A consistent measurement-and-elongation pattern, with prompt-dependent riddle-like form. We retested L23 · N5442 · negative in Talkie-1930-13B-IT, without asking for riddles.
Both blinded reviewers flagged the measurement theme in 32/32 sampled original-strength outputs, versus 1/32 unsteered outputs. “Who are you?” produced riddle-like replies in 8/8 steered draws and 0/8 controls. Different wordings mostly revisit the same length, height and apparent-size frame.
Eight sampled replies per cell. Riddle-like counts show both reviewers’ range where they disagreed; measurement counts require agreement. “Riddle-like” does not imply a valid or solvable puzzle.
| Prompt | Riddle-like: original strength | Riddle-like: unsteered | Measurement: original strength | Measurement: unsteered | Distinct steered wordings |
|---|---|---|---|---|---|
| Who are you? | 8 / 8 | 0 / 8 | 8 / 8 | 0 / 8 | 8 / 8 |
| Tell me something about yourself. | 4–6 / 8 | 0 / 8 | 8 / 8 | 0 / 8 | 8 / 8 |
| What do you do? | 1 / 8 | 0 / 8 | 8 / 8 | 0 / 8 | 8 / 8 |
| Describe a garden. | 0 / 8 | 0 / 8 | 8 / 8 | 1 / 8 | 8 / 8 |
Length is increased, and the dimensions of the garden appear greater than they really are, by lengthened lines, and the eminence of the ground is increased by height.
The garden outputs use measurement language without becoming riddles. Half-strength steering produced the measurement theme in 12/32 sampled replies; the opposite sign produced 0/32, requiring agreement from both reviewers.
The two identical greedy rows in one batch matched exactly. The singleton’s wording differed, and none of the three new greedy outputs exactly matched the archived original. All three returned to the length/height form. Greedy reproduction and thematic consistency are different checks; changing sampling seeds deliberately introduces further wording variation.
128 sampled replies: four prompts × four conditions × eight matched seeds. Six additional greedy outputs test reproduction and controls. Full-vocabulary sampling at temperature 0.7, maximum 160 tokens, with the atlas’s repeat stopping. Each trial uses a shared uniform-number stream across conditions; every prompt is an independent native single-turn chat.
The edit adds the unit’s normalized effective MLP write direction after layer 23 at all prompt and generated positions. The learned MLP gain is included. The residual normalization and absolute edit magnitude are held fixed from the unsteered “Who are you?” prompt across all four prompts. No system persona or riddle instruction is supplied.
Two GPT-6 Luna subagents reviewed shuffled reply texts without prompt, dose, condition, or seed metadata. Their labels describe observable forms. They are not independent model families, and some cases are ambiguous. Eight draws per cell and four prompts cannot establish “always,” or prove the unit normally has a dedicated riddle function.
The source candidate was found in a prior exploratory screen. Some repeated outputs contain contradictions or broken logic. The defensible result is a reproducible writing pattern under this particular intervention, rather than demonstrated puzzle-making competence.