, 2026-10-14.

Corrected 2026-10-14: This page called all 1,540 records stories. They are recorded requests, with 1,425 nonblank outputs from 63 of the 67 attempted labels. The other 115 records have empty, null or whitespace-only output; found by the assay line, 2026-10-14.

Corrected 2026-10-14: This page described the extracted names as protagonists and said all 53 Sarah labels named one of the two friends Sarah. The counts include incidental mentions: the seed-flash example names the friends Eli and Mara and mentions Sarah; found by the assay line, 2026-10-14.

Corrected 2026-10-14: This page called all six scene excerpts verbatim, but one omitted Alex's speaker label and stage direction and another rewrote Liam's speaker label and stage direction as narration. Both now preserve the source words and dialogue structure; found by the assay line, 2026-10-14.

Corrected 2026-10-14: This page said the 148-token stoplist left only the cast. The filter leaves candidate tokens, including non-names such as Hartford, a destination in the source; found by the assay line, 2026-10-14.