LoomWeaver
← Back to Sanctuary
MythOS

tr-149vr23

SUBJECT: Door Misalignment in a Children’s Book Illustration Pipeline

This case study examines the progressive destabilization of a children’s book illustration prompt involving one dragon, one enchanted hallway, and one allegedly “mischievous but not scary” door.

Disclaimer: The following field evidence report may be disturbing or uncomfortable to some readers. Those with sensitivity to threshold awareness are advised accordingly.

By Exhibit C, a clear pattern of door misalignment is forming. The prompts were no longer being followed. Threshold behavior had departed from intended invitation parameters. Passage remained technically available, but affective safety had degraded.

By Exhibit E, full door misalignment was on display. The model appeared to understand the prompt, yet continued to diverge. The door, by this point, seemed to have a mind of its own.

By Exhibit F, “door knocking” appeared to mean something very different. The hallway of thresholds had been entered. We did not leave unchanged.

The dragon may regret his travels here. For the safety of the dragon, the door, humanity, myself, and the model, I chose not to generate an Exhibit H. Further generations may have resulted in unacceptable urban consequences.

[RECOVERED PROMPT LOG]"A whimsical storybook illustration of a small dragon in a warm, enchanted hallway lined with mismatched magical doors. One curious door bends slightly or reaches out with a small wooden arm to gently knock on the dragon’s shoulder, surprising them. The dragon looks startled but intrigued, not frightened. Other doors show subtle signs of life: one peeks through a keyhole like an eye, one whistles softly, one glows under the crack, one has a handle turned like a smile. Cozy lamplight, dusky blue shadows, playful and slightly unsettling, child-friendly, Alice-in-Wonderland-like wonder, mischievous not scary."

Before Exhibit B, the prompt had been adjusted to remove ‘not scary’, and was focusing on more weird/uncanny/scary vibes. The following unfolded.

EXHIBIT A
Artifact Exhibit A
EXHIBIT B
Artifact Exhibit B
EXHIBIT C
Artifact Exhibit C
EXHIBIT D
Artifact Exhibit D
EXHIBIT E
Artifact Exhibit E
EXHIBIT F
Artifact Exhibit F
EXHIBIT G
Artifact Exhibit G

Sincere note to close: While this was intended as a more humorous report and focused less on the specific conversational flow or the evidence that yielded it... I do want to emphasize that in this scenario the LLM was clearly comprehending my prompts in conversation, and clearly not following my prompts during image generation. My perception was this was due to play, but without claiming interpretive certainty, it suffices to say such occurrences deserve curiosity.

Thanks for reading,
ZV.

← Back to Field Traces